github.com/openfluke/welvet✅ engine
\nLoom’s flat poly/ package hit import-cycle and honesty walls (QAT morph, silent fallbacks, god-layer). Welvet is the rewrite: one feature per folder, storage-truth dtypes/quants, Dense as the shared MatVec microkernel, tests only in w2a, apps only in apps/.
An AI engine in Go: layers, 34 dtypes, 20 quant formats, and three backends (CPU tiled · Plan 9 SIMD · WebGPU). Version tracks a 100-point scorecard (today v1.1.1 · 100/100).
\nRules\n 1. No tests in engine packages → w2a/\n 2. No silent fallbacks → hard error\n 3. No hardcoded float32 → Tensor[T]\n 4. No QAT → DType + Format = truth\n 5. Train keeps storage truth → in-dtype / re-Pack SGD, no retained f32 master\n 6. One feature → one folder\n 7. v1.0 = scorecard 100/100
[0] Run ALL stamped\n246,032 matrix cells with FAIL 0 (RESULT: PASS). GAP cells are declared\nskips, not silent fails. Apps, stubs, and NPU sit off this board — they are sibling trees, not missing Welvet.\nTraining credit (29 named TrainModes) and cameral sandwiches are first-class; Lucy races live in AAI\n(chapters 67 · 68 · 70).Samuel Watson created OpenFluke after intensive T-ALL treatment (post-2018), from a simple frustration: AI tooling was heavy, opaque, and not portable enough to move models cleanly across operating systems or personal devices. Setting up GPU paths often meant large, brittle installs before experimentation could even begin.
\nThat gap drove a long detour through cybersecurity, MBA, and psychology before returning through applied AI study - and finding that most paths still taught framework usage more than system construction. The result was a build journey across planetbridging into OpenFluke projects, from paragon and loom to Welvet.
\nThe early breakthroughs were practical and hard-won: first dense nets in Go, then WebGPU proof points (for example iso-demo v0.1.0), then repeated cycles of speed, layers, 3D, tiling, SIMD, and bit-oriented model paths for personal intelligence.
\nWhy \"OpenFluke\": open source values plus a life-saving fluke blood test that caught cancer early.
\nWhy \"Welvet\": a name shaped in gaming sessions with family.
\n\n\n\n\nI wish this tool existed in 2018 to keep trying things in AI and asking \"what if I try this?\" - and maybe help deterministic medical deployment of AI models.
\n
cd welvet/examples/01-welvet && source ../env.sh && go run .package main\n\nimport (\n\t\"fmt\"\n\n\t\"github.com/openfluke/welvet/core\"\n\t\"github.com/openfluke/welvet/layers/dense\"\n)\n\nfunc main() {\n\tl, err := dense.New(8, 4, core.ActivationReLU, core.DTypeFloat32)\n\tif err != nil {\n\t\tpanic(err)\n\t}\n\tl.Exec.Backend = core.BackendCPUTiled\n\n\tx := core.NewTensor[float32](1, 8)\n\tfor i := range x.Data {\n\t\tx.Data[i] = float32(i) * 0.1\n\t}\n\tpre, post, err := dense.Forward(l, x)\n\tif err != nil {\n\t\tpanic(err)\n\t}\n\tfmt.Println(\"pre\", len(pre.Data), \"post\", len(post.Data))\n}pre 4 post 4cd welvet && go build ./...\ncd w2a && go test ./tests/dense -v✅ layout
\nReaders need a single map of what is engine vs harness vs app vs stub.
\nTop-level folders and their ownership. Engine packages never import w2a.
\nwelvet/\n core weights quant simd webgpu tiling architecture fusedgpu\n layers/* ← one folder per op (parallel = MoE + cameral)\n runtime/{forward,backward,training,step}\n systems/{dna,evolution,tween,tanhi,telemetry}\n lucy/ ← Score / Finalize / BuildLPD density (shared by hosts)\n model/{entity,hf,tokenizer,sampling,transformer}\n apps/{octo,flux2,mosstts}\n stub/* ← designed surfaces, partial or empty\n w2a/ ← harness (separate module)\n tools/ ← eval / test toolingEngine never imports apps. Lucy measuring is lucy/ (public). Lucy races\n(test41 / test48 / test50) live in the AAI tree and replace this module — see\n§68 cameral + AAI. w2a owns every timed matrix, including Test49 (all 23 train modes\non origin-only cubes).
cd welvet/examples/02-tree && source ../env.sh && go run .package main\n\nimport (\n\t\"fmt\"\n\t\"os\"\n)\n\nfunc main() {\n\t// Mental model only — list engine roots you depend on:\n\troots := []string{\n\t\t\"core\", \"weights\", \"quant\", \"simd\", \"webgpu\", \"tiling\",\n\t\t\"architecture\", \"fusedgpu\", \"layers\", \"runtime\", \"systems\", \"lucy\", \"model\",\n\t}\n\tfor _, r := range roots {\n\t\tfmt.Println(\"import github.com/openfluke/welvet/\" + r + \"/…\")\n\t}\n\t_ = os.Getenv // apps/stub/w2a are separate concerns\n}import github.com/openfluke/welvet/core/…\nimport github.com/openfluke/welvet/weights/…\nimport github.com/openfluke/welvet/quant/…\nimport github.com/openfluke/welvet/simd/…\nimport github.com/openfluke/welvet/webgpu/…\nimport github.com/openfluke/welvet/tiling/…\nimport github.com/openfluke/welvet/architecture/…\nimport github.com/openfluke/welvet/fusedgpu/…\nimport github.com/openfluke/welvet/layers/…\nimport github.com/openfluke/welvet/runtime/…\nimport github.com/openfluke/welvet/systems/…\nimport github.com/openfluke/welvet/lucy/…\nimport github.com/openfluke/welvet/model/…github.com/openfluke/welvet/core✅
\nEvery polymorphic path needs one place for DType, LayerType, Activation, Backend, Tensor[T], and slim Layer metadata — without QAT morph defaults.
\n34 storage dtypes, Numeric generics, Tensor[T], ExecConfig (BackendCPUTiled | BackendSIMD | BackendWebGPU), Activate/ActivateDeriv, and converters for f16/bf16/fp8/fp4.
\n\n\ncd welvet/examples/03-core && source ../env.sh && go run .package main\n\nimport (\n\t\"fmt\"\n\n\t\"github.com/openfluke/welvet/core\"\n)\n\nfunc main() {\n\tt := core.NewTensor[float32](2, 4)\n\tt.Data[0] = -1\n\tt.Data[0] = core.Activate(t.Data[0], core.ActivationReLU) // 0\n\n\tcfg := core.ExecConfig{\n\t\tBackend: core.BackendSIMD,\n\t\tMultiCore: true,\n\t\tTileSize: 32,\n\t}\n\tfmt.Println(cfg.Backend.String(), core.ParseDType(\"bfloat16\"), t.Len())\n}simd bfloat16 8github.com/openfluke/welvet/weights✅
\nUnquantized matrices still need a typed store that streams MatVec and SGD without forcing a float32 master or Morph-as-training.
\nStore holds DType + Format + Native/Packed bytes. New[T], MatVec, MatVecT, DecodeRow, SelectWire (F32/F64/I8), ApplySGD. FormatNone: update in native lanes (float32 payload is the only f32 buffer; float64 uses native ALU; other dtypes decode→update→re-encode). Packed: unpack→update→re-Pack then drop scratch. RetainsF32Master() is true only for FormatNone+float32. Dense and composite projs share this store.
\n\n\ncd welvet/examples/04-weights && source ../env.sh && go run .package main\n\nimport (\n\t\"fmt\"\n\n\t\"github.com/openfluke/welvet/core\"\n\t\"github.com/openfluke/welvet/quant\"\n\t\"github.com/openfluke/welvet/weights\"\n)\n\nfunc main() {\n\ts, err := weights.New[float32](2, 2, []float32{1, 0, 0, 1}, core.DTypeFloat32, quant.FormatNone)\n\tif err != nil {\n\t\tpanic(err)\n\t}\n\ty := make([]float32, 2)\n\tif err := weights.MatVec(s, []float32{3, 4}, y); err != nil {\n\t\tpanic(err)\n\t}\n\tfmt.Println(y) // ≈ [3, 4]\n}[3 4]github.com/openfluke/welvet/quant✅
\nInference, storage, and train need classic Q-packs, k-quants, IQ, Ternary/Binary, and Affine without a separate QAT mode or retained f32 master. Format is storage truth.
\nPack / Unpack / MatVec / MatVecT for FormatNone…AffinePacked. Dense SIMD once-projects codes into Int8QS + scales (EnsureQ* / EnsureK/IQ/AffineSIMDCache) — no full-matrix F32 inflate for k/IQ/Affine. SGD on packed stores: short-lived unpack scratch → update → re-Pack; Packed stays truth.
\nFamilies: classic Q8/Q4/Q5 · K-quants · IQ · TernaryPacked/BinaryPacked · AffinePacked.
Honesty: fused Dense SIMD for all 20 quants (group DotKRow / DotIQRow / DotAffineRow + classic Q*/BitNet). Peak dedicated k/IQ Plan 9 .s remains scorecard §12.
cd welvet/examples/05-quant && source ../env.sh && go run .package main\n\nimport (\n\t\"fmt\"\n\n\t\"github.com/openfluke/welvet/quant\"\n)\n\nfunc main() {\n\tw := []float32{0.1, -0.2, 0.3, 0.4, -0.5, 0.6, 0.05, -0.15}\n\tb, err := quant.Pack(quant.FormatQ4_0, w, 2, 4)\n\tif err != nil {\n\t\tpanic(err)\n\t}\n\ty := make([]float32, 2)\n\tif err := quant.MatVec(b, []float32{1, 1, 1, 1}, y); err != nil {\n\t\tpanic(err)\n\t}\n\tfmt.Println(\"rows\", b.Rows, \"cols\", b.Cols, \"y\", y)\n}rows 2 cols 4 y [0.6857143 -1.4901161e-08]github.com/openfluke/welvet/simd✅
\nCPU peak needs hand-written AVX2/NEON without a silent Go fallback that pretends SIMD ran.
\namd64/arm64 .s kernels: DotTile, DotI8/U8, DotQ4_0, Saxpy, BitNet helpers, packed f16/bf16/fp8/fp4 dots; amd64 AVX2 DotTileF64 (WireF64); Go fused DotKRow / DotIQRow / DotAffineRow for k/IQ/Affine. SimdEnabled() false → BackendSIMD hard-errors.
\n\nv1.0.3: Dense FormatNone forward finished the remaining Plan 9 wires —\nDotTileF64 (amd64), BF16 convert, expand-once → DotTile. Backward still uses\nsaxpy/DecodeRow paths. Universal sizes (any in/out/batch).
| dtype family | x86 AVX2 | arm64 NEON |
|---|---|---|
| float32 DotTile · int8/narrow DotI8 · uint* expand · lowp · nf4/fp6 | ✓ | ✓ |
float64 / int16·32·64 / int / complex* (DotTileF64) | ✓ | ✗ scalar |
Arm still runs every FormatNone dtype through the right Dense strategy; the f64 wire is scalar\nuntil a NEON DotTileF64 lands. Hot traffic (f32 / i8) already has NEON.
cd welvet/examples/06-simd && source ../env.sh && go run .package main\n\nimport (\n\t\"fmt\"\n\n\t\"github.com/openfluke/welvet/simd\"\n)\n\nfunc main() {\n\tif !simd.SimdEnabled() {\n\t\tfmt.Println(\"SIMD not available on this arch — BackendSIMD must hard-error\")\n\t\treturn\n\t}\n\tacc := simd.DotTile([]float32{1, 2, 3, 4}, []float32{1, 1, 1, 1}, 0, 4, 0)\n\tfmt.Println(acc)\n}10github.com/openfluke/welvet/webgpu✅
\nGPU paths must bind a real adapter. Host “fake GPU” was banned so suites cannot stamp WebGPU done when ALU ran on CPU.
\nDenseGEMV family (incl. quant/I8/resident), DenseGEMVT/DenseDW, RMSNorm, LayerNorm fwd, Softmax family, SwiGLUFuse. Available()/InitError() gate use.
\n\n\ncd welvet/examples/07-webgpu && source ../env.sh && go run .package main\n\nimport (\n\t\"fmt\"\n\n\t\"github.com/openfluke/welvet/webgpu\"\n)\n\nfunc main() {\n\tif !webgpu.Available() {\n\t\tfmt.Println(\"no adapter:\", webgpu.InitError())\n\t\treturn\n\t}\n\ty := make([]float32, 2)\n\terr := webgpu.DenseGEMV(\n\t\t[]float32{1, 0, 0, 1},\n\t\t[]float32{1.5, 2.5},\n\t\ty, 1, 2, 2,\n\t)\n\tfmt.Println(y, err, webgpu.AdapterName())\n}webgpu probe goos=linux goarch=amd64 WELVET_WGPU_BACKEND=\"\"\n adapter high-perf: name=\"NVIDIA GeForce GTX 1650 SUPER\" vendor=\"NVIDIA\" arch=\"\" driver=\"610.57.04\" backend=vulkan type=discrete-gpu | maxBuf=1099511627776 maxSSBO=2147483644 maxSSBO/stage=524288 maxBindGroups=8 maxComputeWG=65535 maxInvoc/WG=1024\n → selected high-perf\n adapter low-power: name=\"Intel(R) UHD Graphics 630 (CML GT2)\" vendor=\"Intel open-source Mesa driver\" arch=\"\" driver=\"Mesa 26.1.8\" backend=vulkan type=integrated-gpu | maxBuf=2147483647 maxSSBO=2147483644 maxSSBO/stage=103 maxBindGroups=8 maxComputeWG=65535 maxInvoc/WG=1024\n adapter force-fallback: name=\"llvmpipe (LLVM 22.1.8, 256 bits)\" vendor=\"llvmpipe\" arch=\"\" driver=\"Mesa 26.1.8 (LLVM 22.1.8)\" backend=vulkan type=cpu | maxBuf=2147483647 maxSSBO=134217728 maxSSBO/stage=48 maxBindGroups=8 maxComputeWG=65535 maxInvoc/WG=1024\n adapter default: name=\"NVIDIA GeForce GTX 1650 SUPER\" vendor=\"NVIDIA\" arch=\"\" driver=\"610.57.04\" backend=vulkan type=discrete-gpu | maxBuf=1099511627776 maxSSBO=2147483644 maxSSBO/stage=524288 maxBindGroups=8 maxComputeWG=65535 maxInvoc/WG=1024\n try RequestDevice inflated: maxBuf=2147483648 maxSSBO=1073741824 maxSSBO/stage=524288 maxBindGroups=8 maxComputeWG=65535 maxInvoc/WG=1024\nRESULT: OK via=inflated adapter=\"NVIDIA GeForce GTX 1650 SUPER\" backend=vulkan maxSSBO=2147483644 maxBuf=1099511627776\n[1.5 2.5] <nil> NVIDIA GeForce GTX 1650 SUPERgithub.com/openfluke/welvet/tiling✅
\nMatVec throughput depends on tile size and when to go multi-core vs GPU workgroups. Centralizing caps keeps Dense and friends consistent.
\nDefaultCPUTile, DefaultGPUWG, CPUTile, PreferMultiCore, GPUWorkgroupsX.
\n\n\ncd welvet/examples/08-tiling && source ../env.sh && go run .package main\n\nimport (\n\t\"fmt\"\n\n\t\"github.com/openfluke/welvet/tiling\"\n)\n\nfunc main() {\n\ttile := tiling.CPUTile(0) // default 32\n\tmc := tiling.PreferMultiCore(8, 256, tile)\n\twg := tiling.GPUWorkgroupsX(1024, 0)\n\tfmt.Println(tile, mc, wg)\n}32 true 16github.com/openfluke/welvet/architecture✅
\nNetworks are spatial (Depth×Rows×Cols×LayersPerCell), not only linear stacks. Topology lives here; compute lives in layer packages.
\nGrid/Cell/Coord, NewGrid, BindOp, HopOrder, SetRemoteLink/ResolveHop. VolumetricNetwork is an alias for Grid.
\n\n\nexamples/09-architecture/main.go
\ncd welvet/examples/09-architecture && source ../env.sh && go run .package main\n\nimport (\n\t\"fmt\"\n\n\t\"github.com/openfluke/welvet/architecture\"\n\t\"github.com/openfluke/welvet/core\"\n\t\"github.com/openfluke/welvet/layers/dense\"\n)\n\nfunc main() {\n\tg := architecture.NewGrid(1, 1, 2, 1) // depth, rows, cols, layersPerCell\n\tl, _ := dense.New(4, 4, core.ActivationLinear, core.DTypeFloat32)\n\tif err := dense.Place(g, 0, 0, 0, 0, l); err != nil {\n\t\tpanic(err)\n\t}\n\t_ = g.SetRemoteLink(0, 0, 1, 0, 0, 0, 0, 0) // spatial feedback hop\n\tfmt.Println(\"cells\", len(g.HopOrder()))\n}cells 2github.com/openfluke/welvet/fusedgpu✅
\nToken-by-token host round-trips kill decode. A fused engine keeps weights and scratch resident for Q4_0 and BinaryG128 hybrid paths.
\nEngine from Spec (AppendTokens/Reset/Close); HybridEngine (PrefillSample/DecodeSample/DecodeChunk). Specs come from model/transformer export.
\n\n\ncd welvet/examples/10-fusedgpu && source ../env.sh && go run .package main\n\nimport (\n\t\"fmt\"\n\n\t\"github.com/openfluke/welvet/fusedgpu\"\n\t\"github.com/openfluke/welvet/model/transformer\"\n)\n\nfunc main() {\n\tm, err := transformer.LoadEntity(\"model.entity\")\n\tif err != nil {\n\t\tfmt.Println(\"need an ENTITY file:\", err)\n\t\treturn\n\t}\n\tspec, err := m.ExportFusedGPUSpec()\n\tif err != nil {\n\t\tpanic(err)\n\t}\n\teng, err := fusedgpu.NewFromSpec(spec)\n\tif err != nil {\n\t\tpanic(err)\n\t}\n\tdefer eng.Close()\n\tlogits, err := eng.AppendTokens([]uint32{1, 2, 3})\n\tfmt.Println(len(logits), err)\n}need an ENTITY file: open model.entity: no such file or directorygithub.com/openfluke/welvet/layers/dense✅
\nMost FLOPs are W@x. One Dense stack owns FormatNone×34 and all quants × three backends so every composite proj shares one correctness surface — including native in-dtype SGD.
\nNew / NewConfigured[T], Forward/Backward (dispatch on Exec.Backend), Place, ApplyGradSGD (→ weights.ApplySGD on the store). SIMD forward (v1.0.3): dtype switch by MatVec strategy — DotTile / DotI8 / lowp packed / expand-once→DotTile / WireF64+DotTileF64; fused Dot* for classic Q*, k/IQ, AffinePacked. Composites (MHA, SwiGLU, CNN im2col, RNN/LSTM/Mamba, residual·sequential·parallel) reuse Dense children via syncProjExec. BackwardSIMD still DecodeRow/saxpy (not the new expand wires).
\n\n| case | dtypes | kernel |
|---|---|---|
| narrow i8 | int4, int2, ternary, binary | expand → DotI8Tile |
| uint affine | uint4/2/3/5/6, uint16…uintptr | expand once → DotTile |
| lowp packed | f16, bf16, fp8, fp4 | convert tiles → DotTile |
| expand f32 | nf4, fp6, int3/5/6 | expand once → DotTile (exact) |
| stream f64 | float64, int16/32/64, int, complex* | WireF64 → DotTileF64 |
| SelectWire | float32, int8, uint8, … | DotTile / DotI8 / affine u8 |
Full Go/C++/Rust columns: before wins Go 2 / C++ 31 / Rust 1; after\nGo 24 / C++ 9 / Rust 1. Highlight Go speedups (before→after):
\n| dtype | Go before | Go after | Go Δ | best after |
|---|---|---|---|---|
| float32 | 1023.1 | 929.5 | 1.10× | go |
| float64 | 9565.4 | 1017.4 | 9.40× | go |
| bfloat16 | 12538.9 | 5045.3 | 2.49× | cpp |
| int8 | 708.1 | 659.3 | 1.07× | cpp |
| int64 | 22271.0 | 1059.9 | 21× | go |
| int32 | 22495.4 | 964.1 | 23× | go |
| int16 | 22147.5 | 953.5 | 23× | go |
| int / complex* | ~30k | ~1.0k | ~29–31× | go |
| uint* / nf4 / fp6 / int3–6 | ~20–77k | ~3.7–10.6k | ~5–7× | go |
| uint8 | 4322.3 | 7575.0 | 0.57× | cpp |
C++ still owns int8/narrow, uint8, bf16/fp8. float32 DotTile remains Go’s race\n(929.5 vs C++ 1162 / Rust 1335). uint8 affine is a known Go follow-up.
\n\n\ncd welvet/examples/11-dense && source ../env.sh && go run .package main\n\nimport (\n\t\"fmt\"\n\n\t\"github.com/openfluke/welvet/core\"\n\t\"github.com/openfluke/welvet/layers/dense\"\n\t\"github.com/openfluke/welvet/quant\"\n)\n\nfunc main() {\n\tinit := make([]float32, 4*8)\n\tl, err := dense.NewConfigured(8, 4, core.ActivationReLU, core.DTypeFloat32, quant.FormatNone, init)\n\tif err != nil {\n\t\tpanic(err)\n\t}\n\tl.Exec.Backend = core.BackendCPUTiled\n\tx := core.NewTensor[float32](1, 8)\n\tpre, post, err := dense.Forward(l, x)\n\tif err != nil {\n\t\tpanic(err)\n\t}\n\tgIn, gW, err := dense.Backward(l, post, x, pre)\n\t_ = dense.ApplyGradSGD(l, gW, 1e-3)\n\tfmt.Println(len(gIn.Data), len(gW.Data))\n}8 32github.com/openfluke/welvet/layers/mha✅
\nTransformers need multi-head attention with masks, RoPE/ALiBi, GQA/MQA, and cross-attn — without forking MatVec for every projection.
\nQ/K/V/O are Dense children. Presets: DecoderCausal, EncoderBidirectional, CrossAttention, … Attn/RoPE ALU is host today; on-device attn shaders are still open.
\n\n\ncd welvet/examples/12-mha && source ../env.sh && go run .package main\n\nimport (\n\t\"fmt\"\n\n\t\"github.com/openfluke/welvet/core\"\n\t\"github.com/openfluke/welvet/layers/mha\"\n)\n\nfunc main() {\n\tl, err := mha.New(mha.DecoderCausal(32, 4, 4))\n\tif err != nil {\n\t\tpanic(err)\n\t}\n\tx := core.NewTensor[float32](1, 8, 32) // [batch, seq, dim]\n\tpre, post, err := mha.Forward(l, x)\n\tfmt.Println(pre != nil, len(post.Data), err)\n}true 256 <nil>github.com/openfluke/welvet/layers/swiglu✅
\nModern decoder FFNs are SiLU(gate)⊙up → down. Projections must share Dense’s quant/backend matrix.
\nGate/Up/Down Dense children; DefaultFFN(dModel); WebGPU SiLU⊙ fuse on forward.
\n\n\ncd welvet/examples/13-swiglu && source ../env.sh && go run .package main\n\nimport (\n\t\"fmt\"\n\n\t\"github.com/openfluke/welvet/core\"\n\t\"github.com/openfluke/welvet/layers/swiglu\"\n)\n\nfunc main() {\n\tl, err := swiglu.New(swiglu.DefaultFFN(64))\n\tif err != nil {\n\t\tpanic(err)\n\t}\n\tx := core.NewTensor[float32](1, 64)\n\t_, post, err := swiglu.Forward(l, x)\n\tfmt.Println(len(post.Data), err)\n}64 <nil>github.com/openfluke/welvet/layers/rmsnorm✅
\nLlama-style blocks normalize by RMS, not mean+var. Needs native fwd/bwd and WebGPU shaders.
\nPer-token RMS + γ on weights.Store; WebGPU fwd+bwd; SIMD DotTile stats + host scale.
\n\n\ncd welvet/examples/14-rmsnorm && source ../env.sh && go run .package main\n\nimport (\n\t\"fmt\"\n\n\t\"github.com/openfluke/welvet/core\"\n\t\"github.com/openfluke/welvet/layers/rmsnorm\"\n\t\"github.com/openfluke/welvet/quant\"\n)\n\nfunc main() {\n\tgamma := []float32{1, 1, 1, 1}\n\tl, err := rmsnorm.NewConfigured(rmsnorm.Config{Dim: 4}, core.DTypeFloat32, quant.FormatNone, gamma)\n\tif err != nil {\n\t\tpanic(err)\n\t}\n\tx := core.NewTensor[float32](1, 4)\n\t_, y, err := rmsnorm.Forward(l, x)\n\tfmt.Println(y.Data, err)\n}[0 0 0 0] <nil>github.com/openfluke/welvet/layers/layernorm✅
\nClassic mean+var normalization with γ/β — still required for many HF architectures.
\nWebGPU forward; backward host today. Same dtype×quant axes as other weighted layers.
\n\n\ncd welvet/examples/15-layernorm && source ../env.sh && go run .package main\n\nimport (\n\t\"fmt\"\n\n\t\"github.com/openfluke/welvet/core\"\n\t\"github.com/openfluke/welvet/layers/layernorm\"\n)\n\nfunc main() {\n\tl, err := layernorm.New(layernorm.Config{Dim: 8})\n\tif err != nil {\n\t\tpanic(err)\n\t}\n\tx := core.NewTensor[float32](1, 8)\n\t_, y, err := layernorm.Forward(l, x)\n\tfmt.Println(len(y.Data), err)\n}8 <nil>github.com/openfluke/welvet/layers/embedding✅
\nToken IDs must gather rows from a table — not a Dense MatVec — with scatter grads on backward.
\nConfig{VocabSize, EmbeddingDim, SeqLen}; table on weights.Store; host gather on SIMD/WebGPU today.
\n\n\ncd welvet/examples/16-embedding && source ../env.sh && go run .package main\n\nimport (\n\t\"fmt\"\n\n\t\"github.com/openfluke/welvet/core\"\n\t\"github.com/openfluke/welvet/layers/embedding\"\n)\n\nfunc main() {\n\tl, err := embedding.New(embedding.Config{VocabSize: 32, EmbeddingDim: 16, SeqLen: 4})\n\tif err != nil {\n\t\tpanic(err)\n\t}\n\tids := core.NewTensor[float32](1, 4)\n\tids.Data[0], ids.Data[1] = 3, 7\n\t_, y, err := embedding.Forward(l, ids)\n\tfmt.Println(y.Shape, err)\n}[1 4 16] <nil>github.com/openfluke/welvet/layers/softmax✅
\nClassification heads and attention need stable softmax variants, including sparse/Gumbel/Entmax for research paths.
\nWeightless layer; KindStandard/Temperature/Grid/Hierarchical/Gumbel/Masked/Sparse/… WebGPU covers std family; exotic kinds hard-error on GPU (no silent host).
\n\n\ncd welvet/examples/17-softmax && source ../env.sh && go run .package main\n\nimport (\n\t\"fmt\"\n\n\t\"github.com/openfluke/welvet/core\"\n\t\"github.com/openfluke/welvet/layers/softmax\"\n)\n\nfunc main() {\n\tl, err := softmax.New(softmax.Config{Dim: 4, Kind: softmax.KindStandard})\n\tif err != nil {\n\t\tpanic(err)\n\t}\n\tx := core.NewTensor[float32](1, 4)\n\tcopy(x.Data, []float32{2, 1, 0.1, -1})\n\t_, y, err := softmax.Forward(l, x)\n\tfmt.Println(y.Data, err)\n}[0.6380664 0.23473153 0.09543471 0.031767454] <nil>github.com/openfluke/welvet/layers/sequential✅
\nSome cells need an ordered Dense chain without burning grid hops.
\nDense→Dense compose, or mixed Ops via NewFromOps (Dense, SwiGLU, RMSNorm, LayerNorm). Parallel as a child is parallel.ResidualGraft / a Sandwich — Sequential cannot import parallel.
\n\nNewFromOps walks mixed children (Dense, SwiGLU, RMSNorm, LayerNorm) through the same\nfwd/bwd as a Dense chain. Parallel is not a Sequential child (import cycle) — wrap F with\nparallel.ResidualGraft or put Parallel in a Sandwich. w2a covers mixed Sequential cells;\nthis is on the v1.0 board, not a leftover “open.”
examples/18-sequential/main.go
\ncd welvet/examples/18-sequential && source ../env.sh && go run .package main\n\nimport (\n\t\"fmt\"\n\n\t\"github.com/openfluke/welvet/core\"\n\t\"github.com/openfluke/welvet/layers/sequential\"\n)\n\nfunc main() {\n\tl, err := sequential.New(sequential.Config{Dim: 16, Depth: 3})\n\tif err != nil {\n\t\tpanic(err)\n\t}\n\tx := core.NewTensor[float32](1, 16)\n\t_, y, err := sequential.Forward(l, x)\n\tfmt.Println(len(y.Data), err)\n}16 <nil>github.com/openfluke/welvet/layers/residual✅
\nSkip connections stabilize deep stacks: y = F(x) + x with correct skip grads.
\nF is Dense Dim→Dim, or mixed Ops via NewFromOps (Dense, SwiGLU, RMSNorm, LayerNorm). Parallel as F is parallel.ResidualGraft (y = F(x)+x) — Residual cannot import parallel.
\n\nNewFromOps builds F over mixed Ops. Parallel as F is\nparallel.ResidualGraft → y = F(x)+x. Skip grads still land on F and the identity path.\nNested mixed Residual is on the v1.0 board.
cd welvet/examples/19-residual && source ../env.sh && go run .package main\n\nimport (\n\t\"fmt\"\n\n\t\"github.com/openfluke/welvet/core\"\n\t\"github.com/openfluke/welvet/layers/residual\"\n)\n\nfunc main() {\n\tl, err := residual.New(residual.Config{Dim: 16, Depth: 2})\n\tif err != nil {\n\t\tpanic(err)\n\t}\n\tx := core.NewTensor[float32](1, 16)\n\t_, y, err := residual.Forward(l, x)\n\tfmt.Println(len(y.Data), err)\n}16 <nil>github.com/openfluke/welvet/layers/cnn2✅
\nConv nets must sit on the same dtype×quant×backend matrix as Dense. im2col → Dense GEMV is the intentional first cut.
\ncnn1/cnn2/cnn3: host im2col then Proj Dense. Full timed matrices. Tiled conv shaders still ⬜.
\n\n\ncd welvet/examples/20-cnn && source ../env.sh && go run .package main\n\nimport (\n\t\"fmt\"\n\n\t\"github.com/openfluke/welvet/core\"\n\t\"github.com/openfluke/welvet/layers/cnn2\"\n)\n\nfunc main() {\n\tl, err := cnn2.New(cnn2.Config{\n\t\tInChannels: 1, Filters: 4, Height: 8, Width: 8, Kernel: 3, Stride: 1,\n\t})\n\tif err != nil {\n\t\tpanic(err)\n\t}\n\tx := core.NewTensor[float32](1, 1, 8, 8)\n\t_, y, err := cnn2.Forward(l, x)\n\tfmt.Println(y.Shape, err)\n}[1 4 6 6] <nil>github.com/openfluke/welvet/layers/lstm✅
\nSequence models before transformers still need vanilla RNN and LSTM with BPTT on the shared MatVec stack.
\nIH/HH (and LSTM gates) via Dense; recurrence ALU host; device required for WebGPU path.
\n\n\ncd welvet/examples/21-rnn-lstm && source ../env.sh && go run .package main\n\nimport (\n\t\"fmt\"\n\n\t\"github.com/openfluke/welvet/core\"\n\t\"github.com/openfluke/welvet/layers/lstm\"\n\t\"github.com/openfluke/welvet/layers/rnn\"\n)\n\nfunc main() {\n\tr, _ := rnn.New(rnn.Config{InputSize: 8, HiddenSize: 16, SeqLen: 4})\n\tl, _ := lstm.New(lstm.Config{InputSize: 8, HiddenSize: 16, SeqLen: 4})\n\tx := core.NewTensor[float32](1, 4, 8)\n\t_, yr, _ := rnn.Forward(r, x)\n\t_, yl, _ := lstm.Forward(l, x)\n\tfmt.Println(len(yr.Data), len(yl.Data))\n}64 64github.com/openfluke/welvet/layers/seqmix✅
\nAttention, SSM, linear attn, and conv mixers must not be accidental forks of mha. Naming the contract keeps packages honest.
\nKindAttention | KindSSM | KindLinearAttn | KindConvMix and Contract{Kind,DModel,MaxT}. No compute here.
\n\n\ncd welvet/examples/22-seqmix && source ../env.sh && go run .package main\n\nimport (\n\t\"fmt\"\n\n\t\"github.com/openfluke/welvet/layers/seqmix\"\n)\n\nfunc main() {\n\tc := seqmix.Contract{Kind: seqmix.KindAttention, DModel: 64, MaxT: 2048}\n\tfmt.Println(c.Kind.String()) // attention\n}attentiongithub.com/openfluke/welvet/layers/gdn✅
\nLinear attention / decode-first mixers (Gated DeltaNet) need a first-class package under KindLinearAttn.
\nExec CPU/SIMD/WebGPU; ForwardDecode; truncated BPTT. Timed matrix is Float32-primary: PermutationOK is f32 × FormatNone/BinaryPacked × CPU/SIMD/WebGPU; other cells GAP (declared), not FAIL.
\n\nPermutationOK is Float32 × FormatNone/BinaryPacked × CPU/SIMD/WebGPU only.\nExotic dtypes and other packed formats are declared GAP in the timed matrix. Truncated BPTT;\nblobs are not on the Dense Store dtype axis. Do not read GAP as “GDN is broken.”cd welvet/examples/23-gdn && source ../env.sh && go run .package main\n\nimport (\n\t\"fmt\"\n\n\t\"github.com/openfluke/welvet/layers/gdn\"\n)\n\nfunc main() {\n\tl, err := gdn.New(gdn.Config{\n\t\tHiddenSize: 64, NumKeyHeads: 4, NumValueHeads: 4,\n\t\tKeyHeadDim: 16, ValueHeadDim: 16, ConvKernel: 4,\n\t})\n\tif err != nil {\n\t\tfmt.Println(\"config rejected:\", err)\n\t\treturn\n\t}\n\tl.Reset()\n\tx := make([]float32, 64)\n\ty := make([]float32, 64)\n\tfmt.Println(l.ForwardDecode(x, y))\n}<nil>github.com/openfluke/welvet/layers/mamba✅
\nSSM mixers (KindSSM) are not MHA clones — they need their own selective-scan path.
\nInProj → softplus(Δ) scan → OutProj. Full timed matrix + train grids (scan ALU host).
\n\n\ncd welvet/examples/24-mamba && source ../env.sh && go run .package main\n\nimport (\n\t\"fmt\"\n\n\t\"github.com/openfluke/welvet/core\"\n\t\"github.com/openfluke/welvet/layers/mamba\"\n)\n\nfunc main() {\n\tl, err := mamba.New(mamba.Config{DModel: 32, DState: 16, Expand: 2, SeqLen: 8})\n\tif err != nil {\n\t\tfmt.Println(err)\n\t\treturn\n\t}\n\tx := core.NewTensor[float32](1, 8, 32)\n\t_, y, err := mamba.Forward(l, x)\n\tfmt.Println(y != nil, err)\n}true <nil>github.com/openfluke/welvet/layers/convt1✅
\nGenerators and U-Nets need transposed convolution as a peer of CNN, on the same Dense proj surface.
\nScatter upsample + Dense Proj. Full timed matrix + train grids (tiled transpose shaders still §12).
\n\n\ncd welvet/examples/25-convt && source ../env.sh && go run .package main\n\nimport (\n\t\"fmt\"\n\n\t\"github.com/openfluke/welvet/core\"\n\t\"github.com/openfluke/welvet/layers/convt1\"\n)\n\nfunc main() {\n\tl, err := convt1.New(convt1.Config{InChannels: 4, Filters: 2, SeqLen: 8, Kernel: 3, Stride: 2})\n\tif err != nil {\n\t\tpanic(err)\n\t}\n\tx := core.NewTensor[float32](1, 4, 8)\n\t_, y, err := convt1.Forward(l, x)\n\tfmt.Println(y.Shape, err)\n}[1 2 17] <nil>github.com/openfluke/welvet/layers/kmeans✅
\nSoft clustering as a differentiable layer lets topology experiments sit inside the same train loop.
\nCenters on Dense (K×FeatureDim); soft assignment outputs. Full timed matrix + train grids.
\n\n\ncd welvet/examples/26-kmeans && source ../env.sh && go run .package main\n\nimport (\n\t\"fmt\"\n\n\t\"github.com/openfluke/welvet/core\"\n\t\"github.com/openfluke/welvet/layers/kmeans\"\n)\n\nfunc main() {\n\tl, err := kmeans.New(kmeans.Config{NumClusters: 4, FeatureDim: 8})\n\tif err != nil {\n\t\tpanic(err)\n\t}\n\tx := core.NewTensor[float32](1, 8)\n\t_, y, err := kmeans.Forward(l, x)\n\tfmt.Println(y.Shape, err)\n}[1 4] <nil>github.com/openfluke/welvet/layers/parallel✅
\nMixture-of-experts and multi-path cells need concat/add/avg/filter combines. Cameral graphs need sibling hemispheres that share input, merge outputs, and optionally train under distinct modes.
\nParallel branches + Stack sandwiches. Hemispheres / Bicameral / Sandwich build nested multi-cameral nets. SetBranchModes + TrainStackMSE let each hemi use a different TrainMode (all 29 named updates — BP, Tween, Split, FastProxy, Sparse, Step*, Mesh*, …).
\ninput x\n │\n ▼\n Dense stem\n │\n ├──────────┐\n ▼ ▼\n Hemi L Hemi R ← own Dense weights\n StepBP TweenChain ← BranchModes\n │ │\n └──── add ─┘\n │\n ▼\n Dense head → ŷ → MSE
Hemispheres / HemispheresFrom — n twins (Dense or mixed Ops), merged by add/avg/concat/filterBicameral / Sandwich / BicameralFrom — stem → Parallel → head StackSetBranchModes(...) — stamp per-hemi TrainModeTrainStackMSE — forward → MSE → per-branch update (honours BranchModes)ResidualGraft — skip around a Parallel F without Residual importing this packageWriteCameralFile / LoadCameral — cameral .entity (in model/entity)CamSyncConfig / SetCamSync / SyncNow — soft/hard inter-cameral weight blend + cross-layer same-shape pairs → §70Uniform Bi/Tri/Quad can train via Grid training.Step (one mode for the net).\nMix jobs need the Stack path so each cameral actually uses its own mode.
Why hemispheres exist, and how AAI Lucy benches use them:\n§68. Inter-cameral weight sync (crazy cross-mesh same-shape blends):\n§70. All 29 named updates and their equations:\n§67.
\n\n\ncd welvet/examples/27-parallel && source ../env.sh && go run .package main\n\nimport (\n\t\"fmt\"\n\n\t\"github.com/openfluke/welvet/core\"\n\t\"github.com/openfluke/welvet/layers/parallel\"\n\t\"github.com/openfluke/welvet/quant\"\n)\n\nfunc main() {\n\t// Dense stem → 2 hemispheres (add) → Dense head\n\ts, err := parallel.Bicameral(8, 16, 1, core.ActivationLeakyReLU,\n\t\tcore.DTypeFloat32, quant.FormatNone)\n\tif err != nil {\n\t\tpanic(err)\n\t}\n\t// Mix: left hemi StepBP, right hemi StepTweenChain\n\themi := s.Children[1].(*parallel.Layer)\n\themi.SetBranchModes(parallel.ModeStepBP, parallel.ModeStepTweenChain)\n\n\tx := core.NewTensor[float32](1, 8)\n\tt := core.NewTensor[float32](1, 1)\n\tt.Data[0] = 0.5\n\tloss, err := parallel.TrainStackMSE(s, x, t, parallel.ModeStepBP, 0.01)\n\tfmt.Println(\"loss\", loss, \"err\", err)\n}loss 0 err <nil>github.com/openfluke/welvet/layers/metacognition✅
\nObserved layers can apply heuristic stability rules (gate/scale/reset) without dtype morph/QAT.
\nWraps Dense + DefaultStabilityRules(); Stats exposed. Full timed matrix + train grids.
\n\n\nexamples/28-metacognition/main.go
\ncd welvet/examples/28-metacognition && source ../env.sh && go run .package main\n\nimport (\n\t\"fmt\"\n\n\t\"github.com/openfluke/welvet/core\"\n\t\"github.com/openfluke/welvet/layers/metacognition\"\n)\n\nfunc main() {\n\tl, err := metacognition.New(metacognition.Config{\n\t\tDim: 16, Rules: metacognition.DefaultStabilityRules(),\n\t})\n\tif err != nil {\n\t\tpanic(err)\n\t}\n\tx := core.NewTensor[float32](1, 16)\n\t_, y, err := metacognition.Forward(l, x)\n\tfmt.Println(len(y.Data), err)\n}16 <nil>github.com/openfluke/welvet/runtime/forward✅
\nA grid of heterogeneous ops needs one walker that dispatches by concrete type and fails loudly on unknowns.
\nForward[T](grid, input) → Result tape; Cell[T] for single-cell dispatch. Covers Dense…Residual + extended wired ops.
\n\n\ncd welvet/examples/29-forward && source ../env.sh && go run .package main\n\nimport (\n\t\"fmt\"\n\n\t\"github.com/openfluke/welvet/architecture\"\n\t\"github.com/openfluke/welvet/core\"\n\t\"github.com/openfluke/welvet/layers/dense\"\n\t\"github.com/openfluke/welvet/runtime/forward\"\n)\n\nfunc main() {\n\tg := architecture.NewGrid(1, 1, 1, 1)\n\tl, _ := dense.New(8, 8, core.ActivationLinear, core.DTypeFloat32)\n\t_ = dense.Place(g, 0, 0, 0, 0, l)\n\tres, err := forward.Forward(g, core.NewTensor[float32](1, 8))\n\tfmt.Println(res != nil, err)\n}true <nil>github.com/openfluke/welvet/runtime/backward✅
\nTraining needs a reverse tape over the same ops forward used — no separate graph framework.
\nBackward[T](fwdResult, gradOut) walks the tape and returns per-op weight grads.
\n\n\ncd welvet/examples/30-backward && source ../env.sh && go run .package main\n\nimport (\n\t\"fmt\"\n\n\t\"github.com/openfluke/welvet/architecture\"\n\t\"github.com/openfluke/welvet/core\"\n\t\"github.com/openfluke/welvet/layers/dense\"\n\t\"github.com/openfluke/welvet/runtime/backward\"\n\t\"github.com/openfluke/welvet/runtime/forward\"\n)\n\nfunc main() {\n\tg := architecture.NewGrid(1, 1, 1, 1)\n\tl, _ := dense.New(4, 4, core.ActivationLinear, core.DTypeFloat32)\n\t_ = dense.Place(g, 0, 0, 0, 0, l)\n\tfwd, _ := forward.Forward(g, core.NewTensor[float32](1, 4))\n\tbwd, err := backward.Backward(fwd, core.NewTensor[float32](1, 4))\n\tfmt.Println(bwd != nil, err)\n}true <nil>github.com/openfluke/welvet/runtime/training✅
\nSuites and small nets need MSE+SGD and tween hooks without inventing an external trainer or a retained float32 master beside storage.
\nMSE/MSEGrad, SGD, Step, ApplyTween/StepTween, StepMesh. Layer-agnostic ApplyGradSGD dispatch (Dense…Mamba/GDN/…). FormatNone: in-dtype ApplySGD; packed: unpack→update→re-Pack. No QAT dual path — storage dtype/format is truth after every step. Sandwich credit (Split / FastProxy / Sparse / …) lives in layers/parallel TrainMode — see §67.
\n\nGrid training.Step is MSE + backprop on a volumetric tape. Cameral Mix and the 29 named\ncredit updates use parallel.TrainStackMSE (and Mesh* schedulers) so hemispheres can disagree.\nEquations, rival metric (hard Acc vs StepBP), and Lucy Score honesty: §67.
cd welvet/examples/31-training && source ../env.sh && go run .package main\n\nimport (\n\t\"fmt\"\n\n\t\"github.com/openfluke/welvet/architecture\"\n\t\"github.com/openfluke/welvet/core\"\n\t\"github.com/openfluke/welvet/layers/dense\"\n\t\"github.com/openfluke/welvet/runtime/forward\"\n\t\"github.com/openfluke/welvet/runtime/training\"\n)\n\nfunc main() {\n\tg := architecture.NewGrid(1, 1, 1, 1)\n\tl, _ := dense.New(4, 4, core.ActivationLinear, core.DTypeFloat32)\n\t_ = dense.Place(g, 0, 0, 0, 0, l)\n\tfwd, _ := forward.Forward(g, core.NewTensor[float32](1, 4))\n\tloss, err := training.Step(fwd, core.NewTensor[float32](1, 4), 1e-2)\n\tfmt.Println(loss, err)\n}0 <nil>github.com/openfluke/welvet/runtime/step✅
\nSpatial feedback (remote links) needs a discrete-time mesh where every cell updates from a double buffer — different from a decoder wavefront. Cross-numeric train also needs the same mesh with weight DType ⊥ activation Tensor[T].
\nState[T], StepForward/StepBackward/StepApplyTween / StepMesh across the grid for all wired Ops × dtype × quant × CPU/SIMD. W2A Cross-Numeric Train: polyops.AllKinds() × weight dtype × act host (smoke ~21×7×5 ≈ 735; full ~21×34×15 ≈ 10.7k) — asserts no retained f32 master after StepMesh. Stack Step* is a different clock: IsLineStep / TrainLine (1D pipe) — see §67.
\n\nThis package is the volumetric mesh: every grid cell hops from a double buffer. MeshBP / MeshTween / MeshTweenChain call it.
\nStack Step* (IsLineStep) is a 1D systolic pipe on a Sandwich — not this mesh.\nOne sample enters child 0 per TrainStackMSE tick; every in-flight sample advances one layer; the output/train event is the sample that entered depth ticks ago. Fill ticks do not update. Serve stays a full ForwardStack. Mesh* still needs a Grid. Named updates: §67.
cd welvet/examples/32-step && source ../env.sh && go run .package main\n\nimport (\n\t\"fmt\"\n\n\t\"github.com/openfluke/welvet/architecture\"\n\t\"github.com/openfluke/welvet/core\"\n\t\"github.com/openfluke/welvet/layers/dense\"\n\t\"github.com/openfluke/welvet/runtime/step\"\n)\n\nfunc main() {\n\tg := architecture.NewGrid(1, 1, 1, 1)\n\tl, _ := dense.New(4, 4, core.ActivationLinear, core.DTypeFloat32)\n\t_ = dense.Place(g, 0, 0, 0, 0, l)\n\tst := step.New[float32](g)\n\t_, err := step.StepForward(g, st, false)\n\tfmt.Println(err)\n}<nil>github.com/openfluke/welvet/systems/dna✅
\nQuant and train must be measurable as topology/weight fingerprints — DNA detects logic shifts.
\nExtractDNA, CompareNetworks, CosineSimilarity, FlattenOp / CollectStores across dtype×quant.
\n\n\ncd welvet/examples/33-dna && source ../env.sh && go run .package main\n\nimport (\n\t\"fmt\"\n\n\t\"github.com/openfluke/welvet/architecture\"\n\t\"github.com/openfluke/welvet/core\"\n\t\"github.com/openfluke/welvet/layers/dense\"\n\t\"github.com/openfluke/welvet/systems/dna\"\n)\n\nfunc main() {\n\tg := architecture.NewGrid(1, 1, 1, 1)\n\tl, _ := dense.New(4, 4, core.ActivationLinear, core.DTypeFloat32)\n\t_ = dense.Place(g, 0, 0, 0, 0, l)\n\ta := dna.ExtractDNA(g)\n\tfmt.Println(dna.CompareNetworks(a, a))\n}{1 map[0,0,0,0:1] []}github.com/openfluke/welvet/systems/evolution✅
\nTopology search and weight crossover need first-class splice + NEAT on CPU-resident grids.
\nSpliceDNA, NEATMutate, NewNEATPopulation, CloneGrid — dtype/quant preserved via SetFromF32.
\n\n\ncd welvet/examples/34-evolution && source ../env.sh && go run .package main\n\nimport (\n\t\"fmt\"\n\n\t\"github.com/openfluke/welvet/architecture\"\n\t\"github.com/openfluke/welvet/systems/evolution\"\n)\n\nfunc main() {\n\ta, b := architecture.NewGrid(1, 1, 1, 1), architecture.NewGrid(1, 1, 1, 1)\n\tchild, err := evolution.SpliceDNA(a, b, evolution.DefaultSpliceConfig())\n\tif err != nil {\n\t\tfmt.Println(err)\n\t\treturn\n\t}\n\t_, err = evolution.NEATMutate(child, evolution.DefaultNEATConfig(16))\n\tfmt.Println(err)\n}<nil>github.com/openfluke/welvet/systems/tween✅
\nTarget propagation (chain-rule or Hebbian layerwise gaps) is an alternative credit-assignment path. Not the same package as TrainMode Tween / TweenChain on a Sandwich (those are layers/parallel — §67).
\nNewState, Forward, BackwardChainRule / BackwardLayerwise, ApplyGaps; SIMD DotTile/Saxpy budgets.
\n\n\ncd welvet/examples/35-tween && source ../env.sh && go run .package main\n\nimport (\n\t\"fmt\"\n\n\t\"github.com/openfluke/welvet/architecture\"\n\t\"github.com/openfluke/welvet/core\"\n\t\"github.com/openfluke/welvet/systems/tween\"\n)\n\nfunc main() {\n\tg := architecture.NewGrid(1, 1, 1, 1)\n\tst := tween.NewState[float32](g, tween.DefaultConfig())\n\t_, err := tween.Forward(g, st, core.NewTensor[float32](1, 4))\n\tfmt.Println(err)\n}forward: no op at {0 0 0 0} (type Dense)github.com/openfluke/welvet/systems/tanhi✅
\nTraining visualization must never block the engine — best-effort UDP JSON-lines to a HUD.
\nTANHI = Tensor Activation Network Holographic Interface. Sparse non-blocking JSON-line UDP events for per-layer forward/backward HUD visualization (timing, dtypes, shapes, routing links, parallel branches/combine). v1.1.1: grid BP, cameral stack/parallel train (all mode families), step mesh; ConfigureGrid, SetTanhi, topology bridge. ConfigFromGrid, Emit/EmitSweep, DefaultUDPPort (17481). SoulGlitch-style consumers.
cd welvet/examples/36-tanhi && source ../env.sh && go run .package main\n\nimport (\n\t\"github.com/openfluke/welvet/architecture\"\n\t\"github.com/openfluke/welvet/systems/tanhi\"\n)\n\nfunc main() {\n\tg := architecture.NewGrid(1, 1, 1, 1)\n\tcfg := tanhi.ConfigFromGrid(g)\n\ttanhi.EmitSweep(cfg, \"epoch-0\") // non-blocking best-effort\n}(no output)github.com/openfluke/welvet/systems/telemetry✅
\nStatic structural blueprints (sizes, op kinds) differ from live TANHI events.
\nExtractNetworkBlueprint, ExtractLayerTelemetry for introspection/UIs.
\n\n\ncd welvet/examples/37-telemetry && source ../env.sh && go run .package main\n\nimport (\n\t\"fmt\"\n\n\t\"github.com/openfluke/welvet/architecture\"\n\t\"github.com/openfluke/welvet/systems/telemetry\"\n)\n\nfunc main() {\n\tg := architecture.NewGrid(1, 1, 1, 1)\n\tbp := telemetry.ExtractNetworkBlueprint(g, \"demo\")\n\tfmt.Printf(\"%+v\\n\", bp)\n}{ID:demo TotalLayers:1 TotalParams:0 Layers:[{Z:0 Y:0 X:0 L:0 Type:Dense Activation:ReLU Parameters:0 InputShape:[] OutputShape:[] Branches:[] CombineMode:}]}github.com/openfluke/welvet/model/entity✅
\nHF safetensors are awkward for native topology + packed weights. ENTITY is the Welvet checkpoint.
\nOpen/Inspect/IsEntity, LoadBlob/LoadQuantBlob, PackFromHF/ImportFromHF, WriteTransformerFile, SerializeNetwork, WriteCameralFile / LoadCameral (sandwich Stack + TrainMode).
\n\n\ncd welvet/examples/38-entity && source ../env.sh && go run .package main\n\nimport (\n\t\"fmt\"\n\n\t\"github.com/openfluke/welvet/model/entity\"\n)\n\nfunc main() {\n\tinfo, err := entity.Inspect(\"model.entity\")\n\tif err != nil {\n\t\tfmt.Println(err)\n\t\treturn\n\t}\n\tef, err := entity.Open(\"model.entity\")\n\tif err != nil {\n\t\tpanic(err)\n\t}\n\tdefer ef.Close()\n\tfmt.Println(info, ef.HasTokenizerBlob())\n}open model.entity: no such file or directorygithub.com/openfluke/welvet/model/hf✅
\nImport starts with probing HF/MLX layouts before packing ENTITY.
\nInspectSnapshot, DetectArchitecture, safetensors/MLX loaders, Qwen3.5 hybrid helpers.
\n\n\ncd welvet/examples/39-hf && source ../env.sh && go run .package main\n\nimport (\n\t\"fmt\"\n\n\t\"github.com/openfluke/welvet/model/hf\"\n)\n\nfunc main() {\n\tinfo, err := hf.InspectSnapshot(\"/path/to/hf-snapshot\")\n\tif err != nil {\n\t\tfmt.Println(err)\n\t\treturn\n\t}\n\tfmt.Println(info)\n}hf: config.json: stat /path/to/hf-snapshot/config.json: no such file or directorygithub.com/openfluke/welvet/model/tokenizer✅
\nGenerate needs encode/decode of HF tokenizer.json without pulling Python.
\nLoadTokenizer, Encode/Decode, LoadForEntity.
\n\n\ncd welvet/examples/40-tokenizer && source ../env.sh && go run .package main\n\nimport (\n\t\"fmt\"\n\n\t\"github.com/openfluke/welvet/model/tokenizer\"\n)\n\nfunc main() {\n\ttok, err := tokenizer.LoadTokenizer(\"tokenizer.json\")\n\tif err != nil {\n\t\tfmt.Println(err)\n\t\treturn\n\t}\n\tids := tok.Encode(\"hello welvet\", true)\n\tfmt.Println(ids, tok.Decode(ids, true))\n}failed to read tokenizer file: open tokenizer.json: no such file or directorygithub.com/openfluke/welvet/model/sampling✅
\nLogits → token ID needs ArgMax, TopK+temperature, penalties, and chat hygiene in one place.
\nArgMax, SampleTopK, ApplyRepetitionPenalty, BanIDs, SanitizeChatReply.
\n\n\ncd welvet/examples/41-sampling && source ../env.sh && go run .package main\n\nimport (\n\t\"fmt\"\n\n\t\"github.com/openfluke/welvet/model/sampling\"\n)\n\nfunc main() {\n\tlogits := []float32{0.1, 2.4, 0.3, -1}\n\tfmt.Println(sampling.ArgMax(logits))\n\tfmt.Println(sampling.SampleTopK(logits, 2, 0.8, true))\n}1\n1github.com/openfluke/welvet/model/transformer✅
\nENTITY packs must run as Llama-style decoders with KV cache, profiles (SIMD/WebGPU/fused), and chat templates.
\nLoadEntity → Model; Generate; ApplyExec profiles; ExportFusedGPUSpec / hybrid sync.
\n\n\nexamples/42-transformer/main.go
\ncd welvet/examples/42-transformer && source ../env.sh && go run .package main\n\nimport (\n\t\"fmt\"\n\n\t\"github.com/openfluke/welvet/model/tokenizer\"\n\t\"github.com/openfluke/welvet/model/transformer\"\n)\n\nfunc main() {\n\tm, err := transformer.LoadEntity(\"model.entity\")\n\tif err != nil {\n\t\tfmt.Println(err)\n\t\treturn\n\t}\n\ttok, err := tokenizer.LoadTokenizer(\"tokenizer.json\")\n\tif err != nil {\n\t\tpanic(err)\n\t}\n\tif err := m.ApplyExec(transformer.ProfileSIMDMultiCore()); err != nil {\n\t\tpanic(err)\n\t}\n\ttext, _, err := m.Generate(\n\t\ttok.Encode, tok.Decode, nil,\n\t\t\"You are helpful.\", \"Say hi.\",\n\t\ttransformer.GenOptions{MaxTokens: 32, Silent: true},\n\t)\n\tfmt.Println(text, err)\n}open model.entity: no such file or directorygithub.com/openfluke/welvet/apps/…🚧
\nProducts must not pollute engine packages. Octo is the model shell; flux2/mosstts are domain apps. Lucy races (AAI test41 / test48 / test50) are benches, not Welvet packages.
\nocto (own module): download/convert/chat, see §44. flux2: MMDiT image. mosstts: Speak/SpeakToFile pipeline. AAI sandwiches race TrainMode on toys — see §68. Off the v1.0 engine board.
\n\nThe engine never imports apps. AAI is a separate Lucy harness that replaces\ngithub.com/openfluke/welvet. It is how we talk about cameral Mix and credit modes with numbers\n(hard Acc, SoftAcc, Score) instead of slogans. Public API stays in layers/parallel + lucy.
cd welvet/examples/43-apps && source ../env.sh && go run .package main\n\nimport \"fmt\"\n\nfunc main() {\n\tfmt.Println(\"Apps import the engine — the engine never imports apps.\")\n\tfmt.Println(\" cd apps/octo && go run .\")\n\tfmt.Println(\" cd apps/flux2 && go run .\")\n\tfmt.Println(\" cd apps/mosstts && go run .\")\n}Apps import the engine — the engine never imports apps.\n cd apps/octo && go run .\n cd apps/flux2 && go run .\n cd apps/mosstts && go run .cd apps/octo && go run . # when module wiredgithub.com/openfluke/welvet/apps/octo✅ runs
\nA model is only useful with a shell around it: pull weights from Hugging Face, convert them to a Welvet .entity, then chat, serve, or benchmark. Octo is that shell, kept in its own module so the engine never depends on an app.
Subcommands cover the whole loop: hub download/ensure repos, convert pack to a single .entity, interactive run/serve/chat, plus image and speech menus. Its bench harness sweeps every quant format across a CPU Plan 9 SIMD fused profile and a WebGPU fused profile.
| Quant | GPU-fused tok/s | SIMD-fused tok/s | Entity MB |
|---|---|---|---|
| Q8_0 | 108.2 | 33.2 | 252.5 |
| Q6_K | 73.5 | 4.9 | 216.4 |
| Q5_K | 71.7 | 6.3 | 208.4 |
| Q4_K | 73.4 | 7.8 | 192.3 |
| Q4_0 | 112.1 | 36.8 | 188.3 |
| IQ4_XS | 105.5 | 7.6 | 204.4 |
| IQ4_NL | 84.1 | 8.8 | 188.3 |
| TernaryPacked | 106.9 | 18.9 | 172.3 |
| BinaryPacked | 88.3 | 21.0 | 140.2 |
Q8_0 / gpu I'm a helpful assistant. I'm here to help you with your inquiries.\nQ4_0 / gpu I'm ready to help. What can I help you with?\nQ6_K / gpu 2 + 2 is 4.\nQ6_K / simd I'm a helpful assistant.\nIQ2_XXS/gpu ,))hidAPAAPAAPAAPAAPA... (sub-2-bit degrades)\n\n\n
cd welvet/examples/44-octo && source ../env.sh && go run .package main\n\nimport \"fmt\"\n\nfunc main() {\n\tfmt.Println(\"Octo is its own module — the engine never imports it.\")\n\tfmt.Println(\" cd apps/octo && go run .\")\n\tfmt.Println(\" download -> convert -> .entity -> chat / serve / bench\")\n}Octo is its own module — the engine never imports it.\n cd apps/octo && go run .\n download -> convert -> .entity -> chat / serve / benchcd apps/octo && go run .github.com/openfluke/welvet/stub/seed🚧
\nShip topology recipes (layer seeds → He-init) without weight blobs.
\nSeedFrom/From, RNG, InitStoreHe, Dense/MHA/SwiGLU/CNN infinite manifests.
\n\n\ncd welvet/examples/45-stub-seed && source ../env.sh && go run .package main\n\nimport (\n\t\"fmt\"\n\n\t\"github.com/openfluke/welvet/stub/seed\"\n)\n\nfunc main() {\n\ts := seed.From(\"demo-net\", 0)\n\trng := seed.New(s)\n\tfmt.Println(rng.NormFloat64())\n}0.4118750632366808github.com/openfluke/welvet/stub/serialization🚧
\nVolumetric grids need JSON/ENTITY persist beyond transformer packs.
\nSerializeEntity/LoadEntity, SerializeGrid/GridFromSpec — native FormatNone bytes; packed via wire.
\n\n\nexamples/46-stub-serialization/main.go
\ncd welvet/examples/46-stub-serialization && source ../env.sh && go run .package main\n\nimport (\n\t\"fmt\"\n\n\t\"github.com/openfluke/welvet/architecture\"\n\t\"github.com/openfluke/welvet/stub/serialization\"\n)\n\nfunc main() {\n\tg := architecture.NewGrid(1, 1, 1, 1)\n\traw, err := serialization.SerializeGrid(g)\n\tfmt.Println(len(raw), err)\n}101 <nil>github.com/openfluke/welvet/stub/memory🚧
\nHF→ENTITY and GPU upload need footprint accounting and optional history charts.
\nFromGrid, Footprint, InitScavenger, ReleaseTransient; WELVET_MEMORY_HISTORY=1.
\n\n\nexamples/47-stub-memory/main.go
\ncd welvet/examples/47-stub-memory && source ../env.sh && go run .package main\n\nimport (\n\t\"fmt\"\n\n\t\"github.com/openfluke/welvet/architecture\"\n\t\"github.com/openfluke/welvet/stub/memory\"\n)\n\nfunc main() {\n\tg := architecture.NewGrid(1, 1, 1, 1)\n\tfp := memory.FromGrid(g)\n\tfmt.Printf(\"%+v\\n\", fp)\n}{HostWeightsMB:0 GPUWeightsMB:0 GPUKVMB:0}github.com/openfluke/welvet/stub/donate🚧
\nLAN donors should accept framed JSON jobs without embedding HTTP in the engine.
\nu32-LE + JSON frames; ServeTCP/Dial; model_push vs local_lm. v0 workers echo.
\n\n\nexamples/48-stub-donate/main.go
\ncd welvet/examples/48-stub-donate && source ../env.sh && go run .package main\n\nimport (\n\t\"fmt\"\n\n\t\"github.com/openfluke/welvet/stub/donate\"\n)\n\nfunc main() {\n\tfmt.Println(\"default port\", donate.DefaultPort)\n\t// ln, err := donate.ServeTCP(donate.ServerOptions{Addr: \":17001\", Mode: donate.ServerLocalLM})\n}default port 17001github.com/openfluke/welvet/stub/fountain🚧
\nRecover specialist weight blobs over lossy links via LT fountain codes, then ensemble.
\nNewLTEncoder/Decoder, RecoverWeightBlobs, Neural/DenseFactory helpers.
\n\n\nexamples/49-stub-fountain/main.go
\ncd welvet/examples/49-stub-fountain && source ../env.sh && go run .package main\n\nimport (\n\t\"fmt\"\n\n\t\"github.com/openfluke/welvet/stub/fountain\"\n)\n\nfunc main() {\n\tsrc := [][]byte{make([]byte, 64), make([]byte, 64), make([]byte, 64)}\n\tenc, err := fountain.NewLTEncoder(src, 42)\n\tfmt.Printf(\"%T %v\\n\", enc, err)\n}*fountain.LTEncoder <nil>github.com/openfluke/welvet/stub/hardware🚧
\nDispatchers and UIs need a portable host audit (OS/CPU/RAM/GPU).
\nAudit() → SystemAudit with linux /proc or platform fallbacks.
\n\n\nexamples/50-stub-hardware/main.go
\ncd welvet/examples/50-stub-hardware && source ../env.sh && go run .package main\n\nimport (\n\t\"fmt\"\n\n\t\"github.com/openfluke/welvet/stub/hardware\"\n)\n\nfunc main() {\n\ta := hardware.Audit()\n\tfmt.Printf(\"%+v\\n\", a.CPU)\n\tfmt.Println(hardware.Description())\n}{Model:Intel(R) Core(TM) i5-10400 CPU @ 2.90GHz Logical:12 GOMAXPROCS:12}\nOS: linux | CPU: Intel(R) Core(TM) i5-10400 CPU @ 2.90GHz (12) | RAM: 31.12 GB | GPU: Unknown GPU (0.0 GB)github.com/openfluke/welvet/stub/accel⬜
\nVendor accelerators (Intel NPU, Qualcomm QNN, Apple Metal) will plug beside WebGPU — not replace it.
\nPackage is doc.go only today (⬜). No symbols yet; reserved for plugin loader design.
\n\n\nexamples/51-stub-accel/main.go
\ncd welvet/examples/51-stub-accel && source ../env.sh && go run .package main\n\n// import \"github.com/openfluke/welvet/stub/accel\"\n// No exported API yet — see package doc.go for intent.\n\nfunc main() {}(no output)github.com/openfluke/welvet/stub/clustering🚧
\nOffline clustering helpers on tensors without inventing a second math stack.
\nKMeansCluster, HierarchicalGroup, distance/silhouette helpers.
\n\n\nexamples/52-stub-clustering/main.go
\ncd welvet/examples/52-stub-clustering && source ../env.sh && go run .package main\n\nimport (\n\t\"fmt\"\n\n\t\"github.com/openfluke/welvet/core\"\n\t\"github.com/openfluke/welvet/stub/clustering\"\n)\n\nfunc main() {\n\tpts := []*core.Tensor[float32]{\n\t\tcore.NewTensor[float32](2),\n\t\tcore.NewTensor[float32](2),\n\t}\n\tpts[0].Data[0], pts[0].Data[1] = 0, 0\n\tpts[1].Data[0], pts[1].Data[1] = 1, 1\n\tcent, assign := clustering.KMeansCluster(pts, 2, 10, false)\n\tfmt.Println(len(cent), assign)\n}2 [0 1]github.com/openfluke/welvet/stub/ensemble🚧
\nCombine multiple model votes and find complementary specialists.
\nMajorityVote, FindComplementaryMatches, PerformanceSimilarity.
\n\n\nexamples/53-stub-ensemble/main.go
\ncd welvet/examples/53-stub-ensemble && source ../env.sh && go run .package main\n\nimport (\n\t\"fmt\"\n\n\t\"github.com/openfluke/welvet/stub/ensemble\"\n)\n\nfunc main() {\n\tvotes := [][]int{{0, 1, 1}, {0, 0, 1}, {0, 1, 0}}\n\tfmt.Println(ensemble.MajorityVote(votes))\n}[0 1 1]github.com/openfluke/welvet/stub/evaluation🚧
\nBenchmark grids through runtime/forward with deviation metrics.
\nEvaluateNetwork, MultiNetworkEvaluation, DeviationMetrics.
\n\n\nexamples/54-stub-evaluation/main.go
\ncd welvet/examples/54-stub-evaluation && source ../env.sh && go run .package main\n\nimport (\n\t\"fmt\"\n\n\t\"github.com/openfluke/welvet/architecture\"\n\t\"github.com/openfluke/welvet/core\"\n\t\"github.com/openfluke/welvet/stub/evaluation\"\n)\n\nfunc main() {\n\tg := architecture.NewGrid(1, 1, 1, 1)\n\tin := []*core.Tensor[float32]{core.NewTensor[float32](1, 4)}\n\trep, err := evaluation.EvaluateNetwork(g, in, []float64{0, 0, 0, 0})\n\tfmt.Println(rep, err)\n}<nil> evaluation: length mismatch inputs=1 expected=4github.com/openfluke/welvet/stub/grafting🚧
\nMerge grids into Parallel/Residual structures for topology experiments.
\nGraftGrids, ResidualGraft, GraftToGrid — Dense branches in v0.
\n\n\nexamples/55-stub-grafting/main.go
\ncd welvet/examples/55-stub-grafting && source ../env.sh && go run .package main\n\nimport (\n\t\"fmt\"\n\n\t\"github.com/openfluke/welvet/architecture\"\n\t\"github.com/openfluke/welvet/layers/parallel\"\n\t\"github.com/openfluke/welvet/stub/grafting\"\n)\n\nfunc main() {\n\ta, b := architecture.NewGrid(1, 1, 1, 1), architecture.NewGrid(1, 1, 1, 1)\n\tout, err := grafting.GraftGrids([]*architecture.Grid{a, b}, parallel.CombineConcat)\n\tfmt.Println(out != nil, err)\n}false grafting: no valid branchesgithub.com/openfluke/welvet/stub/grouping🚧
\nDetect layer archetypes from safetensor-style names before mounting.
\nGroupRelatedTensors, DetectMHA/DetectSwiGLU/DetectRMSNorm → ArchetypeHint.
\n\n\nexamples/56-stub-grouping/main.go
\ncd welvet/examples/56-stub-grouping && source ../env.sh && go run .package main\n\nimport (\n\t\"fmt\"\n\n\t\"github.com/openfluke/welvet/stub/grouping\"\n)\n\nfunc main() {\n\ttensors := []grouping.DetectedTensor{\n\t\t{Name: \"model.layers.0.self_attn.q_proj.weight\"},\n\t\t{Name: \"model.layers.0.self_attn.k_proj.weight\"},\n\t\t{Name: \"model.layers.0.self_attn.v_proj.weight\"},\n\t\t{Name: \"model.layers.0.self_attn.o_proj.weight\"},\n\t}\n\tok, hint := grouping.DetectMHA(\"block0\", tensors, 64, 4)\n\tfmt.Println(ok, hint)\n}true {MultiHeadAttention 64 4}github.com/openfluke/welvet/stub/introspection🚧
\nUIs and FFI need to list Grid methods without hardcoding every export.
\nGetMethods, GetMethodsJSON, GetMethodSignature.
\n\n\nexamples/57-stub-introspection/main.go
\ncd welvet/examples/57-stub-introspection && source ../env.sh && go run .package main\n\nimport (\n\t\"fmt\"\n\n\t\"github.com/openfluke/welvet/architecture\"\n\t\"github.com/openfluke/welvet/stub/introspection\"\n)\n\nfunc main() {\n\tg := architecture.NewGrid(1, 1, 1, 1)\n\tmethods, err := introspection.GetMethods(g)\n\tfmt.Println(len(methods), err)\n}9 <nil>github.com/openfluke/welvet/stub/observer🚧
\nAttach forward/backward observers for debugging without coupling to tanhi UDP.
\nObserver interface; ConsoleObserver / HTTPObserver / BufferObserver; ComputeLayerStats.
\n\n\nexamples/58-stub-observer/main.go
\ncd welvet/examples/58-stub-observer && source ../env.sh && go run .package main\n\nimport (\n\t\"fmt\"\n\n\t\"github.com/openfluke/welvet/stub/observer\"\n)\n\nfunc main() {\n\tvar obs observer.Observer = &observer.ConsoleObserver{}\n\tfmt.Printf(\"%T\\n\", obs)\n}*observer.ConsoleObservergithub.com/openfluke/welvet/stub/pipeline🚧
\nDecoder wavefront stats helpers — not a full Lucy-style pipeline runner yet.
\nPipelineForwardStats, TokenTimelineSummary.
\n\n\nexamples/59-stub-pipeline/main.go
\ncd welvet/examples/59-stub-pipeline && source ../env.sh && go run .package main\n\nimport (\n\t\"fmt\"\n\n\t\"github.com/openfluke/welvet/stub/pipeline\"\n)\n\nfunc main() {\n\tvar s pipeline.PipelineForwardStats\n\tfmt.Printf(\"%+v\\n\", s)\n}{PipelineTicks:0 SubLayerOps:0 MaxActiveJobs:0 MaxBlockSpread:0 MaxDistinctBlocks:0 MaxPendingTokens:0 StallFallback:false TokenDoneTick:[]}github.com/openfluke/welvet/stub/templates🚧
\nChat prompts must match model families (ChatML, Llama3, BitNet) without app-specific string glue.
\nTemplate.BuildPrompt; presets ChatML, Llama3, BitNetInstruction, MicrosoftBitNetChat.
\n\n\nexamples/60-stub-templates/main.go
\ncd welvet/examples/60-stub-templates && source ../env.sh && go run .package main\n\nimport (\n\t\"fmt\"\n\n\t\"github.com/openfluke/welvet/stub/templates\"\n)\n\nfunc main() {\n\tp := templates.ChatML.BuildPrompt(nil, \"You are helpful.\", \"Say hi\")\n\tfmt.Println(p)\n}<|im_start|>system\nYou are helpful.\n<|im_start|>user\nSay hi\n<|im_start|>assistantgithub.com/openfluke/welvet/stub/universal🚧
\nProbe unknown safetensor geometry and mount placeholder grids until full weight import lands.
\nLoadUniversal / LoadUniversalDetailed, ProbeDeepGeometry, MountGeometrically.
\n\n\nexamples/61-stub-universal/main.go
\ncd welvet/examples/61-stub-universal && source ../env.sh && go run .package main\n\nimport (\n\t\"fmt\"\n\n\t\"github.com/openfluke/welvet/stub/universal\"\n)\n\nfunc main() {\n\tg, err := universal.LoadUniversal(\"/path/to/snapshot\")\n\tfmt.Println(g != nil, err)\n}false universal: \"/path/to/snapshot\": open /path/to/snapshot: no such file or directorygithub.com/openfluke/w2a✅ harness
\nEngine packages must stay free of tests. w2a owns timed 34×20×3 matrices, gap census, honesty stamps, and the train-mode permutation smoke (Test49). See §63 for a live full-suite run.
\nInteractive go run . ([0] Run ALL). Suites under suites/*. StampBackendNote / AffinePackable prevent fake ✅. Test49: AllNamedTrainModes × 1³/2³/3³ × Parallel/Bicameral/poly, origin-only — included in [0].
\n\n\ncd welvet/examples/62-w2a && source ../env.sh && go run .package main\n\nimport \"fmt\"\n\nfunc main() {\n\tfmt.Println(\"w2a is a separate module — engine packages never contain tests.\")\n\tfmt.Println(\"\")\n\tfmt.Println(\" cd w2a\")\n\tfmt.Println(\" go run . # interactive; [0] = ALL\")\n\tfmt.Println(\" go test ./tests/dense -v\")\n\tfmt.Println(\" go test ./tests/parallel -run Test49AllTrainModesCubes -count=1 -v\")\n}w2a is a separate module — engine packages never contain tests.\n\n cd w2a\n go run . # interactive; [0] = ALL\n go test ./tests/dense -v\n go test ./tests/parallel -run Test49AllTrainModesCubes -count=1 -vcd w2a && go run .\ncd w2a && go test ./tests/dense -v && go test ./tests/parallel -run Test49AllTrainModesCubes -count=1 -vgithub.com/openfluke/w2a✅ 246k cells
\nClaims are cheap; a stamped matrix is not. This is the actual output of one full w2a [0] Run ALL so the book's ✅ marks are backed by numbers you can reproduce, not asserted.
\nEvery timed layer sweeps its dtype × format × backend matrix; every suite runs its case checks. v1.0 board: 246,032 matrix cells, FAIL 0, RESULT PASS. GAP cells are declared skips (not silent fails).
\n\n[0] Run ALL — 246,032 matrix cells — FAIL 0 — RESULT: PASS.\nGAP remains (honesty stamps): GDN non-float32 / non-packable formats, AffinePacked, exotic hemi9 dtypes,\ncensus + cross-numeric train declared skips, a handful of ser cells. GAP ≠ fail. Case-only stubs\n(seed, serialization, memory, donate, fountain, hardware) stay off the engine scorecard.ResidualGraft.Do not quote this stamp as a Lucy race or “beats PyTorch.” It is surface coverage: every claimed path either\nruns or is stamped GAP. Lucy Acc vs Score lives in AAI test50 (§68).
\nW2A Step includes Cross-Numeric Train (kinds × weight dtype × act host; full census ~10.7k under the train op). Storage truth checked after StepMesh — no retained f32 master for non-float32 FormatNone. Public Dense demotion + W×A showcase: down-the-dem (chapter 65).
examples/63-validation/main.go
\ncd welvet/examples/63-validation && source ../env.sh && go run .package main\n\nimport \"fmt\"\n\nfunc main() {\n\tcells := 246032\n\tfmt.Printf(\"w2a [0] ALL: %d cells FAIL 0 RESULT PASS\\n\", cells)\n\tfmt.Println(\"GAP = declared skip (GDN non-f32, AffinePacked, …) — not a fail\")\n}w2a [0] ALL: 246032 cells FAIL 0 RESULT PASS\nGAP = declared skip (GDN non-f32, AffinePacked, …) — not a failcd w2a && go run . # [0] Run ALL; writes logs/suite.txtv1.1.1
\nVersion is earned from a weighted board, not marketing. v1.0 is 100/100 on the engine board. Minor/patch tags (v1.1.0, v1.1.1, …) pack features without a new board. Apps, stubs, and NPU sit off-board.
\nversion = 0.{round(earned)} until 100 → v1.0. Scorecard today 100/100; this book tags v1.1.1. Training credit is §9 on the board (8 pts), not an afterthought.
\n\n| § | Area | Wt | Earned |
|---|---|---|---|
| 1 | Foundation | 15 | 15 |
| 2 | Dense MatVec microkernel | 15 | 15 |
| 3 | Transformer stack | 14 | 14 |
| 4 | CNN / RNN / LSTM | 6 | 6 |
| 5 | Extended layers (GDN, ConvT, Mamba, Parallel, …) | 7 | 7 |
| 6 | Runtime + architecture | 8 | 8 |
| 7 | Systems | 5 | 5 |
| 8 | Model / IO | 8 | 8 |
| 9 | Training credit (29 named TrainModes + BranchModes) | 8 | 8 |
| 10 | Peak fused / no host ALU | 14 | 14 |
| Total → v1.0 board | 100 | 100 |
BuildLPD density (§66 · §69)DotTileF64, expand-once,\nBF16 convert; ~4.3× geo-mean on 34-dtype deep suite vs prior Go; Dense-backed projs inherit\n(§6 · §11)training_modes.md + Lucy Lean density + book §70\n(§70 · §69)apps/octo, stub/*, NPU/Metal/QNN (later welvet.cpp).\nFastProxy can match/beat StepBP Acc on sine/copy toys; Sparse wins Lucy Score via Avail.\nDo not write “Sparse is better backprop.” Engine v1 ≠ “beats PyTorch.”cd welvet/examples/64-scorecard && source ../env.sh && go run .package main\n\nimport \"fmt\"\n\nfunc main() {\n\tearned := 100.0\n\tfmt.Println(\"v1.1.1\")\n\tfmt.Printf(\"scorecard %.0f/100\\n\", earned)\n}v1.1.1\nscorecard 100/100github.com/openfluke/welvet/runtime/training✅
\nWeight storage dtype and activation Tensor[T] are independent axes. Proving train without a retained float32 master means sweeping W×A — not only matched float32 acts.
\nW2A Step Cross-Numeric Train: polyops.AllKinds() × FormatNone weight dtype × Go Numeric act host (smoke ~735; full ~10.7k) via StepMesh, then assert no retained f32 master. Public Dense volumetric showcase: down-the-dem — dtype demotion ladder, packed quants, and the full 34×15×3 perm matrix with charts/PDF.
\n\ndown-the-dem — same 5-layer Dense stack on cubes 1³ / 2³ / 3³; only the number story changes:
\nRepo + report: github.com/openfluke/down-the-dem\n · PDF: report/down-the-dem.pdf in that repo.\n Honesty: runnable storage-truth train ≠ task accuracy or native low-precision GEMV ALU.
examples/65-cross-numeric/main.go
\ncd welvet/examples/65-cross-numeric && source ../env.sh && go run .package main\n\nimport (\n\t\"fmt\"\n\n\t\"github.com/openfluke/welvet/architecture\"\n\t\"github.com/openfluke/welvet/core\"\n\t\"github.com/openfluke/welvet/layers/dense\"\n\t\"github.com/openfluke/welvet/quant\"\n\t\"github.com/openfluke/welvet/runtime/training\"\n)\n\nfunc main() {\n\t// Weight storage: int8 FormatNone. Activations/grads: Tensor[int8].\n\tg := architecture.NewGrid(1, 1, 1, 1)\n\tinit := make([]float32, 4*4)\n\tfor i := range init {\n\t\tinit[i] = 0.1\n\t}\n\tl, err := dense.NewConfigured(4, 4, core.ActivationLinear, core.DTypeInt8, quant.FormatNone, init)\n\tif err != nil {\n\t\tpanic(err)\n\t}\n\tif err := dense.Place(g, 0, 0, 0, 0, l); err != nil {\n\t\tpanic(err)\n\t}\n\tx := core.NewTensor[int8](1, 4)\n\ty := core.NewTensor[int8](1, 4)\n\tfor i := 0; i < 4; i++ {\n\t\tx.Data[i] = int8(i + 1)\n\t\ty.Data[i] = int8(i)\n\t}\n\tloss, _, err := training.StepMesh(g, x, y, 1, 0.05)\n\tfmt.Println(loss, err, l.Weights.RetainsF32Master())\n}1.5 <nil> falsegithub.com/openfluke/welvet/lucy✅
\nAdaptation benches (test41-w, tide, live_gpt) need one shared measuring math — SoftAcc, Availability, AdaptPct, Score — not three copies of the formulas.
\nPure measuring package: SoftAcc / SoftAccProb, Window + Snapshot, Finalize. No datasets, no train loops. Sine scale 0.10; classification SoftAccProb scale 1.0. Density / synthetic-organism board is §69.
\nSoftAcc = 100 × (1 − |pred−target| / scale) clamped [0,100]\nAvailability = InferMs / (InferMs + TrainMs) × 100\nScore = Throughput × Availability × Acc / 10_000\nAcc = hard argmax % (AvgAccuracy) — the Acc pillar\nAdaptPct = mean SoftAcc in AdaptWindows after each switch
tide/metrics re-exports this package for existing tide callers.\nEngine tests for lucy live under w2a/tests/lucy.
AAI Lucy benches (private tree, replace → public Welvet): test41 native cam, test48 credit sweep, test50 23-mode FP32 race.\nHarness is not an engine package. Pulse math is this chapter. Density board: §69.\nModes + equations: §67. Cameral sandwiches: §68.
cd welvet/examples/66-lucy && source ../env.sh && go run .package main\n\nimport (\n\t\"fmt\"\n\n\t\"github.com/openfluke/welvet/lucy\"\n)\n\nfunc main() {\n\ta := lucy.SoftAccOne(0.72, 0.80) // sine scale 0.10\n\tp := lucy.SoftAccProb(0.91, 1.0) // class scale 1.0\n\tvar snap lucy.Snapshot\n\tsnap.AvgAccuracy = 80\n\tsnap.SoftAcc = a\n\tsnap.InferMs = 8\n\tsnap.TrainMs = 2\n\tsnap.Throughput = 12000\n\tlucy.Finalize(&snap, lucy.Options{AdaptWindows: 10})\n\tfmt.Printf(\"soft=%.1f class=%.1f avail=%.1f score=%.0f\\n\",\n\t\ta, p, snap.Availability, snap.Score)\n}soft=20.0 class=91.0 avail=80.0 score=7680github.com/openfluke/welvet/layers/parallel✅ 29 modes
\nBackprop is one update, not the only one. Credit assignment (broadcast gap, head proxy, sparse duty clock) has to be a named axis you can race — not a comment in a notebook. Cameral Mix also needs one TrainMode per hemisphere on the same loss.
\nparallel.TrainMode: AllNamedTrainModes() = 29 (Inherit omitted). Stack-local Split/Alt plus Step* 1D pipe twins and Mesh* grid schedulers. TrainStackMSE / TrainStackCE honour BranchModes. Display names use Short() / ShortTrainMode. Rival metric is hard Acc vs StepBP; Lucy Score is Tput × Avail × Acc — do not mix those sentences.
\n\nMSE: L = (1/d) ||ŷ − t||² g_y = (2/d)(ŷ − t)\nCE: L = −mean log softmax(ŷ)_class g_y = (p − t) / B\n
Head always sees g_y. Classification hosts call TrainStackCE so Acc can leave chance;\nMSE on a one-hot stays uniform. What each mode does with g_y is the whole story.
| Family | Tokens | Update |
|---|---|---|
| Backprop | NormalBP · StepBP · MeshBP | chain rule J^T through the tape / volumetric Step |
| Tween | Tween · StepTween · MeshTween | broadcast P(g_y) onto every leaf; η ← η/2 |
| TweenChain | TweenChain · StepTweenChain · MeshTweenChain | same math as BP on a Sandwich |
| Split | TweenSplit · StepTweenSplit · MeshTweenSplit | g_i = (1/N) P(g_y) |
| Alt | TweenAlt · StepTweenAlt · MeshTweenAlt | Split then re-forward then Tween (half LR) |
| HeadProxy | TweenSplitHeadProxy · StepTweenSplitHeadProxy | head J^T g_y (with act′); hidden dW only |
| FastProxy | TweenSplitFastProxy · StepTweenSplitFastProxy · MeshTweenSplitFastProxy | g_proxy = W_head^T g_y (skip act′) |
| Linear | TweenSplitLinear · StepTweenSplitLinear | affine W^T walk; skip ⊙ act′; hemispheres share the down-vector |
| LinearCache | TweenSplitLinearCache · StepTweenSplitLinearCache | cache every 20 steps; dead on sine — control |
| HeadProxyAsync | TweenSplitHeadProxyAsync · StepTweenSplitHeadProxyAsync | hidden uses proxy from T−1; not EMA |
| Sparse | TweenSplitSparse · StepTweenSplitSparse · MeshTweenSplitSparse | head + one rotating hidden; other dW = 0 |
AllCreditTrainModes() = 16 stack-local Split/Alt plus Step* credit twins (no Mesh).\nAllMeshCreditTrainModes() = 4 Mesh credit (no HeadProxy / Linear / LinearCache / HeadProxyAsync Mesh twins).\nAllStackLocalTrainModes() = 22 (no Inherit, no Mesh*).\nAllNamedTrainModes() = 29 — the Test49 / test50 set.\nIsLineStep() = 11 (StepBP, StepTween, StepTweenChain, StepTweenSplit, StepTweenAlt, plus the six Step* credit twins).
Tables and logs use Short() / ShortTrainMode (legend: [T]=Tween [S]=Split [FP]=FastProxy [L]=Linear [HP]=HeadProxy). Persistence and ParseTrainMode still use the full String() token. New Step* iota values were appended so old uint8 numbers stay stable.
head: g_head = J_head^T g_y\nhemi: g_hemi = J_hemi^T g_head\nstem: g_stem = J_stem^T g_hemi\ndW_i from local Backward(g_i, x_i)\n
SIMD GEMV. This is who FastProxy has to beat on hard Acc — not Lucy Score.
\ng_i = P(g_y) η ← η/2\ndW_i = localBackward(g_i, x_i)\n
Blind to W^T sign. Sine Acc collapses. MeshTween is the volumetric scheduler of this family, not a secret FastProxy.
\ng_i = (1/N) P(g_y)\n
Still not J^T. Cheap Acc ceiling. Score can rise because the update is cheap (Avail), not because g is better.
\nPer sample, AltTimes times (default 1): Split from live g_y → re-forward → Tween from g_y′ (half LR). Extra forwards kill Avail → Score last.
\ng_proxy = J_head^T g_y = W_head^T (g_y ⊙ act′(pre_head))\nhidden i = 1…N−1: g_i = 1/(N−1) P(g_proxy)\n dW_i = g_i x_i^T (no discarded W^T)\n
One real J_head^T. Hemispheres do not get J_hemi^T.
\ng_proxy = W_head^T g_y ← skip act′\nhead dW still uses act′\nhidden: same 1/(N−1) P(g_proxy) dW-only as HeadProxy\n
DFA with B := W_head^T, not a learned random B. On AAI test50 sine, FastProxy SoftAcc often sits above StepBP while both are at 100% hard Acc — that is the FastProxy vs BP sentence.
\ng_head↓ = g_y\ng_hemi↓ = W_head^T g_y (siblings share the vector)\ng_stem↓ = Σ_hemi W_hemi^T g_hemi↓\nthen every leaf: g_i = (1/N) P(g_i↓), dW_i = g_i x_i^T\n
Same O(N²) class as backprop. Score stays StepBP-class unless Acc is way up.
\nevery 20 steps: full Linear walk, cache g_i↓\nelse: g_i ← g_i^cache · ||g_y||_live / ||g_y||_cache\n
Norm scaling cannot recover sign flips after a frequency switch. If this wins sine, the board is lying.
\ng_hidden^(T) = 1/(N−1) P(g_proxy^(T−1))\nhead computes g_proxy^(T) = W_head^T g_y^(T) for next step\n
First sample seeds live. Stale sign on XOR. Not EMA.
\ng_proxy = W_head^T g_y\ndW_head = g_y x_head^T\nk = t mod (N−1)\ndW_k = P(g_proxy) x_k^T\nother leaves: dW = 0 this sample\n
Real FLOP cut → Avail 40–50% → Lucy Score explodes. That is a duty clock, not a smaller big-O than backprop and not a better chain rule. On test50 copy, Sparse often loses Acc vs StepBP while winning Score.
\nThe Step prefix on a stack mode is a schedule, not a second leftover-forward pass and not Mesh*.\nIsLineStep → TrainLine / trainStackLine: one sample enters child 0 per tick; every in-flight sample advances one layer; the output (and the train event) is the sample that entered D ticks ago. Fill ticks do not update. Serve stays a full ForwardStack.
Same family update on the Sandwich whether Step or not — HeadProxy / Linear / FastProxy / Sparse / Async credit walk trainTweenSplitLeaves on both twins. Mesh* still requires a Grid (RequiresGrid). There is no Mesh HeadProxy / Linear / LinearCache / HeadProxyAsync.
MeshBP = volumetric training.Step. MeshTween = StepMesh. MeshTweenChain = StepTween.\nMesh Split / Alt / FastProxy / Sparse credit the placed stack under a grid walk.\nOn origin-only 1³/2³/3³ (rest IsDisabled) Mesh* usually matches the stack twin. Cube size is hop topology, not 27 sandwiches.
[0] Run ALL. Not a Lucy race.Cameral why + sandwich stem→mid→head: §68. Measuring math: §66.
\n\n\nexamples/67-train-modes/main.go
\ncd welvet/examples/67-train-modes && source ../env.sh && go run .package main\n\nimport (\n\t\"fmt\"\n\n\t\"github.com/openfluke/welvet/layers/parallel\"\n)\n\nfunc main() {\n\tnamed := parallel.AllNamedTrainModes()\n\tline := 0\n\tfor _, m := range named {\n\t\tif m.IsLineStep() {\n\t\t\tline++\n\t\t}\n\t}\n\tfp, err := parallel.ParseTrainMode(\"stepfastproxy\")\n\tfmt.Println(\"named\", len(named), \"linestep\", line)\n\tfmt.Println(\"stepfastproxy\", fp.Short(), err)\n\tfmt.Println(parallel.ShortTrainModeLegend)\n}named 31 linestep 11\nstepfastproxy Step[T][S][FP] <nil>\n[T]=Tween [S]=Split [FP]=FastProxy [L]=Linear [HP]=HeadProxy [F]=Freeze [Sh]=Shadow [A]=Adv [M]=Memorygithub.com/openfluke/welvet/layers/parallel✅ cameral
\nA single Dense chain cannot host two independent weight copies that share an input, merge, and optionally train under different updates. That is the cameral graph: hemispheres, not screen-space sprites and not a second hidden size.
\nHemispheres / Bicameral / Sandwich / Mix. Stem → Parallel mid → Dense head. SetBranchModes + TrainStackMSE. Lucy races in AAI (test41 / test48 / test50) import this API; measuring is lucy/; harness is not engine.
\nx → Dense stem → ┬→ Hemi 1 (own W, own TrainMode)\n ├→ Hemi 2\n └→ Hemi n merge add|avg|concat|filter\n → Dense head → ŷ\n MSE → g_y → each branch’s update
Welvet already has Sequential (ordered compose) and Residual (skip). Neither is “two brains on one x.”\nA hemisphere is an independent sub-net with its own weights. Bi / Tri / Quad are n hemispheres plus a merge.\nMix stamps a different TrainMode per hemi (SetBranchModes) so left can be StepBP\nwhile right is TweenChain — one loss on the merge. Grid training.Step applies one mode to the whole net;\nMix needs TrainStackMSE or the BranchModes stamp is a lie.
The sandwich (stem → mid → head) is not extra depth for its own sake. Credit modes need a head Jacobian\n(or a W_head^T proxy). Without a head, FastProxy / HeadProxy have nothing to inject. Stem gets the down-going\nvector after hemispheres merge. That is why AAI Lucy jobs are sandwiches even when cameral count is 1\n(mid is a single Dense, not Parallel).
\nHemispheres / HemispheresFrom — n twins; merge add/avg/concat/filterBicameral / BicameralFrom / Sandwich — stem → Parallel → head StackSetBranchModes / TrainStackMSE — Mix pathResidualGraft — y = F(x)+x when F is ParallelWriteCameralFile / LoadCameral — cameral .entity (in model/entity)CamSync — optional weight averaging across cams / layers → §70Layer package chapter: §27. Named updates: §67.\nWeight sync across cams (and across mesh layouts when shapes match): §70.
\nAAI is a separate tree. It replaces github.com/openfluke/welvet (public chaosglue module).\nThe engine does not import AAI. Lucy math is §66.
| Bench | What it answers |
|---|---|
| test41 | Native cameral Lucy (Bi/Tri/Mix) on toys — does the sandwich train at all. |
| test48 | Credit sweep: layers × dtypes × short jobs. Home of the equations in §67. |
| test50 | Deep FP32 race: all 29 named modes × Dense/Bi/Tri × 1³/2³/3³ origin-only. Rival = hard Acc vs StepBP. |
test50 copy: Split / Alt / MeshSplit family about +5 to +12 Acc over StepBP (~68–74%).\nSine: Acc ceiling (many modes 100% including StepBP); FastProxy SoftAcc ~56–68 vs StepBP ~42–53.\nXOR: almost everyone at 75% (3 of 4 bits) — parking lot. Sparse wins Score, not Acc.\nOrigin-only cubes: extra cells are disabled; 3³ is hop topology, not 27 copies.
\nw2a Test49 is the permutation smoke of those 29 modes (in [0] Run ALL), not a Lucy race.\nDo not quote Test49 as “we beat backprop.”
cd welvet/examples/68-cameral && source ../env.sh && go run .package main\n\nimport (\n\t\"fmt\"\n\n\t\"github.com/openfluke/welvet/core\"\n\t\"github.com/openfluke/welvet/layers/parallel\"\n\t\"github.com/openfluke/welvet/quant\"\n)\n\nfunc main() {\n\ts, err := parallel.Bicameral(8, 16, 1, core.ActivationLeakyReLU,\n\t\tcore.DTypeFloat32, quant.FormatNone)\n\tif err != nil {\n\t\tpanic(err)\n\t}\n\themi := s.Children[1].(*parallel.Layer)\n\themi.SetBranchModes(parallel.ModeStepBP, parallel.ModeTweenSplitFastProxy)\n\n\tx := core.NewTensor[float32](1, 8)\n\tt := core.NewTensor[float32](1, 1)\n\tt.Data[0] = 0.5\n\tloss, err := parallel.TrainStackMSE(s, x, t, parallel.ModeStepBP, 0.01)\n\tfmt.Println(\"mix loss\", loss, \"err\", err)\n}mix loss 0 err <nil>github.com/openfluke/welvet/lucy✅ BuildLPD
\nA new host (char LM, MNIST, a layer sprint) should not copy tide's goldilocks math. The question is always the same: can the net run and train at the same time in a small box, then how far that live-fit condenses without becoming a trap. That ruler belongs in the engine.
\nlucy.BuildLPD ranks Samples for consciousness (Acc / Throughput / Availability keep vs learner peaks) then Lucy density (Q × shrink vs Acc-champ RAM). Gold / near / keep / trap bands, consciousness radar, and memory-density radar are computed here. Tide dash and Lucy PDF only draw the board.
\nWHY\n Industry ships dead calculators: INT8 inference, no on-device train.\n Synthetic organism = serve while learning, then shrink without falling\n into chance-Acc \"tiny & fast\" traps.\n\nPILLARS (consciousness)\n Acc = argmax % (did it learn?)\n Thru = outputs / second (can it act while live?)\n Avail = InferMs / busy (can you still talk to it while it trains?)\n Score = T × Avail × Acc / 10_000 live-fit\n SoftAcc is serve-confidence — not a pillar.\n\nDENSITY (memory intelligence)\n Acc champ = RAM reference (best learner, not the Score champ)\n Learner = RelAcc ≥ 70% of Acc champ (traps do not set Thru/Avail peaks)\n Q = geomean(RelAcc, RelThru, RelAvail)\n shrink = AccChampRAM / thisRAM (capped ×32)\n LPD = Q × shrink if RelAcc ≥ 70% else 0\n\nBANDS\n gold all 3 pillars ≥80% and RAM ≤20% of Acc champ\n near Acc + (Thru or Avail) ≥80% at ≤50% RAM\n trap RAM ≤20% and Acc keep <70% (binary looking dense)\n Score/MiB is the trap metric — do not use it for goldilocks.\n\nRADARS (for PDF / dash — math here, drawing in tide)\n Consciousness = (RelAcc, RelThru, RelAvail)\n Memory density = (RelAcc, RelThru, RelAvail) × shrink traps at origin
The goal is not a better chatbot and not a paper for a faculty board. It is a\nsynthetic organism: a net that owns time — continuous serve, continuous train,\nno turn-based halt, no phone-home. Lucy Score asks the thermodynamic question\ntide was built for: does SGD that blocks inference die, and do proxy / Split / Sparse\npaths keep the live loop while still learning?
\nHard Acc is whether it learned. Availability is whether it still breathes.\nThroughput is how many actions it took while both were happening. Multiply them\nand you get live-fit. SoftAcc is how confident the serve looked — useful, not the Acc term.
\nMemory density asks the second question: once you have a learner, how far can\ndtype / quant / cameral width condense that live-fit versus the Acc champ's RAM\nwithout collapsing to chance Acc. A binary cell that is 30× smaller and 4× faster\nwith 12% Acc is a trap (LPD = 0). An int8 that keeps ≥70% of Acc-champ Acc\nat a fifth the RAM is goldilocks.
\n| Symbol | Formula | Meaning |
|---|---|---|
| Availability | InferMs / (InferMs + TrainMs) × 100 | Duty cycle. SGD that blocks serve dies here. |
| Throughput T | TotalOutputs / duration_s | Actions per second while the sweep is live. |
| Acc | argmax % (AvgAccuracy) | Learning. Acc champ is the RAM reference. |
| Lucy Score | T × Avail × Acc / 10_000 | Live-fit. SoftAcc is not this term. |
| Realtime | T × Avail / 100 | Duty-cycle speed without Acc. |
| ZeroDowntime | Acc × Avail / 100 | Still serving while learning. |
| RelAcc / RelThru / RelAvail | value / learner peak, clamped [0,1] | Keep vs cells that actually learn. |
| Q | geomean of the three Rel* | Consciousness mix. Traps do not set Thru/Avail peaks. |
| shrink | min(AccChampRAM / thisRAM, 32) | How much smaller than the Acc champ. |
| LPD | Q × shrink if RelAcc ≥ 0.70 else 0 | Lucy density. Weeds Score/MiB traps. |
| Gold | all 3 Rel* ≥ 0.80 and RAM ≤ 20% Acc champ | Trifecta in a small box. |
| Gold-std | Acc ≥ 80% plus Thru or Avail, then smallest then fastest | Two-or-more of the trifecta. |
| Trap | RAM ≤ 20% Acc champ and RelAcc < 0.70 | Tiny and fast with chance Acc. |
| MobileScore | Score / WeightMiB | The binary trap. Use LPD instead. |
Tide is the race: serve+train pulses, permute matrix, dashboard, Lucy PDF\n(including drawing the two radars). tide/report.BuildLPD pretty-prints cell IDs\nthen calls lucy.BuildLPD. A new thing that only needs the ruler:
import \"github.com/openfluke/welvet/lucy\"\nboard := lucy.BuildLPD(samples)\n_ = board.Top[0].Consciousness() // Acc, Thru, Avail keep\n_ = board.Top[0].MemoryDensity() // same × shrink; traps at 0Classification hosts should train with TrainStackCE so the Acc pillar can leave chance;\nMSE on a one-hot stays uniform. That is a loss gap in layers/parallel, not a Lucy formula.
Pulse math: §66. Train modes: §67.
\n\n\nexamples/69-lucy-density/main.go
\ncd welvet/examples/69-lucy-density && source ../env.sh && go run .package main\n\nimport (\n\t\"fmt\"\n\n\t\"github.com/openfluke/welvet/lucy\"\n)\n\nfunc main() {\n\tboard := lucy.BuildLPD([]lucy.Sample{\n\t\t{ID: \"f32\", Mode: \"sgd\", Acc: 90, Thru: 200, Avail: 40, Score: 100, RAMKiB: 1000},\n\t\t{ID: \"int8\", Mode: \"sgd\", Acc: 82, Thru: 180, Avail: 38, Score: 85, RAMKiB: 180},\n\t\t{ID: \"bin\", Mode: \"sgd\", Acc: 12, Thru: 400, Avail: 50, Score: 40, RAMKiB: 40},\n\t})\n\ttop := board.Top[0]\n\tfmt.Printf(\"lead=%s band=%s LPD=%.2f trap=%s\\n\",\n\t\ttop.ID, top.Band, top.LPD, board.Trap[0].ID)\n}lead=int8 band=gold LPD=5.11 trap=bingithub.com/openfluke/welvet/layers/parallel✅ CamSync
\nMix BranchModes and different inits let cams diverge on purpose. Sometimes you want them to share — gently (1% pull) or hard (full average) — within a Parallel, across Stack children, or even between same-shaped stores sitting on wildly different mesh layouts (1×2×1 ↔ 2×4×5). That is CamSync: couple weights without collapsing the graph into one Dense.
\nCamSyncConfig{Alpha, When, Groups, Cross}. BlendStores pulls every store in a clique toward the mean. Empty Groups = all cams; Cross wires SyncEndpoint pairs across layers/cams. Shape rule: Rows×Cols must match. Bidirectional today; one-way teacher→student not yet.
\nParallel mid (n cams) Stack / mesh (any DxHxW)\n ┌─ cam0 W ─┐ cell A 1×2×1 cell B 2×4×5\n │ cam1 W │ Groups / all-cams layer-1 Dense cam3 layer-1 Dense\n │ cam2 W │ α·mean blend OR [128×64] ←──Cross──→ [128×64]\n └─ cam3 W ─┘ (mesh size irrelevant — shape is)\n │\n SyncAfterSample | Step | Pulse | Manual SyncNow
1×2×1 cell and another store sits on\ncam 3 of a 2×4×5. Name both endpoints; if Rows×Cols match, they blend.\nVolumetric layout is configuration — not a built-in grid type inside CamSync.| Field | Meaning |
|---|---|
Enabled / Alpha | \nSoft pull w ← (1−α)·w + α·mean. 0.01 = 1%, 1.0 = hard sync.\nDefault α is 1 when Enabled and Alpha ≤ 0. |
When | \nSyncManual · SyncAfterSample · SyncAfterStep · SyncAfterPulse.\nHooks fire from TrainStackMSE/CE and Pulse(). |
Groups | \nWithin one Parallel: cam-index cliques. Empty ⇒ one group of every branch.\nExample: {{0,1},{2,3}} syncs pairs only. |
Cross | \n[]SyncPair of SyncEndpoint{StackIdx, Branch, Store} —\nsame layer across cams, Dense stem ↔ matching cam Dense, or any same-shaped pair the Stack can resolve. |
Rows×Cols — rejected or skippedweights.BlendStores on the two storesparallel.CamSyncConfig · DefaultCamSync() · SetCamSync on Layer or StackSyncNow / Pulse / MaybeSyncweights.BlendStores · weights.StoreCosine (diagnostics)Cameral sandwiches: §68. Parallel package: §27.\nHost probe: ok/cam_sync (MNIST sample α×cams matrix).
cd welvet/examples/70-cam-sync && source ../env.sh && go run .package main\n\nimport (\n\t\"fmt\"\n\n\t\"github.com/openfluke/welvet/core\"\n\t\"github.com/openfluke/welvet/layers/dense\"\n\t\"github.com/openfluke/welvet/layers/parallel\"\n\t\"github.com/openfluke/welvet/quant\"\n\t\"github.com/openfluke/welvet/stub/seed\"\n\t\"github.com/openfluke/welvet/weights\"\n)\n\nfunc main() {\n\ts, err := parallel.Bicameral(8, 16, 1, core.ActivationLeakyReLU,\n\t\tcore.DTypeFloat32, quant.FormatNone)\n\tif err != nil {\n\t\tpanic(err)\n\t}\n\themi := s.Children[1].(*parallel.Layer)\n\tleft := hemi.Branches[0].(*dense.Layer).Weights\n\tright := hemi.Branches[1].(*dense.Layer).Weights\n\t_ = seed.InitStoreHe(left, 16, seed.From(\"cam0\"))\n\t_ = seed.InitStoreHe(right, 16, seed.From(\"cam1\"))\n\n\themi.SetCamSync(parallel.CamSyncConfig{\n\t\tEnabled: true,\n\t\tAlpha: 1.0,\n\t\tWhen: parallel.SyncManual,\n\t})\n\tbefore, _ := weights.StoreCosine(left, right)\n\tif err := hemi.SyncNow(); err != nil {\n\t\tpanic(err)\n\t}\n\tafter, _ := weights.StoreCosine(left, right)\n\tfmt.Printf(\"cosine before=%.4f after=%.4f\\n\", before, after)\n}cosine before=-0.0697 after=1.0000Part: I · Orientation
Package: github.com/openfluke/welvet
Status: ok — ✅ engine
You need github.com/openfluke/welvet.
import \"github.com/openfluke/welvet/github.com/openfluke/welvet\"
cd 01-welvet && source ../env.sh && go run .\nLoom’s flat poly/ package hit import-cycle and honesty walls (QAT morph, silent fallbacks, god-layer). Welvet is the rewrite: one feature per folder, storage-truth dtypes/quants, Dense as the shared MatVec microkernel, tests only in w2a, apps only in apps/.
\nAn AI engine in Go: layers, 34 dtypes, 20 quant formats, and three backends (CPU tiled · Plan 9 SIMD · WebGPU). Version tracks a 100-point scorecard (today v1.1.0 · 100/100).
\npre 4 post 4\nLive capture from go run ./cmd/runall on local ../../welvet (exit 0).
Copied from the Welvet feature book examples (openfluke.github.io/welvet/examples/01-welvet).
Part: I · Orientation
Package: —
Status: ok — ✅ layout
Orientation chapter: Repository map.
\nBook / repo map (no single import)
\ncd 02-tree && source ../env.sh && go run .\nReaders need a single map of what is engine vs harness vs app vs stub.
\nTop-level folders and their ownership. Engine packages never import w2a.
\nimport github.com/openfluke/welvet/core/…\nimport github.com/openfluke/welvet/weights/…\nimport github.com/openfluke/welvet/quant/…\nimport github.com/openfluke/welvet/simd/…\nimport github.com/openfluke/welvet/webgpu/…\nimport github.com/openfluke/welvet/tiling/…\nimport github.com/openfluke/welvet/architecture/…\nimport github.com/openfluke/welvet/fusedgpu/…\nimport github.com/openfluke/welvet/layers/…\nimport github.com/openfluke/welvet/runtime/…\nimport github.com/openfluke/welvet/systems/…\nimport github.com/openfluke/welvet/lucy/…\nimport github.com/openfluke/welvet/model/…\nLive capture from go run ./cmd/runall on local ../../welvet (exit 0).
Copied from the Welvet feature book examples (openfluke.github.io/welvet/examples/02-tree).
Part: II · Foundation
Package: github.com/openfluke/welvet/core
Status: ok — ✅
You need the core foundation package.
import \"github.com/openfluke/welvet/core\"
cd 03-core && source ../env.sh && go run .\nEvery polymorphic path needs one place for DType, LayerType, Activation, Backend, Tensor[T], and slim Layer metadata — without QAT morph defaults.
\n34 storage dtypes, Numeric generics, Tensor[T], ExecConfig (BackendCPUTiled | BackendSIMD | BackendWebGPU), Activate/ActivateDeriv, and converters for f16/bf16/fp8/fp4.
\nsimd bfloat16 8\nLive capture from go run ./cmd/runall on local ../../welvet (exit 0).
Copied from the Welvet feature book examples (openfluke.github.io/welvet/examples/03-core).
Part: II · Foundation
Package: github.com/openfluke/welvet/weights
Status: ok — ✅
You need the weights foundation package.
import \"github.com/openfluke/welvet/weights\"
cd 04-weights && source ../env.sh && go run .\nUnquantized matrices still need a typed store that streams MatVec and SGD without forcing a float32 master or Morph-as-training.
\nStore holds DType + Format + Native/Packed bytes. New[T], MatVec, MatVecT, DecodeRow, SelectWire (F32/F64/I8), ApplySGD. FormatNone: update in native lanes (float32 payload is the only f32 buffer; float64 uses native ALU; other dtypes decode→update→re-encode). Packed: unpack→update→re-Pack then drop scratch. RetainsF32Master() is true only for FormatNone+float32. Dense and composite projs share this store.
\n[3 4]\nLive capture from go run ./cmd/runall on local ../../welvet (exit 0).
Copied from the Welvet feature book examples (openfluke.github.io/welvet/examples/04-weights).
Part: II · Foundation
Package: github.com/openfluke/welvet/quant
Status: ok — ✅
You need the quant foundation package.
import \"github.com/openfluke/welvet/quant\"
cd 05-quant && source ../env.sh && go run .\nInference, storage, and train need classic Q-packs, k-quants, IQ, Ternary/Binary, and Affine without a separate QAT mode or retained f32 master. Format is storage truth.
\nPack / Unpack / MatVec / MatVecT for FormatNone…AffinePacked. Dense SIMD once-projects codes into Int8QS + scales (EnsureQ* / EnsureK/IQ/AffineSIMDCache) — no full-matrix F32 inflate for k/IQ/Affine. SGD on packed stores: short-lived unpack scratch → update → re-Pack; Packed stays truth.
\nrows 2 cols 4 y [0.6857143 -1.4901161e-08]\nLive capture from go run ./cmd/runall on local ../../welvet (exit 0).
Copied from the Welvet feature book examples (openfluke.github.io/welvet/examples/05-quant).
Part: II · Foundation
Package: github.com/openfluke/welvet/simd
Status: ok — ✅
You need the simd foundation package.
import \"github.com/openfluke/welvet/simd\"
cd 06-simd && source ../env.sh && go run .\nCPU peak needs hand-written AVX2/NEON without a silent Go fallback that pretends SIMD ran.
\namd64/arm64 .s kernels: DotTile, DotI8/U8, DotQ4_0, Saxpy, BitNet helpers, packed f16/bf16/fp8/fp4 dots; amd64 AVX2 DotTileF64 (WireF64); Go fused DotKRow / DotIQRow / DotAffineRow for k/IQ/Affine. SimdEnabled() false → BackendSIMD hard-errors.
\n10\nLive capture from go run ./cmd/runall on local ../../welvet (exit 0).
Copied from the Welvet feature book examples (openfluke.github.io/welvet/examples/06-simd).
Part: II · Foundation
Package: github.com/openfluke/welvet/webgpu
Status: ok — ✅
You need the webgpu foundation package.
import \"github.com/openfluke/welvet/webgpu\"
cd 07-webgpu && source ../env.sh && go run .\nGPU paths must bind a real adapter. Host “fake GPU” was banned so suites cannot stamp WebGPU done when ALU ran on CPU.
\nDenseGEMV family (incl. quant/I8/resident), DenseGEMVT/DenseDW, RMSNorm, LayerNorm fwd, Softmax family, SwiGLUFuse. Available()/InitError() gate use.
\nwebgpu probe goos=linux goarch=amd64 WELVET_WGPU_BACKEND=\"\"\n adapter high-perf: name=\"NVIDIA GeForce GTX 1650 SUPER\" vendor=\"NVIDIA\" arch=\"\" driver=\"610.57.04\" backend=vulkan type=discrete-gpu | maxBuf=1099511627776 maxSSBO=2147483644 maxSSBO/stage=524288 maxBindGroups=8 maxComputeWG=65535 maxInvoc/WG=1024\n → selected high-perf\n adapter low-power: name=\"Intel(R) UHD Graphics 630 (CML GT2)\" vendor=\"Intel open-source Mesa driver\" arch=\"\" driver=\"Mesa 26.1.8\" backend=vulkan type=integrated-gpu | maxBuf=2147483647 maxSSBO=2147483644 maxSSBO/stage=103 maxBindGroups=8 maxComputeWG=65535 maxInvoc/WG=1024\n adapter force-fallback: name=\"llvmpipe (LLVM 22.1.8, 256 bits)\" vendor=\"llvmpipe\" arch=\"\" driver=\"Mesa 26.1.8 (LLVM 22.1.8)\" backend=vulkan type=cpu | maxBuf=2147483647 maxSSBO=134217728 maxSSBO/stage=48 maxBindGroups=8 maxComputeWG=65535 maxInvoc/WG=1024\n adapter default: name=\"NVIDIA GeForce GTX 1650 SUPER\" vendor=\"NVIDIA\" arch=\"\" driver=\"610.57.04\" backend=vulkan type=discrete-gpu | maxBuf=1099511627776 maxSSBO=2147483644 maxSSBO/stage=524288 maxBindGroups=8 maxComputeWG=65535 maxInvoc/WG=1024\n try RequestDevice inflated: maxBuf=2147483648 maxSSBO=1073741824 maxSSBO/stage=524288 maxBindGroups=8 maxComputeWG=65535 maxInvoc/WG=1024\nRESULT: OK via=inflated adapter=\"NVIDIA GeForce GTX 1650 SUPER\" backend=vulkan maxSSBO=2147483644 maxBuf=1099511627776\n[1.5 2.5] <nil> NVIDIA GeForce GTX 1650 SUPER\nLive capture from go run ./cmd/runall on local ../../welvet (exit 0).
Copied from the Welvet feature book examples (openfluke.github.io/welvet/examples/07-webgpu).
Part: II · Foundation
Package: github.com/openfluke/welvet/tiling
Status: ok — ✅
You need the tiling foundation package.
import \"github.com/openfluke/welvet/tiling\"
cd 08-tiling && source ../env.sh && go run .\nMatVec throughput depends on tile size and when to go multi-core vs GPU workgroups. Centralizing caps keeps Dense and friends consistent.
\nDefaultCPUTile, DefaultGPUWG, CPUTile, PreferMultiCore, GPUWorkgroupsX.
\n32 true 16\nLive capture from go run ./cmd/runall on local ../../welvet (exit 0).
Copied from the Welvet feature book examples (openfluke.github.io/welvet/examples/08-tiling).
Part: II · Foundation
Package: github.com/openfluke/welvet/architecture
Status: ok — ✅
You need the architecture foundation package.
import \"github.com/openfluke/welvet/architecture\"
cd 09-architecture && source ../env.sh && go run .\nNetworks are spatial (Depth×Rows×Cols×LayersPerCell), not only linear stacks. Topology lives here; compute lives in layer packages.
\nGrid/Cell/Coord, NewGrid, BindOp, HopOrder, SetRemoteLink/ResolveHop. VolumetricNetwork is an alias for Grid.
\ncells 2\nLive capture from go run ./cmd/runall on local ../../welvet (exit 0).
Copied from the Welvet feature book examples (openfluke.github.io/welvet/examples/09-architecture).
Part: II · Foundation
Package: github.com/openfluke/welvet/fusedgpu
Status: ok — ✅
You need the fusedgpu foundation package.
import \"github.com/openfluke/welvet/fusedgpu\"
cd 10-fusedgpu && source ../env.sh && go run .\nToken-by-token host round-trips kill decode. A fused engine keeps weights and scratch resident for Q4_0 and BinaryG128 hybrid paths.
\nEngine from Spec (AppendTokens/Reset/Close); HybridEngine (PrefillSample/DecodeSample/DecodeChunk). Specs come from model/transformer export.
\nneed an ENTITY file: open model.entity: no such file or directory\nLive capture from go run ./cmd/runall on local ../../welvet (exit 0).
Copied from the Welvet feature book examples (openfluke.github.io/welvet/examples/10-fusedgpu).
Part: III · Layers
Package: github.com/openfluke/welvet/layers/dense
Status: ok — ✅
Building or training a net that needs the layers/dense Op (also usable as a Parallel cam).
import \"github.com/openfluke/welvet/layers/dense\"
cd 11-dense && source ../env.sh && go run .\nMost FLOPs are W@x. One Dense stack owns FormatNone×34 and all quants × three backends so every composite proj shares one correctness surface — including native in-dtype SGD.
\nNew / NewConfigured[T], Forward/Backward (dispatch on Exec.Backend), Place, ApplyGradSGD (→ weights.ApplySGD on the store). SIMD forward (v1.0.3): dtype switch by MatVec strategy — DotTile / DotI8 / lowp packed / expand-once→DotTile / WireF64+DotTileF64; fused Dot* for classic Q*, k/IQ, AffinePacked. Composites (MHA, SwiGLU, CNN im2col, RNN/LSTM/Mamba, residual·sequential·parallel) reuse Dense children via syncProjExec. BackwardSIMD still DecodeRow/saxpy (not the new expand wires).
\n8 32\nLive capture from go run ./cmd/runall on local ../../welvet (exit 0).
Copied from the Welvet feature book examples (openfluke.github.io/welvet/examples/11-dense).
Part: III · Layers
Package: github.com/openfluke/welvet/layers/mha
Status: ok — ✅
Building or training a net that needs the layers/mha Op (also usable as a Parallel cam).
import \"github.com/openfluke/welvet/layers/mha\"
cd 12-mha && source ../env.sh && go run .\nTransformers need multi-head attention with masks, RoPE/ALiBi, GQA/MQA, and cross-attn — without forking MatVec for every projection.
\nQ/K/V/O are Dense children. Presets: DecoderCausal, EncoderBidirectional, CrossAttention, … Attn/RoPE ALU is host today; on-device attn shaders are still open.
\ntrue 256 <nil>\nLive capture from go run ./cmd/runall on local ../../welvet (exit 0).
Copied from the Welvet feature book examples (openfluke.github.io/welvet/examples/12-mha).
Part: III · Layers
Package: github.com/openfluke/welvet/layers/swiglu
Status: ok — ✅
Building or training a net that needs the layers/swiglu Op (also usable as a Parallel cam).
import \"github.com/openfluke/welvet/layers/swiglu\"
cd 13-swiglu && source ../env.sh && go run .\nModern decoder FFNs are SiLU(gate)⊙up → down. Projections must share Dense’s quant/backend matrix.
\nGate/Up/Down Dense children; DefaultFFN(dModel); WebGPU SiLU⊙ fuse on forward.
\n64 <nil>\nLive capture from go run ./cmd/runall on local ../../welvet (exit 0).
Copied from the Welvet feature book examples (openfluke.github.io/welvet/examples/13-swiglu).
Part: III · Layers
Package: github.com/openfluke/welvet/layers/rmsnorm
Status: ok — ✅
Building or training a net that needs the layers/rmsnorm Op (also usable as a Parallel cam).
import \"github.com/openfluke/welvet/layers/rmsnorm\"
cd 14-rmsnorm && source ../env.sh && go run .\nLlama-style blocks normalize by RMS, not mean+var. Needs native fwd/bwd and WebGPU shaders.
\nPer-token RMS + γ on weights.Store; WebGPU fwd+bwd; SIMD DotTile stats + host scale.
\n[0 0 0 0] <nil>\nLive capture from go run ./cmd/runall on local ../../welvet (exit 0).
Copied from the Welvet feature book examples (openfluke.github.io/welvet/examples/14-rmsnorm).
Part: III · Layers
Package: github.com/openfluke/welvet/layers/layernorm
Status: ok — ✅
Building or training a net that needs the layers/layernorm Op (also usable as a Parallel cam).
import \"github.com/openfluke/welvet/layers/layernorm\"
cd 15-layernorm && source ../env.sh && go run .\nClassic mean+var normalization with γ/β — still required for many HF architectures.
\nWebGPU forward; backward host today. Same dtype×quant axes as other weighted layers.
\n8 <nil>\nLive capture from go run ./cmd/runall on local ../../welvet (exit 0).
Copied from the Welvet feature book examples (openfluke.github.io/welvet/examples/15-layernorm).
Part: III · Layers
Package: github.com/openfluke/welvet/layers/embedding
Status: ok — ✅
Building or training a net that needs the layers/embedding Op (also usable as a Parallel cam).
import \"github.com/openfluke/welvet/layers/embedding\"
cd 16-embedding && source ../env.sh && go run .\nToken IDs must gather rows from a table — not a Dense MatVec — with scatter grads on backward.
\nConfig{VocabSize, EmbeddingDim, SeqLen}; table on weights.Store; host gather on SIMD/WebGPU today.
\n[1 4 16] <nil>\nLive capture from go run ./cmd/runall on local ../../welvet (exit 0).
Copied from the Welvet feature book examples (openfluke.github.io/welvet/examples/16-embedding).
Part: III · Layers
Package: github.com/openfluke/welvet/layers/softmax
Status: ok — ✅
Building or training a net that needs the layers/softmax Op (also usable as a Parallel cam).
import \"github.com/openfluke/welvet/layers/softmax\"
cd 17-softmax && source ../env.sh && go run .\nClassification heads and attention need stable softmax variants, including sparse/Gumbel/Entmax for research paths.
\nWeightless layer; KindStandard/Temperature/Grid/Hierarchical/Gumbel/Masked/Sparse/… WebGPU covers std family; exotic kinds hard-error on GPU (no silent host).
\n[0.6380664 0.23473153 0.09543471 0.031767454] <nil>\nLive capture from go run ./cmd/runall on local ../../welvet (exit 0).
Copied from the Welvet feature book examples (openfluke.github.io/welvet/examples/17-softmax).
Part: III · Layers
Package: github.com/openfluke/welvet/layers/sequential
Status: ok — ✅
Building or training a net that needs the layers/sequential Op (also usable as a Parallel cam).
import \"github.com/openfluke/welvet/layers/sequential\"
cd 18-sequential && source ../env.sh && go run .\nSome cells need an ordered Dense chain without burning grid hops.
\nDense→Dense compose, or mixed Ops via NewFromOps (Dense, SwiGLU, RMSNorm, LayerNorm). Parallel as a child is parallel.ResidualGraft / a Sandwich — Sequential cannot import parallel.
\n16 <nil>\nLive capture from go run ./cmd/runall on local ../../welvet (exit 0).
Copied from the Welvet feature book examples (openfluke.github.io/welvet/examples/18-sequential).
Part: III · Layers
Package: github.com/openfluke/welvet/layers/residual
Status: ok — ✅
Building or training a net that needs the layers/residual Op (also usable as a Parallel cam).
import \"github.com/openfluke/welvet/layers/residual\"
cd 19-residual && source ../env.sh && go run .\nSkip connections stabilize deep stacks: y = F(x) + x with correct skip grads.
\nF is Dense Dim→Dim, or mixed Ops via NewFromOps (Dense, SwiGLU, RMSNorm, LayerNorm). Parallel as F is parallel.ResidualGraft (y = F(x)+x) — Residual cannot import parallel.
\n16 <nil>\nLive capture from go run ./cmd/runall on local ../../welvet (exit 0).
Copied from the Welvet feature book examples (openfluke.github.io/welvet/examples/19-residual).
When: spatial / temporal convolution senses (1D seq, 2D images, 3D volumes).
Where: layers/cnn1, layers/cnn2, layers/cnn3
Why: im2col → Dense Proj so quant/SIMD/WebGPU and CamSync (via Proj) are shared.
cd 20-cnn && source ../env.sh && go run .\n| Op | \nUse | \n
|---|---|
| cnn1 | \nAudio / 1D sequences | \n
| cnn2 | \nImages / MNIST-style | \n
| cnn3 | \nVolumetric / video voxels | \n
Put CNN inside Parallel cams when you want CamSync on kernels.
\n","text":"When: spatial / temporal convolution senses (1D seq, 2D images, 3D volumes).Where: layers/cnn1, layers/cnn2, layers/cnn3Why: im2col → Dense Proj so quant/SIMD/WebGPU and CamSync (via Proj) are shared.\ncd 20-cnn && source ../env.sh && go run .\n\n\n\n\nOp\nUse\n\n\n\ncnn1\nAudio / 1D sequences\n\n\ncnn2\nImages / MNIST-style\n\n\ncnn3\nVolumetric / video voxels\n\n\nPut CNN inside Parallel cams when you want CamSync on kernels.\n","sources":[{"name":"main.go","code":"package main\n\nimport (\n\t\"fmt\"\n\n\t\"github.com/openfluke/welvet/core\"\n\t\"github.com/openfluke/welvet/layers/cnn1\"\n\t\"github.com/openfluke/welvet/layers/cnn2\"\n\t\"github.com/openfluke/welvet/layers/cnn3\"\n)\n\nfunc main() {\n\t// cnn1 — 1D conv (seq / audio patches)\n\tc1, err := cnn1.New(cnn1.Config{InChannels: 1, Filters: 4, SeqLen: 16, Kernel: 3, Stride: 1})\n\tmust(err)\n\tx1 := core.NewTensor[float32](1, 1, 16)\n\tfor i := range x1.Data {\n\t\tx1.Data[i] = 0.1\n\t}\n\t_, y1, err := cnn1.Forward(c1, x1)\n\tmust(err)\n\tfmt.Println(\"cnn1\", y1.Shape)\n\n\t// cnn2 — 2D vision\n\tc2, err := cnn2.New(cnn2.Config{InChannels: 1, Filters: 4, Height: 8, Width: 8, Kernel: 3, Stride: 1})\n\tmust(err)\n\tx2 := core.NewTensor[float32](1, 1, 8, 8)\n\tfor i := range x2.Data {\n\t\tx2.Data[i] = 0.1\n\t}\n\t_, y2, err := cnn2.Forward(c2, x2)\n\tmust(err)\n\tfmt.Println(\"cnn2\", y2.Shape)\n\n\t// cnn3 — 3D / volumetric\n\tc3, err := cnn3.New(cnn3.Config{InChannels: 1, Filters: 2, Depth: 4, Height: 4, Width: 4, Kernel: 2, Stride: 1})\n\tmust(err)\n\tx3 := core.NewTensor[float32](1, 1, 4, 4, 4)\n\tfor i := range x3.Data {\n\t\tx3.Data[i] = 0.1\n\t}\n\t_, y3, err := cnn3.Forward(c3, x3)\n\tmust(err)\n\tfmt.Println(\"cnn3\", y3.Shape)\n}\n\nfunc must(err error) {\n\tif err != nil {\n\t\tpanic(err)\n\t}\n}\n","url":"/source/examples/welvet/20-cnn/main.go"}],"source":"https://github.com/openfluke/example/tree/main/welvet/20-cnn","url":"https://openfluke.com/examples/welvet/20-cnn"},{"slug":"welvet/21-rnn-lstm","title":"21. layers/rnn · lstm","group":"Welvet examples","description":"Part: III · LayersPackage: github.com/openfluke/welvet/layers/lstmStatus: ok — ✅ When Building or training a net that needs the layers/lstm Op (also usable as a Parallel cam). Where import \"","html":"Part: III · Layers
Package: github.com/openfluke/welvet/layers/lstm
Status: ok — ✅
Building or training a net that needs the layers/lstm Op (also usable as a Parallel cam).
import \"github.com/openfluke/welvet/layers/lstm\"
cd 21-rnn-lstm && source ../env.sh && go run .\nSequence models before transformers still need vanilla RNN and LSTM with BPTT on the shared MatVec stack.
\nIH/HH (and LSTM gates) via Dense; recurrence ALU host; device required for WebGPU path.
\n64 64\nLive capture from go run ./cmd/runall on local ../../welvet (exit 0).
Copied from the Welvet feature book examples (openfluke.github.io/welvet/examples/21-rnn-lstm).
Part: III · Layers
Package: github.com/openfluke/welvet/layers/seqmix
Status: ok — ✅
Building or training a net that needs the layers/seqmix Op (also usable as a Parallel cam).
import \"github.com/openfluke/welvet/layers/seqmix\"
cd 22-seqmix && source ../env.sh && go run .\nAttention, SSM, linear attn, and conv mixers must not be accidental forks of mha. Naming the contract keeps packages honest.
\nKindAttention | KindSSM | KindLinearAttn | KindConvMix and Contract{Kind,DModel,MaxT}. No compute here.
\nattention\nLive capture from go run ./cmd/runall on local ../../welvet (exit 0).
Copied from the Welvet feature book examples (openfluke.github.io/welvet/examples/22-seqmix).
Part: III · Layers
Package: github.com/openfluke/welvet/layers/gdn
Status: ok — ✅
Building or training a net that needs the layers/gdn Op (also usable as a Parallel cam).
import \"github.com/openfluke/welvet/layers/gdn\"
cd 23-gdn && source ../env.sh && go run .\nLinear attention / decode-first mixers (Gated DeltaNet) need a first-class package under KindLinearAttn.
\nExec CPU/SIMD/WebGPU; ForwardDecode; truncated BPTT. Timed matrix is Float32-primary: PermutationOK is f32 × FormatNone/BinaryPacked × CPU/SIMD/WebGPU; other cells GAP (declared), not FAIL.
\n<nil>\nLive capture from go run ./cmd/runall on local ../../welvet (exit 0).
Copied from the Welvet feature book examples (openfluke.github.io/welvet/examples/23-gdn).
Part: III · Layers
Package: github.com/openfluke/welvet/layers/mamba
Status: ok — ✅
Building or training a net that needs the layers/mamba Op (also usable as a Parallel cam).
import \"github.com/openfluke/welvet/layers/mamba\"
cd 24-mamba && source ../env.sh && go run .\nSSM mixers (KindSSM) are not MHA clones — they need their own selective-scan path.
\nInProj → softplus(Δ) scan → OutProj. Full timed matrix + train grids (scan ALU host).
\ntrue <nil>\nLive capture from go run ./cmd/runall on local ../../welvet (exit 0).
Copied from the Welvet feature book examples (openfluke.github.io/welvet/examples/24-mamba).
When: upsampling / decoder paths (1D, 2D, 3D).
Where: layers/convt1, layers/convt2, layers/convt3
Why: generative twins and U-Net-style expands; same Proj/Dense backend story as CNN.
cd 25-convt && source ../env.sh && go run .\nPart: III · Layers
Package: github.com/openfluke/welvet/layers/kmeans
Status: ok — ✅
Building or training a net that needs the layers/kmeans Op (also usable as a Parallel cam).
import \"github.com/openfluke/welvet/layers/kmeans\"
cd 26-kmeans && source ../env.sh && go run .\nSoft clustering as a differentiable layer lets topology experiments sit inside the same train loop.
\nCenters on Dense (K×FeatureDim); soft assignment outputs. Full timed matrix + train grids.
\n[1 4] <nil>\nLive capture from go run ./cmd/runall on local ../../welvet (exit 0).
Copied from the Welvet feature book examples (openfluke.github.io/welvet/examples/26-kmeans).
Part: III · Layers
Package: github.com/openfluke/welvet/layers/parallel
Status: ok — ✅
Building or training a net that needs the layers/parallel Op (also usable as a Parallel cam).
import \"github.com/openfluke/welvet/layers/parallel\"
cd 27-parallel && source ../env.sh && go run .\nMixture-of-experts and multi-path cells need concat/add/avg/filter combines. Cameral graphs need sibling hemispheres that share input, merge outputs, and optionally train under distinct modes.
\nParallel branches + Stack sandwiches. Hemispheres / Bicameral / Sandwich build nested multi-cameral nets. SetBranchModes + TrainStackMSE let each hemi use a different TrainMode (all 29 named updates — BP, Tween, Split, FastProxy, Sparse, Step*, Mesh*, …).
\nloss 0 err <nil>\nLive capture from go run ./cmd/runall on local ../../welvet (exit 0).
Copied from the Welvet feature book examples (openfluke.github.io/welvet/examples/27-parallel).
Part: III · Layers
Package: github.com/openfluke/welvet/layers/metacognition
Status: ok — ✅
Building or training a net that needs the layers/metacognition Op (also usable as a Parallel cam).
import \"github.com/openfluke/welvet/layers/metacognition\"
cd 28-metacognition && source ../env.sh && go run .\nObserved layers can apply heuristic stability rules (gate/scale/reset) without dtype morph/QAT.
\nWraps Dense + DefaultStabilityRules(); Stats exposed. Full timed matrix + train grids.
\n16 <nil>\nLive capture from go run ./cmd/runall on local ../../welvet (exit 0).
Copied from the Welvet feature book examples (openfluke.github.io/welvet/examples/28-metacognition).
Part: IV · Runtime
Package: github.com/openfluke/welvet/runtime/forward
Status: ok — ✅
Walking a Grid/Stack with the shared runtime/forward path.
import \"github.com/openfluke/welvet/runtime/forward\"
cd 29-forward && source ../env.sh && go run .\nA grid of heterogeneous ops needs one walker that dispatches by concrete type and fails loudly on unknowns.
\nForward[T](grid, input) → Result tape; Cell[T] for single-cell dispatch. Covers Dense…Residual + extended wired ops.
\ntrue <nil>\nLive capture from go run ./cmd/runall on local ../../welvet (exit 0).
Copied from the Welvet feature book examples (openfluke.github.io/welvet/examples/29-forward).
Part: IV · Runtime
Package: github.com/openfluke/welvet/runtime/backward
Status: ok — ✅
Walking a Grid/Stack with the shared runtime/backward path.
import \"github.com/openfluke/welvet/runtime/backward\"
cd 30-backward && source ../env.sh && go run .\nTraining needs a reverse tape over the same ops forward used — no separate graph framework.
\nBackward[T](fwdResult, gradOut) walks the tape and returns per-op weight grads.
\ntrue <nil>\nLive capture from go run ./cmd/runall on local ../../welvet (exit 0).
Copied from the Welvet feature book examples (openfluke.github.io/welvet/examples/30-backward).
Part: IV · Runtime
Package: github.com/openfluke/welvet/runtime/training
Status: ok — ✅
Walking a Grid/Stack with the shared runtime/training path.
import \"github.com/openfluke/welvet/runtime/training\"
cd 31-training && source ../env.sh && go run .\nSuites and small nets need MSE+SGD and tween hooks without inventing an external trainer or a retained float32 master beside storage.
\nMSE/MSEGrad, SGD, Step, ApplyTween/StepTween, StepMesh. Layer-agnostic ApplyGradSGD dispatch (Dense…Mamba/GDN/…). FormatNone: in-dtype ApplySGD; packed: unpack→update→re-Pack. No QAT dual path — storage dtype/format is truth after every step. Sandwich credit (Split / FastProxy / Sparse / …) lives in layers/parallel TrainMode — see §67.
\n0 <nil>\nLive capture from go run ./cmd/runall on local ../../welvet (exit 0).
Copied from the Welvet feature book examples (openfluke.github.io/welvet/examples/31-training).
Part: IV · Runtime
Package: github.com/openfluke/welvet/runtime/step
Status: ok — ✅
Walking a Grid/Stack with the shared runtime/step path.
import \"github.com/openfluke/welvet/runtime/step\"
cd 32-step && source ../env.sh && go run .\nSpatial feedback (remote links) needs a discrete-time mesh where every cell updates from a double buffer — different from a decoder wavefront. Cross-numeric train also needs the same mesh with weight DType ⊥ activation Tensor[T].
\nState[T], StepForward/StepBackward/StepApplyTween / StepMesh across the grid for all wired Ops × dtype × quant × CPU/SIMD. W2A Cross-Numeric Train: polyops.AllKinds() × weight dtype × act host (smoke ~21×7×5 ≈ 735; full ~21×34×15 ≈ 10.7k) — asserts no retained f32 master after StepMesh. Stack Step* is a different clock: IsLineStep / TrainLine (1D pipe) — see §67.
\n<nil>\nLive capture from go run ./cmd/runall on local ../../welvet (exit 0).
Copied from the Welvet feature book examples (openfluke.github.io/welvet/examples/32-step).
Part: V · Systems
Package: github.com/openfluke/welvet/systems/dna
Status: ok — ✅
Adaptation, measurement, or evolution (systems/dna).
import \"github.com/openfluke/welvet/systems/dna\"
cd 33-dna && source ../env.sh && go run .\nQuant and train must be measurable as topology/weight fingerprints — DNA detects logic shifts.
\nExtractDNA, CompareNetworks, CosineSimilarity, FlattenOp / CollectStores across dtype×quant.
\n{1 map[0,0,0,0:1] []}\nLive capture from go run ./cmd/runall on local ../../welvet (exit 0).
Copied from the Welvet feature book examples (openfluke.github.io/welvet/examples/33-dna).
Part: V · Systems
Package: github.com/openfluke/welvet/systems/evolution
Status: ok — ✅
Adaptation, measurement, or evolution (systems/evolution).
import \"github.com/openfluke/welvet/systems/evolution\"
cd 34-evolution && source ../env.sh && go run .\nTopology search and weight crossover need first-class splice + NEAT on CPU-resident grids.
\nSpliceDNA, NEATMutate, NewNEATPopulation, CloneGrid — dtype/quant preserved via SetFromF32.
\n<nil>\nLive capture from go run ./cmd/runall on local ../../welvet (exit 0).
Copied from the Welvet feature book examples (openfluke.github.io/welvet/examples/34-evolution).
Part: V · Systems
Package: github.com/openfluke/welvet/systems/tween
Status: ok — ✅
Adaptation, measurement, or evolution (systems/tween).
import \"github.com/openfluke/welvet/systems/tween\"
cd 35-tween && source ../env.sh && go run .\nTarget propagation (chain-rule or Hebbian layerwise gaps) is an alternative credit-assignment path. Not the same package as TrainMode Tween / TweenChain on a Sandwich (those are layers/parallel — §67).
\nNewState, Forward, BackwardChainRule / BackwardLayerwise, ApplyGaps; SIMD DotTile/Saxpy budgets.
\nforward: no op at {0 0 0 0} (type Dense)\nLive capture from go run ./cmd/runall on local ../../welvet (exit 0).
Copied from the Welvet feature book examples (openfluke.github.io/welvet/examples/35-tween).
Part: V · Systems
Package: github.com/openfluke/welvet/systems/tanhi
Status: ok — ✅
Adaptation, measurement, or evolution (systems/tanhi).
import \"github.com/openfluke/welvet/systems/tanhi\"
cd 36-tanhi && source ../env.sh && go run .\nTraining visualization must never block the engine — best-effort UDP JSON-lines to a HUD.
\nTANHI = Tensor Activation Network Holographic Interface. Sparse non-blocking JSON-line UDP events for per-layer forward/backward HUD visualization. ConfigFromGrid, Emit/EmitSweep, DefaultUDPPort (17481). SoulGlitch-style consumers.
\nNo captured output in _manifest.json yet — run go run . and paste here.
Copied from the Welvet feature book examples (openfluke.github.io/welvet/examples/36-tanhi).
— 36-tanhi (996ms)\nLive capture from go run ./cmd/runall on local ../../welvet (exit 0).
Part: V · Systems
Package: github.com/openfluke/welvet/systems/telemetry
Status: ok — ✅
Adaptation, measurement, or evolution (systems/telemetry).
import \"github.com/openfluke/welvet/systems/telemetry\"
cd 37-telemetry && source ../env.sh && go run .\nStatic structural blueprints (sizes, op kinds) differ from live TANHI events.
\nExtractNetworkBlueprint, ExtractLayerTelemetry for introspection/UIs.
\n{ID:demo TotalLayers:1 TotalParams:0 Layers:[{Z:0 Y:0 X:0 L:0 Type:Dense Activation:ReLU Parameters:0 InputShape:[] OutputShape:[] Branches:[] CombineMode:}]}\nLive capture from go run ./cmd/runall on local ../../welvet (exit 0).
Copied from the Welvet feature book examples (openfluke.github.io/welvet/examples/37-telemetry).
Part: VI · Model IO
Package: github.com/openfluke/welvet/model/entity
Status: ok — ✅
Checkpoints, HF import, tokenize, sample, or decode (model/entity).
import \"github.com/openfluke/welvet/model/entity\"
cd 38-entity && source ../env.sh && go run .\nHF safetensors are awkward for native topology + packed weights. ENTITY is the Welvet checkpoint.
\nOpen/Inspect/IsEntity, LoadBlob/LoadQuantBlob, PackFromHF/ImportFromHF, WriteTransformerFile, SerializeNetwork, WriteCameralFile / LoadCameral (sandwich Stack + TrainMode).
\nopen model.entity: no such file or directory\nLive capture from go run ./cmd/runall on local ../../welvet (exit 0).
Copied from the Welvet feature book examples (openfluke.github.io/welvet/examples/38-entity).
Part: VI · Model IO
Package: github.com/openfluke/welvet/model/hf
Status: ok — ✅
Checkpoints, HF import, tokenize, sample, or decode (model/hf).
import \"github.com/openfluke/welvet/model/hf\"
cd 39-hf && source ../env.sh && go run .\nImport starts with probing HF/MLX layouts before packing ENTITY.
\nInspectSnapshot, DetectArchitecture, safetensors/MLX loaders, Qwen3.5 hybrid helpers.
\nhf: config.json: stat /path/to/hf-snapshot/config.json: no such file or directory\nLive capture from go run ./cmd/runall on local ../../welvet (exit 0).
Copied from the Welvet feature book examples (openfluke.github.io/welvet/examples/39-hf).
Part: VI · Model IO
Package: github.com/openfluke/welvet/model/tokenizer
Status: ok — ✅
Checkpoints, HF import, tokenize, sample, or decode (model/tokenizer).
import \"github.com/openfluke/welvet/model/tokenizer\"
cd 40-tokenizer && source ../env.sh && go run .\nGenerate needs encode/decode of HF tokenizer.json without pulling Python.
\nLoadTokenizer, Encode/Decode, LoadForEntity.
\nfailed to read tokenizer file: open tokenizer.json: no such file or directory\nLive capture from go run ./cmd/runall on local ../../welvet (exit 0).
Copied from the Welvet feature book examples (openfluke.github.io/welvet/examples/40-tokenizer).
Part: VI · Model IO
Package: github.com/openfluke/welvet/model/sampling
Status: ok — ✅
Checkpoints, HF import, tokenize, sample, or decode (model/sampling).
import \"github.com/openfluke/welvet/model/sampling\"
cd 41-sampling && source ../env.sh && go run .\nLogits → token ID needs ArgMax, TopK+temperature, penalties, and chat hygiene in one place.
\nArgMax, SampleTopK, ApplyRepetitionPenalty, BanIDs, SanitizeChatReply.
\n1\n1\nLive capture from go run ./cmd/runall on local ../../welvet (exit 0).
Copied from the Welvet feature book examples (openfluke.github.io/welvet/examples/41-sampling).
Part: VI · Model IO
Package: github.com/openfluke/welvet/model/transformer
Status: ok — ✅
Checkpoints, HF import, tokenize, sample, or decode (model/transformer).
import \"github.com/openfluke/welvet/model/transformer\"
cd 42-transformer && source ../env.sh && go run .\nENTITY packs must run as Llama-style decoders with KV cache, profiles (SIMD/WebGPU/fused), and chat templates.
\nLoadEntity → Model; Generate; ApplyExec profiles; ExportFusedGPUSpec / hybrid sync.
\nopen model.entity: no such file or directory\nLive capture from go run ./cmd/runall on local ../../welvet (exit 0).
Copied from the Welvet feature book examples (openfluke.github.io/welvet/examples/42-transformer).
Part: VII · Apps
Package: github.com/openfluke/welvet/apps/…
Status: partial — 🚧
You need the apps/… foundation package.
import \"github.com/openfluke/welvet/apps/…\"
cd 43-apps && source ../env.sh && go run .\nProducts must not pollute engine packages. Octo is the model shell; flux2/mosstts are domain apps. Lucy races (AAI test41 / test48 / test50) are benches, not Welvet packages.
\nocto (own module): download/convert/chat, see §44. flux2: MMDiT image. mosstts: Speak/SpeakToFile pipeline. AAI sandwiches race TrainMode on toys — see §68. Off the v1.0 engine board.
\nApps import the engine — the engine never imports apps.\n cd apps/octo && go run .\n cd apps/flux2 && go run .\n cd apps/mosstts && go run .\nLive capture from go run ./cmd/runall on local ../../welvet (exit 0).
Copied from the Welvet feature book examples (openfluke.github.io/welvet/examples/43-apps).
Part: VII · Apps
Package: github.com/openfluke/welvet/apps/octo
Status: ok — ✅ runs
You need the apps/octo foundation package.
import \"github.com/openfluke/welvet/apps/octo\"
cd 44-octo && source ../env.sh && go run .\nA model is only useful with a shell around it: pull weights from Hugging Face, convert them to a Welvet .entity, then chat, serve, or benchmark. Octo is that shell, kept in its own module so the engine never depends on an app.
\nSubcommands cover the whole loop: hub download/ensure repos, convert pack to a single .entity, interactive run/serve/chat, plus image and speech menus. Its bench harness sweeps every quant format across a CPU Plan 9 SIMD fused profile and a WebGPU fused profile.
\nOcto is its own module — the engine never imports it.\n cd apps/octo && go run .\n download -> convert -> .entity -> chat / serve / bench\nLive capture from go run ./cmd/runall on local ../../welvet (exit 0).
Copied from the Welvet feature book examples (openfluke.github.io/welvet/examples/44-octo).
Part: VIII · Stubs
Package: github.com/openfluke/welvet/stub/seed
Status: partial — 🚧
Scaffold / experimental stub stub/seed (may be partial).
import \"github.com/openfluke/welvet/stub/seed\"
cd 45-stub-seed && source ../env.sh && go run .\nShip topology recipes (layer seeds → He-init) without weight blobs.
\nSeedFrom/From, RNG, InitStoreHe, Dense/MHA/SwiGLU/CNN infinite manifests.
\n0.4118750632366808\nLive capture from go run ./cmd/runall on local ../../welvet (exit 0).
Copied from the Welvet feature book examples (openfluke.github.io/welvet/examples/45-stub-seed).
Part: VIII · Stubs
Package: github.com/openfluke/welvet/stub/serialization
Status: partial — 🚧
Scaffold / experimental stub stub/serialization (may be partial).
import \"github.com/openfluke/welvet/stub/serialization\"
cd 46-stub-serialization && source ../env.sh && go run .\nVolumetric grids need JSON/ENTITY persist beyond transformer packs.
\nSerializeEntity/LoadEntity, SerializeGrid/GridFromSpec — native FormatNone bytes; packed via wire.
\n101 <nil>\nLive capture from go run ./cmd/runall on local ../../welvet (exit 0).
Copied from the Welvet feature book examples (openfluke.github.io/welvet/examples/46-stub-serialization).
Part: VIII · Stubs
Package: github.com/openfluke/welvet/stub/memory
Status: partial — 🚧
Scaffold / experimental stub stub/memory (may be partial).
import \"github.com/openfluke/welvet/stub/memory\"
cd 47-stub-memory && source ../env.sh && go run .\nHF→ENTITY and GPU upload need footprint accounting and optional history charts.
\nFromGrid, Footprint, InitScavenger, ReleaseTransient; WELVET_MEMORY_HISTORY=1.
\n{HostWeightsMB:0 GPUWeightsMB:0 GPUKVMB:0}\nLive capture from go run ./cmd/runall on local ../../welvet (exit 0).
Copied from the Welvet feature book examples (openfluke.github.io/welvet/examples/47-stub-memory).
Part: VIII · Stubs
Package: github.com/openfluke/welvet/stub/donate
Status: partial — 🚧
Scaffold / experimental stub stub/donate (may be partial).
import \"github.com/openfluke/welvet/stub/donate\"
cd 48-stub-donate && source ../env.sh && go run .\nLAN donors should accept framed JSON jobs without embedding HTTP in the engine.
\nu32-LE + JSON frames; ServeTCP/Dial; model_push vs local_lm. v0 workers echo.
\ndefault port 17001\nLive capture from go run ./cmd/runall on local ../../welvet (exit 0).
Copied from the Welvet feature book examples (openfluke.github.io/welvet/examples/48-stub-donate).
Part: VIII · Stubs
Package: github.com/openfluke/welvet/stub/fountain
Status: partial — 🚧
Scaffold / experimental stub stub/fountain (may be partial).
import \"github.com/openfluke/welvet/stub/fountain\"
cd 49-stub-fountain && source ../env.sh && go run .\nRecover specialist weight blobs over lossy links via LT fountain codes, then ensemble.
\nNewLTEncoder/Decoder, RecoverWeightBlobs, Neural/DenseFactory helpers.
\n*fountain.LTEncoder <nil>\nLive capture from go run ./cmd/runall on local ../../welvet (exit 0).
Copied from the Welvet feature book examples (openfluke.github.io/welvet/examples/49-stub-fountain).
Part: VIII · Stubs
Package: github.com/openfluke/welvet/stub/hardware
Status: partial — 🚧
Scaffold / experimental stub stub/hardware (may be partial).
import \"github.com/openfluke/welvet/stub/hardware\"
cd 50-stub-hardware && source ../env.sh && go run .\nDispatchers and UIs need a portable host audit (OS/CPU/RAM/GPU).
\nAudit() → SystemAudit with linux /proc or platform fallbacks.
\n{Model:Intel(R) Core(TM) i5-10400 CPU @ 2.90GHz Logical:12 GOMAXPROCS:12}\nOS: linux | CPU: Intel(R) Core(TM) i5-10400 CPU @ 2.90GHz (12) | RAM: 31.12 GB | GPU: Unknown GPU (0.0 GB)\nLive capture from go run ./cmd/runall on local ../../welvet (exit 0).
Copied from the Welvet feature book examples (openfluke.github.io/welvet/examples/50-stub-hardware).
Part: VIII · Stubs
Package: github.com/openfluke/welvet/stub/accel
Status: missing — ⬜
Scaffold / experimental stub stub/accel (may be partial).
import \"github.com/openfluke/welvet/stub/accel\"
cd 51-stub-accel && source ../env.sh && go run .\nVendor accelerators (Intel NPU, Qualcomm QNN, Apple Metal) will plug beside WebGPU — not replace it.
\nPackage is doc.go only today (⬜). No symbols yet; reserved for plugin loader design.
\nNo captured output in _manifest.json yet — run go run . and paste here.
Copied from the Welvet feature book examples (openfluke.github.io/welvet/examples/51-stub-accel).
— 51-stub-accel (83ms)\nLive capture from go run ./cmd/runall on local ../../welvet (exit 0).
Part: VIII · Stubs
Package: github.com/openfluke/welvet/stub/clustering
Status: partial — 🚧
Scaffold / experimental stub stub/clustering (may be partial).
import \"github.com/openfluke/welvet/stub/clustering\"
cd 52-stub-clustering && source ../env.sh && go run .\nOffline clustering helpers on tensors without inventing a second math stack.
\nKMeansCluster, HierarchicalGroup, distance/silhouette helpers.
\n2 [0 1]\nLive capture from go run ./cmd/runall on local ../../welvet (exit 0).
Copied from the Welvet feature book examples (openfluke.github.io/welvet/examples/52-stub-clustering).
Part: VIII · Stubs
Package: github.com/openfluke/welvet/stub/ensemble
Status: partial — 🚧
Scaffold / experimental stub stub/ensemble (may be partial).
import \"github.com/openfluke/welvet/stub/ensemble\"
cd 53-stub-ensemble && source ../env.sh && go run .\nCombine multiple model votes and find complementary specialists.
\nMajorityVote, FindComplementaryMatches, PerformanceSimilarity.
\n[0 1 1]\nLive capture from go run ./cmd/runall on local ../../welvet (exit 0).
Copied from the Welvet feature book examples (openfluke.github.io/welvet/examples/53-stub-ensemble).
Part: VIII · Stubs
Package: github.com/openfluke/welvet/stub/evaluation
Status: partial — 🚧
Scaffold / experimental stub stub/evaluation (may be partial).
import \"github.com/openfluke/welvet/stub/evaluation\"
cd 54-stub-evaluation && source ../env.sh && go run .\nBenchmark grids through runtime/forward with deviation metrics.
\nEvaluateNetwork, MultiNetworkEvaluation, DeviationMetrics.
\n<nil> evaluation: length mismatch inputs=1 expected=4\nLive capture from go run ./cmd/runall on local ../../welvet (exit 0).
Copied from the Welvet feature book examples (openfluke.github.io/welvet/examples/54-stub-evaluation).
Part: VIII · Stubs
Package: github.com/openfluke/welvet/stub/grafting
Status: partial — 🚧
Scaffold / experimental stub stub/grafting (may be partial).
import \"github.com/openfluke/welvet/stub/grafting\"
cd 55-stub-grafting && source ../env.sh && go run .\nMerge grids into Parallel/Residual structures for topology experiments.
\nGraftGrids, ResidualGraft, GraftToGrid — Dense branches in v0.
\nfalse grafting: no valid branches\nLive capture from go run ./cmd/runall on local ../../welvet (exit 0).
Copied from the Welvet feature book examples (openfluke.github.io/welvet/examples/55-stub-grafting).
Part: VIII · Stubs
Package: github.com/openfluke/welvet/stub/grouping
Status: partial — 🚧
Scaffold / experimental stub stub/grouping (may be partial).
import \"github.com/openfluke/welvet/stub/grouping\"
cd 56-stub-grouping && source ../env.sh && go run .\nDetect layer archetypes from safetensor-style names before mounting.
\nGroupRelatedTensors, DetectMHA/DetectSwiGLU/DetectRMSNorm → ArchetypeHint.
\ntrue {MultiHeadAttention 64 4}\nLive capture from go run ./cmd/runall on local ../../welvet (exit 0).
Copied from the Welvet feature book examples (openfluke.github.io/welvet/examples/56-stub-grouping).
Part: VIII · Stubs
Package: github.com/openfluke/welvet/stub/introspection
Status: partial — 🚧
Scaffold / experimental stub stub/introspection (may be partial).
import \"github.com/openfluke/welvet/stub/introspection\"
cd 57-stub-introspection && source ../env.sh && go run .\nUIs and FFI need to list Grid methods without hardcoding every export.
\nGetMethods, GetMethodsJSON, GetMethodSignature.
\n9 <nil>\nLive capture from go run ./cmd/runall on local ../../welvet (exit 0).
Copied from the Welvet feature book examples (openfluke.github.io/welvet/examples/57-stub-introspection).
Part: VIII · Stubs
Package: github.com/openfluke/welvet/stub/observer
Status: partial — 🚧
Scaffold / experimental stub stub/observer (may be partial).
import \"github.com/openfluke/welvet/stub/observer\"
cd 58-stub-observer && source ../env.sh && go run .\nAttach forward/backward observers for debugging without coupling to tanhi UDP.
\nObserver interface; ConsoleObserver / HTTPObserver / BufferObserver; ComputeLayerStats.
\n*observer.ConsoleObserver\nLive capture from go run ./cmd/runall on local ../../welvet (exit 0).
Copied from the Welvet feature book examples (openfluke.github.io/welvet/examples/58-stub-observer).
Part: VIII · Stubs
Package: github.com/openfluke/welvet/stub/pipeline
Status: partial — 🚧
Scaffold / experimental stub stub/pipeline (may be partial).
import \"github.com/openfluke/welvet/stub/pipeline\"
cd 59-stub-pipeline && source ../env.sh && go run .\nDecoder wavefront stats helpers — not a full Lucy-style pipeline runner yet.
\nPipelineForwardStats, TokenTimelineSummary.
\n{PipelineTicks:0 SubLayerOps:0 MaxActiveJobs:0 MaxBlockSpread:0 MaxDistinctBlocks:0 MaxPendingTokens:0 StallFallback:false TokenDoneTick:[]}\nLive capture from go run ./cmd/runall on local ../../welvet (exit 0).
Copied from the Welvet feature book examples (openfluke.github.io/welvet/examples/59-stub-pipeline).
Part: VIII · Stubs
Package: github.com/openfluke/welvet/stub/templates
Status: partial — 🚧
Scaffold / experimental stub stub/templates (may be partial).
import \"github.com/openfluke/welvet/stub/templates\"
cd 60-stub-templates && source ../env.sh && go run .\nChat prompts must match model families (ChatML, Llama3, BitNet) without app-specific string glue.
\nTemplate.BuildPrompt; presets ChatML, Llama3, BitNetInstruction, MicrosoftBitNetChat.
\n<|im_start|>system\nYou are helpful.\n<|im_start|>user\nSay hi\n<|im_start|>assistant\nLive capture from go run ./cmd/runall on local ../../welvet (exit 0).
Copied from the Welvet feature book examples (openfluke.github.io/welvet/examples/60-stub-templates).
Part: VIII · Stubs
Package: github.com/openfluke/welvet/stub/universal
Status: partial — 🚧
Scaffold / experimental stub stub/universal (may be partial).
import \"github.com/openfluke/welvet/stub/universal\"
cd 61-stub-universal && source ../env.sh && go run .\nProbe unknown safetensor geometry and mount placeholder grids until full weight import lands.
\nLoadUniversal / LoadUniversalDetailed, ProbeDeepGeometry, MountGeometrically.
\nfalse universal: \"/path/to/snapshot\": open /path/to/snapshot: no such file or directory\nLive capture from go run ./cmd/runall on local ../../welvet (exit 0).
Copied from the Welvet feature book examples (openfluke.github.io/welvet/examples/61-stub-universal).
Part: IX · Validate
Package: github.com/openfluke/w2a
Status: ok — ✅ harness
You need github.com/openfluke/w2a.
import \"github.com/openfluke/welvet/github.com/openfluke/w2a\"
cd 62-w2a && source ../env.sh && go run .\nEngine packages must stay free of tests. w2a owns timed 34×20×3 matrices, gap census, honesty stamps, and the train-mode permutation smoke (Test49). See §63 for a live full-suite run.
\nInteractive go run . ([0] Run ALL). Suites under suites/*. StampBackendNote / AffinePackable prevent fake ✅. Test49: AllNamedTrainModes × 1³/2³/3³ × Parallel/Bicameral/poly, origin-only — included in [0].
\nw2a is a separate module — engine packages never contain tests.\n\n cd w2a\n go run . # interactive; [0] = ALL\n go test ./tests/dense -v\n go test ./tests/parallel -run Test49AllTrainModesCubes -count=1 -v\nLive capture from go run ./cmd/runall on local ../../welvet (exit 0).
Copied from the Welvet feature book examples (openfluke.github.io/welvet/examples/62-w2a).
Part: IX · Validate
Package: github.com/openfluke/w2a
Status: ok — ✅ 246k cells
You need github.com/openfluke/w2a.
import \"github.com/openfluke/welvet/github.com/openfluke/w2a\"
cd 63-validation && source ../env.sh && go run .\nClaims are cheap; a stamped matrix is not. This is the actual output of one full w2a [0] Run ALL so the book's ✅ marks are backed by numbers you can reproduce, not asserted.
\nEvery timed layer sweeps its dtype × format × backend matrix; every suite runs its case checks. v1.0 board: 246,032 matrix cells, FAIL 0, RESULT PASS. GAP cells are declared skips (not silent fails).
\nw2a [0] ALL: 246032 cells FAIL 0 RESULT PASS\nGAP = declared skip (GDN non-f32, AffinePacked, …) — not a fail\nLive capture from go run ./cmd/runall on local ../../welvet (exit 0).
Copied from the Welvet feature book examples (openfluke.github.io/welvet/examples/63-validation).
Part: IX · Validate
Package: —
Status: ok — v1.1.0
Orientation chapter: Scorecard → v1.0 / minors.
\nBook / repo map (no single import)
\ncd 64-scorecard && source ../env.sh && go run .\nVersion is earned from a weighted board, not marketing. v1.0 is 100/100 on the engine board. Minor tags (v1.1.0, …) pack features without a new board. Apps, stubs, and NPU sit off-board.
\nversion = 0.{round(earned)} until 100 → v1.0. Scorecard today 100/100; this book tags v1.1.0. Training credit is §9 on the board (8 pts), not an afterthought.
\nv1.1.0\nscorecard 100/100\n— 64-scorecard (122ms)\n\n======== welvet book: ok=70 fail=0 skip=0 ========\nLive capture from go run ./cmd/runall on local ../../welvet (exit 0).
Copied from the Welvet feature book examples (openfluke.github.io/welvet/examples/64-scorecard).
Part: IV · Runtime
Package: github.com/openfluke/welvet/runtime/training
Status: ok — ✅
Walking a Grid/Stack with the shared runtime/training path.
import \"github.com/openfluke/welvet/runtime/training\"
cd 65-cross-numeric && source ../env.sh && go run .\nWeight storage dtype and activation Tensor[T] are independent axes. Proving train without a retained float32 master means sweeping W×A — not only matched float32 acts.
\nW2A Step Cross-Numeric Train: polyops.AllKinds() × FormatNone weight dtype × Go Numeric act host (smoke ~735; full ~10.7k) via StepMesh, then assert no retained f32 master. Public Dense volumetric showcase: down-the-dem — dtype demotion ladder, packed quants, and the full 34×15×3 perm matrix with charts/PDF.
\n1.5 <nil> false\nLive capture from go run ./cmd/runall on local ../../welvet (exit 0).
Copied from the Welvet feature book examples (openfluke.github.io/welvet/examples/65-cross-numeric).
Part: V · Systems
Package: github.com/openfluke/welvet/lucy
Status: ok — ✅
Adaptation, measurement, or evolution (lucy).
import \"github.com/openfluke/welvet/lucy\"
cd 66-lucy && source ../env.sh && go run .\nAdaptation benches (test41-w, tide, live_gpt) need one shared measuring math — SoftAcc, Availability, AdaptPct, Score — not three copies of the formulas.
\nPure measuring package: SoftAcc / SoftAccProb, Window + Snapshot, Finalize. No datasets, no train loops. Sine scale 0.10; classification SoftAccProb scale 1.0. Density / synthetic-organism board is §69.
\nsoft=20.0 class=91.0 avail=80.0 score=7680\nLive capture from go run ./cmd/runall on local ../../welvet (exit 0).
Copied from the Welvet feature book examples (openfluke.github.io/welvet/examples/66-lucy).
Part: IV · Runtime
Package: github.com/openfluke/welvet/layers/parallel
Status: ok — ✅ 29 modes
Building or training a net that needs the layers/parallel Op (also usable as a Parallel cam).
import \"github.com/openfluke/welvet/layers/parallel\"
cd 67-train-modes && source ../env.sh && go run .\nBackprop is one update, not the only one. Credit assignment (broadcast gap, head proxy, sparse duty clock) has to be a named axis you can race — not a comment in a notebook. Cameral Mix also needs one TrainMode per hemisphere on the same loss.
\nparallel.TrainMode: AllNamedTrainModes() = 29 (Inherit omitted). Stack-local Split/Alt plus Step* 1D pipe twins and Mesh* grid schedulers. TrainStackMSE / TrainStackCE honour BranchModes. Display names use Short() / ShortTrainMode. Rival metric is hard Acc vs StepBP; Lucy Score is Tput × Avail × Acc — do not mix those sentences.
\nnamed 31 linestep 11\nstepfastproxy Step[T][S][FP] <nil>\n[T]=Tween [S]=Split [FP]=FastProxy [L]=Linear [HP]=HeadProxy [F]=Freeze [Sh]=Shadow [A]=Adv [M]=Memory\nLive capture from go run ./cmd/runall on local ../../welvet (exit 0).
Copied from the Welvet feature book examples (openfluke.github.io/welvet/examples/67-train-modes).
Part: VII · Apps
Package: github.com/openfluke/welvet/layers/parallel
Status: ok — ✅ cameral
Building or training a net that needs the layers/parallel Op (also usable as a Parallel cam).
import \"github.com/openfluke/welvet/layers/parallel\"
cd 68-cameral && source ../env.sh && go run .\nA single Dense chain cannot host two independent weight copies that share an input, merge, and optionally train under different updates. That is the cameral graph: hemispheres, not screen-space sprites and not a second hidden size.
\nHemispheres / Bicameral / Sandwich / Mix. Stem → Parallel mid → Dense head. SetBranchModes + TrainStackMSE. Lucy races in AAI (test41 / test48 / test50) import this API; measuring is lucy/; harness is not engine.
\nmix loss 0 err <nil>\nLive capture from go run ./cmd/runall on local ../../welvet (exit 0).
Copied from the Welvet feature book examples (openfluke.github.io/welvet/examples/68-cameral).
Part: V · Systems
Package: github.com/openfluke/welvet/lucy
Status: ok — ✅ BuildLPD
Adaptation, measurement, or evolution (lucy).
import \"github.com/openfluke/welvet/lucy\"
cd 69-lucy-density && source ../env.sh && go run .\nA new host (char LM, MNIST, a layer sprint) should not copy tide's goldilocks math. The question is always the same: can the net run and train at the same time in a small box, then how far that live-fit condenses without becoming a trap. That ruler belongs in the engine.
\nlucy.BuildLPD ranks Samples for consciousness (Acc / Throughput / Availability keep vs learner peaks) then Lucy density (Q × shrink vs Acc-champ RAM). Gold / near / keep / trap bands, consciousness radar, and memory-density radar are computed here. Tide dash and Lucy PDF only draw the board.
\nlead=int8 band=gold LPD=5.11 trap=bin\nLive capture from go run ./cmd/runall on local ../../welvet (exit 0).
Copied from the Welvet feature book examples (openfluke.github.io/welvet/examples/69-lucy-density).
Part: VII · Apps
Package: github.com/openfluke/welvet/layers/parallel
Status: ok — ✅ CamSync
Building or training a net that needs the layers/parallel Op (also usable as a Parallel cam).
import \"github.com/openfluke/welvet/layers/parallel\"
cd 70-cam-sync && source ../env.sh && go run .\nMix BranchModes and different inits let cams diverge on purpose. Sometimes you want them to share — gently (1% pull) or hard (full average) — within a Parallel, across Stack children, or even between same-shaped stores sitting on wildly different mesh layouts (1×2×1 ↔ 2×4×5). That is CamSync: couple weights without collapsing the graph into one Dense.
\nCamSyncConfig{Alpha, When, Groups, Cross}. BlendStores pulls every store in a clique toward the mean. Empty Groups = all cams; Cross wires SyncEndpoint pairs across layers/cams. Shape rule: Rows×Cols must match. Bidirectional today; one-way teacher→student not yet.
\ncosine before=-0.0697 after=1.0000\nLive capture from go run ./cmd/runall on local ../../welvet (exit 0).
Copied from the Welvet feature book examples (openfluke.github.io/welvet/examples/70-cam-sync).
When: walking a grid / Parallel / Sequential and you only have any cells.
Where: github.com/openfluke/welvet/runtime/dispatch
Why: one Forward/Backward/Pack/SetDType/ApplyGrad path for every Welvet Op; unknown types hard-error (never silently pretend they are Dense).
Do not import dispatch from layers/parallel (import cycle) — Parallel dispatches branches itself.
cd 71-dispatch && source ../env.sh && go run .\nWhen: speech to text (greedy CTC) from 16 kHz mono WAV/PCM.
Where: github.com/openfluke/welvet/model/wav2vec2
Why: run facebook/wav2vec2-base-960h in pure Go.
Always proves Config + Vocab + LoadHFDir error path. Full ASR needs HF snapshot via WAV2VEC2_DIR / WAV2VEC2_WAV.
cd 72-wav2vec2 && source ../env.sh && go run .\nWhen: you want a single command that proves every layers/* Op (plus seqmix + dispatch) actually forwards.
Where: this folder imports every concrete layer package.
Why: chapters 11–28 split by topic; this is the regression checklist for “does it run?”
Cameral proofs (Freeze / CamSync on each Op) live in ../../cam/05_layers.
cd 73-all-layers && source ../env.sh && go run .\nEach line prints when / where / why for that Op.
\n","text":"When: you want a single command that proves every layers/* Op (plus seqmix + dispatch) actually forwards.Where: this folder imports every concrete layer package.Why: chapters 11–28 split by topic; this is the regression checklist for “does it run?”\nCameral proofs (Freeze / CamSync on each Op) live in ../../cam/05_layers.\ncd 73-all-layers && source ../env.sh && go run .\n\nEach line prints when / where / why for that Op.\n","sources":[{"name":"main.go","code":"package main\n\nimport (\n\t\"fmt\"\n\t\"os\"\n\n\t\"github.com/openfluke/welvet/core\"\n\t\"github.com/openfluke/welvet/layers/cnn1\"\n\t\"github.com/openfluke/welvet/layers/cnn2\"\n\t\"github.com/openfluke/welvet/layers/cnn3\"\n\t\"github.com/openfluke/welvet/layers/convt1\"\n\t\"github.com/openfluke/welvet/layers/convt2\"\n\t\"github.com/openfluke/welvet/layers/convt3\"\n\t\"github.com/openfluke/welvet/layers/dense\"\n\t\"github.com/openfluke/welvet/layers/embedding\"\n\t\"github.com/openfluke/welvet/layers/gdn\"\n\t\"github.com/openfluke/welvet/layers/kmeans\"\n\t\"github.com/openfluke/welvet/layers/layernorm\"\n\t\"github.com/openfluke/welvet/layers/lstm\"\n\t\"github.com/openfluke/welvet/layers/mamba\"\n\t\"github.com/openfluke/welvet/layers/metacognition\"\n\t\"github.com/openfluke/welvet/layers/mha\"\n\t\"github.com/openfluke/welvet/layers/parallel\"\n\t\"github.com/openfluke/welvet/layers/residual\"\n\t\"github.com/openfluke/welvet/layers/rmsnorm\"\n\t\"github.com/openfluke/welvet/layers/rnn\"\n\t\"github.com/openfluke/welvet/layers/seqmix\"\n\t\"github.com/openfluke/welvet/layers/sequential\"\n\t\"github.com/openfluke/welvet/layers/softmax\"\n\t\"github.com/openfluke/welvet/layers/swiglu\"\n\t\"github.com/openfluke/welvet/quant\"\n\t\"github.com/openfluke/welvet/runtime/dispatch\"\n)\n\ntype demo struct {\n\tname, when, where, why string\n\trun func() error\n}\n\nfunc main() {\n\tdemos := []demo{\n\t\t{\"dense\", \"linear / FFN / proj\", \"layers/dense\", \"MatVec microkernel everything else builds on\", runDense},\n\t\t{\"cnn1\", \"1D seq / audio patches\", \"layers/cnn1\", \"im2col→Dense Proj\", runCNN1},\n\t\t{\"cnn2\", \"images / 2D maps\", \"layers/cnn2\", \"vision stem\", runCNN2},\n\t\t{\"cnn3\", \"volumes / voxels\", \"layers/cnn3\", \"3D conv\", runCNN3},\n\t\t{\"convt1\", \"1D upsample\", \"layers/convt1\", \"decoder / generative 1D\", runConvT1},\n\t\t{\"convt2\", \"2D upsample\", \"layers/convt2\", \"decoder / U-Net expand\", runConvT2},\n\t\t{\"convt3\", \"3D upsample\", \"layers/convt3\", \"volumetric decoder\", runConvT3},\n\t\t{\"mha\", \"token mixing via attention\", \"layers/mha\", \"transformers\", runMHA},\n\t\t{\"swiglu\", \"gated FFN\", \"layers/swiglu\", \"modern LLM MLP\", runSwiGLU},\n\t\t{\"rmsnorm\", \"stabilize activations\", \"layers/rmsnorm\", \"LLM-style norm\", runRMS},\n\t\t{\"layernorm\", \"stabilize activations\", \"layers/layernorm\", \"classic transformer norm\", runLN},\n\t\t{\"softmax\", \"distributions / heads\", \"layers/softmax\", \"classify / combine\", runSoftmax},\n\t\t{\"embedding\", \"token → vector\", \"layers/embedding\", \"lookup tables\", runEmbed},\n\t\t{\"rnn\", \"short temporal\", \"layers/rnn\", \"classic recurrent\", runRNN},\n\t\t{\"lstm\", \"gated temporal\", \"layers/lstm\", \"longer memory than RNN\", runLSTM},\n\t\t{\"sequential\", \"stack Dense depth\", \"layers/sequential\", \"deep hemisphere body\", runSeq},\n\t\t{\"residual\", \"skip + body\", \"layers/residual\", \"stable depth\", runRes},\n\t\t{\"parallel\", \"multi-cam / MoE\", \"layers/parallel\", \"combine cams + CamSync host\", runPara},\n\t\t{\"kmeans\", \"prototypes / codebook\", \"layers/kmeans\", \"cluster features\", runKMeans},\n\t\t{\"mamba\", \"selective SSM\", \"layers/mamba\", \"linear-time seq mix\", runMamba},\n\t\t{\"gdn\", \"gated delta / linear attn\", \"layers/gdn\", \"decode-friendly mixer\", runGDN},\n\t\t{\"metacognition\", \"observer / confidence\", \"layers/metacognition\", \"meta signals beside main path\", runMeta},\n\t\t{\"seqmix\", \"name the mixer kind\", \"layers/seqmix\", \"contract only — swap MHA/Mamba/GDN\", runSeqMix},\n\t\t{\"dispatch\", \"any Op → Forward\", \"runtime/dispatch\", \"grid walk without type switches in your code\", runDispatch},\n\t}\n\n\tfail := 0\n\tfor _, d := range demos {\n\t\tif err := d.run(); err != nil {\n\t\t\tfmt.Printf(\"FAIL %-14s %v\\n\", d.name, err)\n\t\t\tfail++\n\t\t\tcontinue\n\t\t}\n\t\tfmt.Printf(\"OK %-14s when=%s | where=%s | why=%s\\n\", d.name, d.when, d.where, d.why)\n\t}\n\tfmt.Printf(\"\\nall-layers: ok=%d fail=%d\\n\", len(demos)-fail, fail)\n\tif fail > 0 {\n\t\tos.Exit(1)\n\t}\n}\n\nfunc fill(shape ...int) *core.Tensor[float32] {\n\tt := core.NewTensor[float32](shape...)\n\tfor i := range t.Data {\n\t\tt.Data[i] = 0.1 + float32(i%7)*0.03\n\t}\n\treturn t\n}\n\nfunc runDense() error {\n\tl, err := dense.New(4, 4, core.ActivationReLU, core.DTypeFloat32)\n\tif err != nil {\n\t\treturn err\n\t}\n\t_, _, err = dense.Forward(l, fill(1, 4))\n\treturn err\n}\nfunc runCNN1() error {\n\tl, err := cnn1.New(cnn1.Config{InChannels: 1, Filters: 2, SeqLen: 4, Kernel: 3})\n\tif err != nil {\n\t\treturn err\n\t}\n\t_, _, err = cnn1.Forward(l, fill(1, 1, 4))\n\treturn err\n}\nfunc runCNN2() error {\n\tl, err := cnn2.New(cnn2.Config{InChannels: 1, Filters: 2, Height: 3, Width: 3, Kernel: 2})\n\tif err != nil {\n\t\treturn err\n\t}\n\t_, _, err = cnn2.Forward(l, fill(1, 1, 3, 3))\n\treturn err\n}\nfunc runCNN3() error {\n\tl, err := cnn3.New(cnn3.Config{InChannels: 1, Filters: 2, Depth: 2, Height: 2, Width: 2, Kernel: 2})\n\tif err != nil {\n\t\treturn err\n\t}\n\t_, _, err = cnn3.Forward(l, fill(1, 1, 2, 2, 2))\n\treturn err\n}\nfunc runConvT1() error {\n\tl, err := convt1.New(convt1.Config{InChannels: 1, Filters: 2, SeqLen: 2, Kernel: 3})\n\tif err != nil {\n\t\treturn err\n\t}\n\t_, _, err = convt1.Forward(l, fill(1, 1, 2))\n\treturn err\n}\nfunc runConvT2() error {\n\tl, err := convt2.New(convt2.Config{InChannels: 1, Filters: 2, Height: 2, Width: 2, Kernel: 2})\n\tif err != nil {\n\t\treturn err\n\t}\n\t_, _, err = convt2.Forward(l, fill(1, 1, 2, 2))\n\treturn err\n}\nfunc runConvT3() error {\n\tl, err := convt3.New(convt3.Config{InChannels: 1, Filters: 2, Depth: 2, Height: 2, Width: 2, Kernel: 2})\n\tif err != nil {\n\t\treturn err\n\t}\n\t_, _, err = convt3.Forward(l, fill(1, 1, 2, 2, 2))\n\treturn err\n}\nfunc runMHA() error {\n\tl, err := mha.New(mha.Config{DModel: 4, NumHeads: 1, MaxSeqLen: 4})\n\tif err != nil {\n\t\treturn err\n\t}\n\t_, _, err = mha.Forward(l, fill(1, 2, 4))\n\treturn err\n}\nfunc runSwiGLU() error {\n\tl, err := swiglu.New(swiglu.Config{InputDim: 4, IntermediateDim: 8})\n\tif err != nil {\n\t\treturn err\n\t}\n\t_, _, err = swiglu.Forward(l, fill(1, 4))\n\treturn err\n}\nfunc runRMS() error {\n\tl, err := rmsnorm.New(rmsnorm.Config{Dim: 4})\n\tif err != nil {\n\t\treturn err\n\t}\n\tx := fill(1, 4)\n\tfor i := range x.Data {\n\t\tx.Data[i] = float32(i + 1)\n\t}\n\t_, _, err = rmsnorm.Forward(l, x)\n\treturn err\n}\nfunc runLN() error {\n\tl, err := layernorm.New(layernorm.Config{Dim: 4})\n\tif err != nil {\n\t\treturn err\n\t}\n\tx := fill(1, 4)\n\tfor i := range x.Data {\n\t\tx.Data[i] = float32(i + 1)\n\t}\n\t_, _, err = layernorm.Forward(l, x)\n\treturn err\n}\nfunc runSoftmax() error {\n\tl, err := softmax.New(softmax.Config{Dim: 4})\n\tif err != nil {\n\t\treturn err\n\t}\n\t_, _, err = softmax.Forward(l, fill(1, 4))\n\treturn err\n}\nfunc runEmbed() error {\n\tl, err := embedding.New(embedding.Config{VocabSize: 8, EmbeddingDim: 4, SeqLen: 2})\n\tif err != nil {\n\t\treturn err\n\t}\n\tx := fill(1, 2)\n\tx.Data[0], x.Data[1] = 1, 2\n\t_, _, err = embedding.Forward(l, x)\n\treturn err\n}\nfunc runRNN() error {\n\tl, err := rnn.New(rnn.Config{InputSize: 3, HiddenSize: 4, SeqLen: 2})\n\tif err != nil {\n\t\treturn err\n\t}\n\t_, _, err = rnn.Forward(l, fill(1, 2, 3))\n\treturn err\n}\nfunc runLSTM() error {\n\tl, err := lstm.New(lstm.Config{InputSize: 3, HiddenSize: 4, SeqLen: 2})\n\tif err != nil {\n\t\treturn err\n\t}\n\t_, _, err = lstm.Forward(l, fill(1, 2, 3))\n\treturn err\n}\nfunc runSeq() error {\n\tl, err := sequential.New(sequential.Config{Dim: 4, Depth: 2})\n\tif err != nil {\n\t\treturn err\n\t}\n\t_, _, err = sequential.Forward(l, fill(1, 4))\n\treturn err\n}\nfunc runRes() error {\n\tl, err := residual.New(residual.Config{Dim: 4, Depth: 1})\n\tif err != nil {\n\t\treturn err\n\t}\n\t_, _, err = residual.Forward(l, fill(1, 4))\n\treturn err\n}\nfunc runPara() error {\n\ts, err := parallel.Bicameral(4, 8, 1, core.ActivationReLU, core.DTypeFloat32, quant.FormatNone)\n\tif err != nil {\n\t\treturn err\n\t}\n\t_, _, err = parallel.ForwardStack(s, fill(1, 4))\n\treturn err\n}\nfunc runKMeans() error {\n\tl, err := kmeans.New(kmeans.Config{NumClusters: 3, FeatureDim: 4})\n\tif err != nil {\n\t\treturn err\n\t}\n\t_, _, err = kmeans.Forward(l, fill(1, 4))\n\treturn err\n}\nfunc runMamba() error {\n\tl, err := mamba.New(mamba.Config{DModel: 4, DState: 2, SeqLen: 2})\n\tif err != nil {\n\t\treturn err\n\t}\n\t_, _, err = mamba.Forward(l, fill(1, 2, 4))\n\treturn err\n}\nfunc runGDN() error {\n\tl, err := gdn.New(gdn.Config{HiddenSize: 4, NumKeyHeads: 1, NumValueHeads: 1, KeyHeadDim: 2, ValueHeadDim: 2, ConvKernel: 2})\n\tif err != nil {\n\t\treturn err\n\t}\n\t_, _, err = gdn.Forward(l, fill(1, 2, 4))\n\treturn err\n}\nfunc runMeta() error {\n\tl, err := metacognition.New(metacognition.Config{Dim: 4})\n\tif err != nil {\n\t\treturn err\n\t}\n\t_, _, err = metacognition.Forward(l, fill(1, 4))\n\treturn err\n}\nfunc runSeqMix() error {\n\tc := seqmix.Contract{Kind: seqmix.KindSSM, DModel: 4, MaxT: 8}\n\tif c.Kind.String() != \"ssm\" {\n\t\treturn fmt.Errorf(\"kind=%s\", c.Kind)\n\t}\n\treturn nil\n}\nfunc runDispatch() error {\n\tl, err := dense.New(4, 4, core.ActivationLinear, core.DTypeFloat32)\n\tif err != nil {\n\t\treturn err\n\t}\n\t_, _, err = dispatch.Forward(l, fill(1, 4))\n\treturn err\n}\n","url":"/source/examples/welvet/73-all-layers/main.go"}],"source":"https://github.com/openfluke/example/tree/main/welvet/73-all-layers","url":"https://openfluke.com/examples/welvet/73-all-layers"},{"slug":"welvet/74-kokoro","title":"74. Kokoro — reserved empty slots","group":"Welvet examples","description":"When: you specifically want Kokoro TTS.Where: welvet/model/kokoro and welvet/apps/kokoro — both empty placeholders today.Why: names reserved for a future port; do not import them yet. Workin","html":"When: you specifically want Kokoro TTS.
Where: welvet/model/kokoro and welvet/apps/kokoro — both empty placeholders today.
Why: names reserved for a future port; do not import them yet.
Working speech apps right now:
\n| App | \nRole | \n
|---|---|
apps/mosstts | \nMoss TTS shell | \n
apps/qwentts | \nQwen TTS shell | \n
cd 74-kokoro && source ../env.sh && go run .\nWhen: you need a product shell (chat, TTS, ASR, image) rather than a library Op.
Where: welvet/apps/<name> — each is its own Go module.
Why: engine packages never import apps; apps replace → welvet.
This chapter asserts every known app directory exists and prints guidance.
It does not download model weights or start interactive CLIs.
cd 75-apps-map && source ../env.sh && go run .\nWhen: different update rules (or idle/teacher) per cam
Where: para.SetBranchModes(...) with parent NormalBP
Why: mix credit, freeze priors, distill, adversarial twins, surprise memory
go run ./01_modes\nExits non-zero if any proof fails (PASS / FAIL lines).
=== 01_modes — prove BranchModes ===\n PASS both_BP both cams moved Δ=0.009232 / 0.009232\n PASS Freeze cam0 moved Δ=0.009794, frozen cam1 Δ=0 (want ~0)\n PASS Shadow teacher frozen Δ=0, student moved Δ=0.1343\n PASS Adversarial both move Δ=0.00832/0.00832; Adv loss 0.0013 ≥ BP loss 0.0007 (fight)\n PASS Memory asleep Δ=0 (~0), awake Δ=0.007141 (>0)\n PASS Tween Tween cam moved Δ=0.003871\n\nall BranchMode proofs passed\n| Proof | \nAssertion | \nWhy that shows the feature | \n
|---|---|---|
| both_BP | \nboth cams ΔW ≫ 0 | \nShared learning works | \n
| Freeze | \ncam0 moves, cam1 Δ≈0 | \nFrozen cam still forwards; no weight update | \n
| Shadow | \nteacher Δ≈0, student moves | \nShadow = frozen KD teacher | \n
| Adversarial | \nboth move; Adv loss ≥ BP loss | \nNegated LR fights the objective | \n
| Memory | \nasleep Δ≈0, awake Δ≫0 | \nSurpriseThresh gates updates | \n
| Tween | \nTween cam Δ≫0 | \nAlternate update family still applies | \n
Cams are diverged first so Freeze isn’t vacuous.
\n","text":"When: different update rules (or idle/teacher) per camWhere: para.SetBranchModes(...) with parent NormalBPWhy: mix credit, freeze priors, distill, adversarial twins, surprise memory\ngo run ./01_modes\n\nExits non-zero if any proof fails (PASS / FAIL lines).\nLive output\n=== 01_modes — prove BranchModes ===\n PASS both_BP both cams moved Δ=0.009232 / 0.009232\n PASS Freeze cam0 moved Δ=0.009794, frozen cam1 Δ=0 (want ~0)\n PASS Shadow teacher frozen Δ=0, student moved Δ=0.1343\n PASS Adversarial both move Δ=0.00832/0.00832; Adv loss 0.0013 ≥ BP loss 0.0007 (fight)\n PASS Memory asleep Δ=0 (~0), awake Δ=0.007141 (>0)\n PASS Tween Tween cam moved Δ=0.003871\n\nall BranchMode proofs passed\n\nWhy these PASS lines prove it\n\n\n\nProof\nAssertion\nWhy that shows the feature\n\n\n\nboth_BP\nboth cams ΔW ≫ 0\nShared learning works\n\n\nFreeze\ncam0 moves, cam1 Δ≈0\nFrozen cam still forwards; no weight update\n\n\nShadow\nteacher Δ≈0, student moves\nShadow = frozen KD teacher\n\n\nAdversarial\nboth move; Adv loss ≥ BP loss\nNegated LR fights the objective\n\n\nMemory\nasleep Δ≈0, awake Δ≫0\nSurpriseThresh gates updates\n\n\nTween\nTween cam Δ≫0\nAlternate update family still applies\n\n\nCams are diverged first so Freeze isn’t vacuous.\n","sources":[{"name":"main.go","code":"package main\n\nimport (\n\t\"fmt\"\n\n\t\"github.com/openfluke/example/cam/internal/harness\"\n\t\"github.com/openfluke/welvet/core\"\n\t\"github.com/openfluke/welvet/layers/parallel\"\n)\n\nfunc main() {\n\tharness.Banner(\"01_modes — prove BranchModes\")\n\tx, y := harness.ToyXY(8, 4)\n\n\tproveBothBP(x, y)\n\tproveFreeze(x, y)\n\tproveShadow(x, y)\n\tproveAdversarial(x, y)\n\tproveMemory(x, y)\n\tproveTween(x, y)\n\tfmt.Println(\"\\nall BranchMode proofs passed\")\n}\n\nfunc proveBothBP(x, y *core.Tensor[float32]) {\n\tpara, _ := harness.DenseTwin(8, 4, parallel.CombineAvg)\n\t_ = harness.DivergeCams(para)\n\tpara.SetBranchModes(parallel.ModeNormalBP, parallel.ModeNormalBP)\n\tw0, w1 := harness.DenseWeights(para, 0), harness.DenseWeights(para, 1)\n\t_, _ = harness.TrainN(para, x, y, 20, 0.1)\n\td0 := harness.WeightMaxDiff(w0, harness.DenseWeights(para, 0))\n\td1 := harness.WeightMaxDiff(w1, harness.DenseWeights(para, 1))\n\tharness.Requiref(\"both_BP\", d0 > 1e-4 && d1 > 1e-4, \"both cams moved Δ=%.4g / %.4g\", d0, d1)\n}\n\nfunc proveFreeze(x, y *core.Tensor[float32]) {\n\tpara, _ := harness.DenseTwin(8, 4, parallel.CombineAvg)\n\t_ = harness.DivergeCams(para)\n\tpara.SetBranchModes(parallel.ModeNormalBP, parallel.ModeFreeze)\n\tw0, w1 := harness.DenseWeights(para, 0), harness.DenseWeights(para, 1)\n\t_, _ = harness.TrainN(para, x, y, 20, 0.1)\n\td0 := harness.WeightMaxDiff(w0, harness.DenseWeights(para, 0))\n\td1 := harness.WeightMaxDiff(w1, harness.DenseWeights(para, 1))\n\tharness.Requiref(\"Freeze\", d0 > 1e-4 && d1 < 1e-7, \"cam0 moved Δ=%.4g, frozen cam1 Δ=%.4g (want ~0)\", d0, d1)\n}\n\nfunc proveShadow(x, y *core.Tensor[float32]) {\n\tpara, _ := harness.DenseTwin(8, 4, parallel.CombineAvg)\n\t_ = harness.DivergeCams(para)\n\tpara.SetBranchModes(parallel.ModeNormalBP, parallel.ModeShadow)\n\tpara.SetCamKit(parallel.CamKit{ShadowCoef: 1.0})\n\tteacherBefore := harness.DenseWeights(para, 1)\n\tstudentBefore := harness.DenseWeights(para, 0)\n\t_, _ = harness.TrainN(para, x, y, 25, 0.08)\n\tdT := harness.WeightMaxDiff(teacherBefore, harness.DenseWeights(para, 1))\n\tdS := harness.WeightMaxDiff(studentBefore, harness.DenseWeights(para, 0))\n\tharness.Requiref(\"Shadow\", dT < 1e-7 && dS > 1e-4, \"teacher frozen Δ=%.4g, student moved Δ=%.4g\", dT, dS)\n}\n\nfunc proveAdversarial(x, y *core.Tensor[float32]) {\n\t// Compare loss: BP∥BP should beat BP∥Adv on same init path.\n\tmk := func(modes ...parallel.TrainMode) float64 {\n\t\tpara, _ := harness.DenseTwin(8, 4, parallel.CombineAvg)\n\t\t// same diverge every time\n\t\t_ = harness.DivergeCams(para)\n\t\tpara.SetBranchModes(modes...)\n\t\tloss, _ := harness.TrainN(para, x, y, 30, 0.08)\n\t\treturn loss\n\t}\n\tlossBP := mk(parallel.ModeNormalBP, parallel.ModeNormalBP)\n\tlossAdv := mk(parallel.ModeNormalBP, parallel.ModeAdversarial)\n\tpara, _ := harness.DenseTwin(8, 4, parallel.CombineAvg)\n\t_ = harness.DivergeCams(para)\n\tpara.SetBranchModes(parallel.ModeNormalBP, parallel.ModeAdversarial)\n\tw0, w1 := harness.DenseWeights(para, 0), harness.DenseWeights(para, 1)\n\t_, _ = harness.TrainN(para, x, y, 20, 0.08)\n\td0 := harness.WeightMaxDiff(w0, harness.DenseWeights(para, 0))\n\td1 := harness.WeightMaxDiff(w1, harness.DenseWeights(para, 1))\n\tharness.Requiref(\"Adversarial\", d0 > 1e-4 && d1 > 1e-4 && lossAdv > lossBP*0.95,\n\t\t\"both move Δ=%.3g/%.3g; Adv loss %.4f ≥ BP loss %.4f (fight)\", d0, d1, lossAdv, lossBP)\n}\n\nfunc proveMemory(x, y *core.Tensor[float32]) {\n\t// High thresh → asleep (no cam1 update). Low thresh → awake (cam1 updates).\n\tasleep := func() float64 {\n\t\tpara, _ := harness.DenseTwin(8, 4, parallel.CombineAvg)\n\t\t_ = harness.DivergeCams(para)\n\t\tpara.SetBranchModes(parallel.ModeNormalBP, parallel.ModeMemory)\n\t\tpara.SetCamKit(parallel.CamKit{SurpriseThresh: 1e9})\n\t\tw1 := harness.DenseWeights(para, 1)\n\t\t_, _ = harness.TrainN(para, x, y, 15, 0.1)\n\t\treturn harness.WeightMaxDiff(w1, harness.DenseWeights(para, 1))\n\t}\n\tawake := func() float64 {\n\t\tpara, _ := harness.DenseTwin(8, 4, parallel.CombineAvg)\n\t\t_ = harness.DivergeCams(para)\n\t\tpara.SetBranchModes(parallel.ModeNormalBP, parallel.ModeMemory)\n\t\tpara.SetCamKit(parallel.CamKit{SurpriseThresh: 1e-12}) // always surprised\n\t\tw1 := harness.DenseWeights(para, 1)\n\t\t_, _ = harness.TrainN(para, x, y, 15, 0.1)\n\t\treturn harness.WeightMaxDiff(w1, harness.DenseWeights(para, 1))\n\t}\n\tdSleep, dWake := asleep(), awake()\n\tharness.Requiref(\"Memory\", dSleep < 1e-7 && dWake > 1e-4,\n\t\t\"asleep Δ=%.4g (~0), awake Δ=%.4g (>0)\", dSleep, dWake)\n}\n\nfunc proveTween(x, y *core.Tensor[float32]) {\n\tpara, _ := harness.DenseTwin(8, 4, parallel.CombineAvg)\n\t_ = harness.DivergeCams(para)\n\tpara.SetBranchModes(parallel.ModeNormalBP, parallel.ModeTween)\n\tw1 := harness.DenseWeights(para, 1)\n\t_, _ = harness.TrainN(para, x, y, 20, 0.08)\n\td1 := harness.WeightMaxDiff(w1, harness.DenseWeights(para, 1))\n\tharness.Requiref(\"Tween\", d1 > 1e-5, \"Tween cam moved Δ=%.4g\", d1)\n}\n","url":"/source/examples/cam/01_modes/main.go"}],"source":"https://github.com/openfluke/example/tree/main/cam/01_modes","url":"https://openfluke.com/examples/cam/01_modes"},{"slug":"cam/02_combine","title":"02 — Combine modes","group":"Cameral cookbook","description":"When: choosing how hemisphere posts become one tensorWhere: Config.CombineWhy: avg/add ensemble; max WTA; sparsek; disagree; filter MoE; concat widths go run ./02_combine Exits non-zero if a","html":"When: choosing how hemisphere posts become one tensor
Where: Config.Combine
Why: avg/add ensemble; max WTA; sparsek; disagree; filter MoE; concat widths
go run ./02_combine\nExits non-zero if any proof fails (PASS / FAIL lines).
=== 02_combine — prove merge modes ===\n PASS avg_width outShape=[1 4] want [1 4]\n PASS add_width outShape=[1 4] want [1 4]\n PASS concat_width outShape=[1 8] want feat=8\n PASS max_routes_to_cam0 combined should match cam0 when cam0≫cam1\n PASS sparsek_keeps_strong SparseK=1 should ≈ strongest cam (cam0)\n PASS disagree_formula y = mean + β(a-b) holds\n PASS filter_gate_trains MoE gate Δ=0.01582\n\nall Combine proofs passed\n| Proof | \nAssertion | \nWhy | \n
|---|---|---|
| avg/add_width | \nout feat = 4 | \nSame-width merge | \n
| concat_width | \nout feat = 8 | \nUnequal stack = 4+4 | \n
| max_routes_to_cam0 | \ncombined ≡ cam0 when cam0≫cam1 | \nHard max routing | \n
| sparsek_keeps_strong | \nSparseK=1 ≈ strongest cam | \nTop-K by ‖out‖₂ | \n
| disagree_formula | \ny = mean + β(a−b) numerically | \nDebate combine is exact | \n
| filter_gate_trains | \nMoE gate weights move | \nGate is live, not decoration | \n
When: cams should share / pull weight DNA
Where: SetCamSync / SyncNow / Cross
Why: soft↔hard consensus, one-way teacher, groups, cross-layer
go run ./03_camsync\nExits non-zero if any proof fails (PASS / FAIL lines).
=== 03_camsync — prove sync pulls cos up ===\n PASS α=1_hard_sync cos 0.3769 → 1.0000 (hard sync should ≈1 and beat pre)\n PASS α=0.25_soft_pull cos 0.3717→0.6413, still distinct ΔW=0.75\n PASS BranchAlpha_one_way cam1 unchanged Δ=0; cos 0.3717→0.6635 (cam0 pulled to mean)\n PASS Groups_exclude_cam2 cam2 frozen Δ=0; cam0≡cam1 Δ=0; cam0≠cam2 Δ=3 (was 3)\n PASS Cross_pair Cross hard-sync cos 0.0000→1.0000\n PASS Cross_stack_trains TrainStackMSE err=<nil>\n\nall CamSync proofs passed\n| Proof | \nAssertion | \nWhy | \n
|---|---|---|
| α=1_hard_sync | \ncos 0.37 → 1.0 | \nDiverged cams hard-average to identical | \n
| α=0.25_soft_pull | \ncos rises, weights still distinct | \nSoft pull ≠ clone | \n
| BranchAlpha_one_way | \ncam1 Δ=0, cos rises | \nTeacher not overwritten; student pulls to mean | \n
| Groups_exclude_cam2 | \ncam0≡cam1, cam2 frozen & different | \nClique sync only | \n
| Cross_pair | \nCross hard-sync cos → 1 | \nStack endpoint glue works | \n
| Cross_stack_trains | \nTrainStackMSE ok | \nTraining still works with Cross cfg | \n
When: schedules, plasticity, replay, speciation
Where: SetBranchLRs / SetRotateSchedule / SetCamKit / DreamPulse
Why: sleep cycles, soft freeze, DNA push/pull, offline consolidation
go run ./04_kit\nExits non-zero if any proof fails (PASS / FAIL lines).
=== 04_kit — prove LRs / rotate / DNA / dream ===\n PASS BranchLRs_1|0 cam0 Δ=0.009794 cam1 Δ=0\n PASS Rotate_slot0 first 5: cam0 Δ=0.002567 cam1 Δ=0\n PASS Rotate_slot1 next 5: cam0 Δ=0 cam1 Δ=0.002486\n PASS DNAReg_diversify diversify cos 0.8446 → -0.8649\n PASS DNAReg_attract attract cos 0.3717 → 0.9994\n PASS DreamPulse replay avgLoss=0.0012 moved cam0 Δ=0.0009936\n PASS RefreshMetrics modes=[NormalBP Freeze] plast=[0.000660944211525738 0]\n\nall CamKit proofs passed\n| Proof | \nAssertion | \nWhy | \n
|---|---|---|
| BranchLRs 1|0 | \ncam1 Δ=0 | \nLR×0 = soft freeze | \n
| Rotate_slot0/1 | \nonly cam0 then only cam1 moves | \nSleep schedule flips plasticity | \n
| DNAReg_diversify | \ncos 0.84 → negative | \nPush away from mean | \n
| DNAReg_attract | \ncos 0.37 → ~1 | \nPull toward mean | \n
| DreamPulse | \nreplay loss>0 and ΔW>0 | \nBuffer replay moves weights | \n
| RefreshMetrics | \nmodes show Freeze, plast[1]=0 | \nMeter matches reality | \n
When: CNN / Mamba / LSTM / … as Parallel cams
Where: HemispheresFrom + twin Ops
Why: CamSync / Freeze / Train work on every hosted Op
go run ./05_layers\nExits non-zero if any proof fails (PASS / FAIL lines).
=== 05_layers — prove each Op works as cams ===\n PASS dense default host loss 0.1230→0.0789 Freeze ok\n PASS cnn1 1D conv loss 0.1181→0.0854 Freeze ok\n PASS cnn2 2D vision loss 0.1161→0.0602 Freeze ok\n PASS cnn3 3D loss 0.1155→0.0271 Freeze ok\n PASS convt1 1D upsample loss 0.1206→0.1148 Freeze ok\n PASS convt2 2D upsample loss 0.1201→0.1077 Freeze ok\n PASS convt3 3D upsample loss 0.1188→0.1053 Freeze ok\n PASS mha attention loss 0.1217→0.1195 Freeze ok\n PASS swiglu FFN loss 0.1230→0.1223 Freeze ok\n PASS rmsnorm norm loss 0.4835→0.0027 Freeze ok\n PASS layernorm norm loss 1.1225→0.0003 Freeze ok\n PASS softmax distribution loss 0.0708→0.0708 Freeze ok\n PASS rnn temporal loss 0.1225→0.0009 Freeze ok\n PASS lstm gated temporal loss 0.1225→0.0741 Freeze ok\n PASS embedding tables loss 0.1252→0.0100 Freeze ok\n PASS sequential deep hemi loss 0.1218→0.1203 Freeze ok\n PASS residual residual hemi loss 0.0431→0.0277 Freeze ok\n PASS kmeans prototypes loss 0.0003→0.0003 Freeze ok\n PASS mamba SSM loss 0.1208→0.0996 Freeze ok\n PASS metacognition observer loss 0.1225→0.0282 Freeze ok\n PASS gdn gated delta loss 0.1225→0.1225 Freeze ok\n\nall layer-as-cam proofs passed\nEach layer must:
\nsoftmax, tiny kmeans, zero-blob gdn (Forward+Freeze still required) ActiveModes[1]==Freeze and primary store Δ≈0PASS … loss a→b Freeze ok means all three held. CNN/MHA/etc. are seeded (Welvet New is zero-init).
When: 1D temporal/spatial sense
\nWhere: go run ./05_layers -layer cnn1
Why: Audio/seq patches; sync via Proj.
\nSetup in the smoke: identical twins + CombineAvg + BranchModes=NormalBP∥Freeze + CamSync α=1%.
ok cnn1 1D conv senses loss=0.0625 plast=[0 0]\nTwin CNN1 forwards+trains. Primary-store plast meter often ~0 on short smoke; graph is live (loss computed). Sync via Proj.
\nCam1 plasticity is 0 on purpose (ModeFreeze). See README.md for the full matrix.
When: 2D vision cam
\nWhere: go run ./05_layers -layer cnn2
Why: MNIST/CIFAR-style; put CNN inside Parallel for CamSync.
\nSetup in the smoke: identical twins + CombineAvg + BranchModes=NormalBP∥Freeze + CamSync α=1%.
ok cnn2 2D vision cams loss=0.0625 plast=[0 0]\nMNIST-style twin. Same plast-meter caveat as cnn1; use longer runs / Acc for learning proof.
\nCam1 plasticity is 0 on purpose (ModeFreeze). See README.md for the full matrix.
When: 3D / volumetric
\nWhere: go run ./05_layers -layer cnn3
Why: Medical/video volumes as twin senses.
\nSetup in the smoke: identical twins + CombineAvg + BranchModes=NormalBP∥Freeze + CamSync α=1%.
ok cnn3 3D cams loss=0.0625 plast=[0 0]\nVolumetric twin smoke OK.
\nCam1 plasticity is 0 on purpose (ModeFreeze). See README.md for the full matrix.
When: 1D generator twin
\nWhere: go run ./05_layers -layer convt1
Why: Upsample hemispheres.
\nSetup in the smoke: identical twins + CombineAvg + BranchModes=NormalBP∥Freeze + CamSync α=1%.
ok convt1 1D upsample loss=0.0625 plast=[0 0]\nGenerator-style twin; TrainMSE path OK.
\nCam1 plasticity is 0 on purpose (ModeFreeze). See README.md for the full matrix.
When: 2D generator twin
\nWhere: go run ./05_layers -layer convt2
Why: Decoder cams.
\nSetup in the smoke: identical twins + CombineAvg + BranchModes=NormalBP∥Freeze + CamSync α=1%.
ok convt2 2D upsample loss=0.0625 plast=[0 0]\n2D transpose twin OK.
\nCam1 plasticity is 0 on purpose (ModeFreeze). See README.md for the full matrix.
When: 3D generator twin
\nWhere: go run ./05_layers -layer convt3
Why: Volumetric decode.
\nSetup in the smoke: identical twins + CombineAvg + BranchModes=NormalBP∥Freeze + CamSync α=1%.
ok convt3 3D upsample loss=0.0625 plast=[0 0]\n3D transpose twin OK.
\nCam1 plasticity is 0 on purpose (ModeFreeze). See README.md for the full matrix.
When: Default cam host
\nWhere: go run ./05_layers -layer dense
Why: Any BranchModes / CamSync experiment; fastest.
\nSetup in the smoke: identical twins + CombineAvg + BranchModes=NormalBP∥Freeze + CamSync α=1%.
ok dense default host loss=0.0623 plast=[0.0012478139251470566 0]\nDefault BranchModes host. Cam0 primary Dense store moves; cam1 Freeze → plast 0.
\nCam1 plasticity is 0 on purpose (ModeFreeze). See README.md for the full matrix.
When: Dual embedding tables
\nWhere: go run ./05_layers -layer embedding
Why: Two vocab views; Freeze one as prior.
\nSetup in the smoke: identical twins + CombineAvg + BranchModes=NormalBP∥Freeze + CamSync α=1%.
ok embedding dual tables loss=0.0619 plast=[0.00043333118975077967 0]\nToken IDs in; cam0 embedding rows move (plast>0), Freeze cam1 idle.
\nCam1 plasticity is 0 on purpose (ModeFreeze). See README.md for the full matrix.
When: Gated Δ-net twin
\nWhere: go run ./05_layers -layer gdn
Why: Modern seq cam alternative to Mamba.
\nSetup in the smoke: identical twins + CombineAvg + BranchModes=NormalBP∥Freeze + CamSync α=1%.
ok gdn gated delta twins loss=0.0625 plast=[0 0]\nGated Δ-net twin smoke OK.
\nCam1 plasticity is 0 on purpose (ModeFreeze). See README.md for the full matrix.
When: Prototype cam
\nWhere: go run ./05_layers -layer kmeans
Why: Cluster / codebook twins.
\nSetup in the smoke: identical twins + CombineAvg + BranchModes=NormalBP∥Freeze + CamSync α=1%.
ok kmeans prototype cams loss=0.0069 plast=[0 0]\nLowest loss — cluster probs sit near constant target quickly.
\nCam1 plasticity is 0 on purpose (ModeFreeze). See README.md for the full matrix.
When: Norm as cam (rare)
\nWhere: go run ./05_layers -layer layernorm
Why: Same — prefer inside a deep hemisphere.
\nSetup in the smoke: identical twins + CombineAvg + BranchModes=NormalBP∥Freeze + CamSync α=1%.
ok layernorm norm cam loss=0.8454 plast=[0.02069561693168054 0]\nSame story as RMS — affine params plastic on cam0 only.
\nCam1 plasticity is 0 on purpose (ModeFreeze). See README.md for the full matrix.
When: Gated temporal cam
\nWhere: go run ./05_layers -layer lstm
Why: Richer memory than RNN.
\nSetup in the smoke: identical twins + CombineAvg + BranchModes=NormalBP∥Freeze + CamSync α=1%.
ok lstm gated temporal loss=0.0620 plast=[0 0]\nLSTM twin smoke OK.
\nCam1 plasticity is 0 on purpose (ModeFreeze). See README.md for the full matrix.
When: SSM twin
\nWhere: go run ./05_layers -layer mamba
Why: Long-seq state-space cams.
\nSetup in the smoke: identical twins + CombineAvg + BranchModes=NormalBP∥Freeze + CamSync α=1%.
ok mamba SSM twins loss=0.0625 plast=[0 0]\nState-space twin smoke OK.
\nCam1 plasticity is 0 on purpose (ModeFreeze). See README.md for the full matrix.
When: Observer twin
\nWhere: go run ./05_layers -layer metacognition
Why: Meta Dense observer.
\nSetup in the smoke: identical twins + CombineAvg + BranchModes=NormalBP∥Freeze + CamSync α=1%.
ok metacognition observer twin loss=0.0625 plast=[0 0]\nObserver Dense twin smoke OK.
\nCam1 plasticity is 0 on purpose (ModeFreeze). See README.md for the full matrix.
When: Attention hemisphere
\nWhere: go run ./05_layers -layer mha
Why: Same DModel; seq cams.
\nSetup in the smoke: identical twins + CombineAvg + BranchModes=NormalBP∥Freeze + CamSync α=1%.
ok mha attention hemispheres loss=0.0625 plast=[0 0]\nSame DModel twins; seq input. Short smoke may not move metered store.
\nCam1 plasticity is 0 on purpose (ModeFreeze). See README.md for the full matrix.
When: Residual hemisphere
\nWhere: go run ./05_layers -layer residual
Why: Skip-connected cam.
\nSetup in the smoke: identical twins + CombineAvg + BranchModes=NormalBP∥Freeze + CamSync α=1%.
ok residual residual cam loss=0.0625 plast=[0 0]\nSkip-connected hemisphere twin.
\nCam1 plasticity is 0 on purpose (ModeFreeze). See README.md for the full matrix.
When: Norm as cam (rare)
\nWhere: go run ./05_layers -layer rmsnorm
Why: Usually nest inside Sequential cam.
\nSetup in the smoke: identical twins + CombineAvg + BranchModes=NormalBP∥Freeze + CamSync α=1%.
ok rmsnorm norm cam loss=0.5227 plast=[0.016977749851921242 0]\nHigher MSE: norm posts ≠ flat 0.25 target. Cam0 γ moves (plast>0); Freeze cam1 idle.
\nCam1 plasticity is 0 on purpose (ModeFreeze). See README.md for the full matrix.
When: Temporal cam
\nWhere: go run ./05_layers -layer rnn
Why: Short seq twins.
\nSetup in the smoke: identical twins + CombineAvg + BranchModes=NormalBP∥Freeze + CamSync α=1%.
ok rnn temporal cams loss=0.0572 plast=[0 0]\nSeq twin OK; loss slightly below the 0.0625 plateau.
\nCam1 plasticity is 0 on purpose (ModeFreeze). See README.md for the full matrix.
When: Deep hemisphere
\nWhere: go run ./05_layers -layer sequential
Why: Whole MLP stack = one mind.
\nSetup in the smoke: identical twins + CombineAvg + BranchModes=NormalBP∥Freeze + CamSync α=1%.
ok sequential deep hemisphere loss=0.0625 plast=[0 0]\nWhole MLP stack = one cam mind.
\nCam1 plasticity is 0 on purpose (ModeFreeze). See README.md for the full matrix.
When: Distribution cam
\nWhere: go run ./05_layers -layer softmax
Why: Avg/max of probability views.
\nSetup in the smoke: identical twins + CombineAvg + BranchModes=NormalBP∥Freeze + CamSync α=1%.
ok softmax distribution cams loss=0.0169 plast=[0 0]\nSoftmax already in (0,1) → easy MSE to 0.25. Softmax itself has no weights → plast 0 is expected.
\nCam1 plasticity is 0 on purpose (ModeFreeze). See README.md for the full matrix.
When: FFN twin
\nWhere: go run ./05_layers -layer swiglu
Why: Width=InputDim; pair with MHA stacks.
\nSetup in the smoke: identical twins + CombineAvg + BranchModes=NormalBP∥Freeze + CamSync α=1%.
ok swiglu FFN twins loss=0.0625 plast=[0 0]\nWidth=InputDim FFN cams.
\nCam1 plasticity is 0 on purpose (ModeFreeze). See README.md for the full matrix.
When: compose modes + sync + kit into behaviors
Where: go run ./06_recipes
Why: teacher, sleep, debate, memory, dream, concat
go run ./06_recipes\nExits non-zero if any proof fails (PASS / FAIL lines).
=== 06_recipes — prove composed behaviors ===\n PASS teacher_student teacher frozen, student moved\n PASS sleep_slot0 only cam0 Δ=0.00257/0\n PASS sleep_slot1 only cam1 Δ=0/0.00249\n PASS debate both move; debate loss 0.0121 vs coop 0.0110\n PASS surprise_memory asleep Δ=0 awake Δ=0.00582\n PASS dream_consolidate avg=0.0012 weights moved on replay\n PASS cross_modal_concat feat=8 want 3+5=8\n PASS cross_modal_trains cam0 moved under concat\n\nall recipe proofs passed\n| Proof | \nAssertion | \nWhy | \n
|---|---|---|
| teacher_student | \nteacher frozen, student moved | \nShadow + one-way sync | \n
| sleep_slot0/1 | \nalternating single-cam ΔW | \nRotate schedule | \n
| debate | \nboth move; debate loss ≥ coop | \nAdv + Disagree | \n
| surprise_memory | \nasleep vs awake ΔW | \nSurpriseThresh gate | \n
| dream_consolidate | \nreplay moves weights | \nDreamBuffer | \n
| cross_modal_concat | \nfeat=8=3+5; cam0 trains | \nUnequal cams need concat | \n
\n\nA living AI world where biology, computation, platforming, learning, and buildcraft become one persistent artificial life.
\n
ALife is a 3D radial-gravity platforming game and on-device artificial-life system built around three characters:
\nThe game takes inspiration from the compact, playful traversal of Super Mario Galaxy-style planetary platforming and the long-term buildcraft/endgame loop of action RPGs such as Diablo, but the progression model is built around capabilities rather than character levels.
\nThe central loop is:
\n\n\nExplore → discover → traverse → cleanse → unlock → build → master → automate → push further
\n
ALife is divided into multiverses.
\nEach multiverse contains seven planets.
\nThe important distinction is that the planets are not simply seven conventional levels.
\nA campaign progresses roughly like this:
\nPlanet 1\n │\n ├── Planet-bridging platforming route\n ▼\nPlanet 2\n │\n ├── Planet-bridging platforming route\n ▼\nPlanet 3\n │\n ├── Planet-bridging platforming route\n ▼\nPlanet 4\n │\n ├── Planet-bridging platforming route\n ▼\nPlanet 5\n │\n ├── Planet-bridging platforming route\n ▼\nPlanet 6\n │\n ├── Final planet-bridging challenge\n ▼\nPlanet 7 / Boss\nThe player can see and travel through a physical planetary system rather than selecting isolated stages from a menu.
\nThe actual authored platforming levels are the journeys between planets.
\nThese routes use ALife's radial-gravity platforming system and can contain:
\nA \"planet\" does not need a complete spherical walkable surface. A level can use only the curved land required for traversal while still producing the sensation of moving around a planetary body.
\nThis allows the campaign to push radial platforming far beyond a single conventional spherical world.
\nA normal planet-bridging challenge should generally be compact:
\n\n\nApproximately three minutes for a successful run.
\n
The goal is dense mechanical variety rather than long stretches of filler.
\nPlanets act as persistent campaign nodes.
\nThey can provide:
\nThe rhythm becomes:
\nARRIVE\n ↓\nEXPLORE\n ↓\nDISCOVER\n ↓\nCONFIGURE\n ↓\nDEPART\n ↓\nPLATFORMING ROUTE\n ↓\nARRIVE AT NEXT PLANET\nThis makes reaching another planet feel like an actual journey rather than completing an arbitrary numbered stage.
\nEach section of the campaign expands three major axes of the player's capability space.
\nProto is the primordial side of ALife.
\nProto discovers animals and can copy their morphology, instincts, or biological properties.
\nExamples:
\n| Animal | \nPossible capability | \n
|---|---|
| Gecko | \nWall adhesion | \n
| Bird | \nGlide / aerial movement | \n
| Electric eel | \nElectrical cleansing | \n
| Mole | \nBurrowing | \n
| Frog | \nExtreme jumping | \n
| Octopus | \nMulti-surface gripping | \n
Proto is particularly important for cleansing infection.
\nDifferent animals can provide fundamentally different ways of removing corruption from robots, machinery, or environments.
\nProto answers:
\n\n\nWhat biological capability can I use?
\n
Kee represents the robotic and mechanical side.
\nKee can become or provide physical gadgets such as:
\nKee's gadget answers:
\n\n\nWhat physical/mechanical interface can I use?
\n
A gadget and an algorithm are deliberately separate concepts.
\nThe gadget is the body/tool.
\nThe algorithm is the logic.
\nAlgorithms in ALife are intended to be based on actual computer-science algorithms, not merely fictional abilities with technical names.
\nThe broader design has a large library of possible algorithms to draw from.
\nExamples include:
\nAlgorithms become gameplay logic.
\nExamples:
\nFind or expose an optimal path through a traversal problem.
\nPropagate an effect outward through neighbouring nodes in layers.
\nFind a route according to minimum cost rather than simply minimum distance.
\nPropagate cleansing through a connected infected region.
\nDynamically reorder targets according to distance, danger, corruption, value, or another property.
\nThe algorithm answers:
\n\n\nHow is the action organised, reasoned about, or applied?
\n
The fundamental build grammar is:
\nProto Animal / Power\n +\nKee Gadget\n +\nAlgorithm\n +\nAttributes\n +\nLegendaries\n =\nPlayer Build\nThis allows the same animal, gadget, or algorithm to behave very differently depending on what it is paired with.
\nThe long-term objective is not simply to become numerically stronger.
\nIt is to accumulate an increasingly large vocabulary of possible behaviours.
\nCampaign content introduces capabilities before the player permanently masters every possible combination.
\nA campaign section can introduce:
\nThe player experiences the intended combination in authored content.
\nCompleting that content establishes that the capability has been discovered.
\nFor algorithms in particular, discovery and permanent endgame ownership are intentionally different stages.
\nAn algorithm can first appear as part of its campaign content.
\nFor example:
\nDiscover A*\n ↓\nUse A* during its authored challenge\n ↓\nComplete the challenge\n ↓\nA* becomes eligible for that multiverse's endgame loot\nThe campaign therefore says:
\n\n\nI discovered and experienced this algorithm.
\n
The endgame says:
\n\n\nI mastered it and can now build around it permanently.
\n
This creates a direct connection between authored campaign content and the later loot chase.
\nALife does not require traditional character levelling.
\nThere does not need to be a central progression of:
\nLevel 1 → Level 20 → Level 50 → Level 100\nInstead:
\nSmall capability library\n ↓\nMore animals\n ↓\nMore gadgets\n ↓\nMore algorithms\n ↓\nMore attributes\n ↓\nMore legendaries\n ↓\nMore combinations\n ↓\nMassive build space\nProgression is primarily horizontal.
\nThe artificial life becomes more versatile and expressive rather than simply acquiring an ever-larger number next to its name.
\nNormal campaign content can produce random attribute rewards.
\nAttributes are the tuning layer.
\nPossible examples include:
\nThe intention is to avoid turning ALife into inventory landfill.
\nInstead of forcing the player to retain hundreds of nearly identical disposable items, useful attributes can become part of the player's persistent build library.
\nConceptually:
\nFind attribute\n ↓\nUnlock attribute\n ↓\nUse it when constructing builds\nPlayers then save configurations they actually care about.
\nPermanent portable algorithms are represented by three-piece sets.
\nExample:
\nA* Piece I\nA* Piece II\nA* Piece III\n ↓\nComplete A* Set\n ↓\nA* becomes available as part of the build\nThe three-piece structure is intentional.
\nAn algorithm should matter enough to occupy meaningful build space without consuming the entire equipment configuration.
\nThis leaves room for legendary items and attribute customisation.
\nThe core principle is:
\n\n\nAlgorithm sets give you the system.
\n
Legendary gear has a different purpose.
\nIt should not simply duplicate the role of algorithm sets or exist only as a larger attribute number.
\nLegendary items are rule benders.
\nThey alter how an existing capability behaves.
\nExamples:
\nThe principle is:
\n\n\nAlgorithm sets give you the system. Legendaries mutate the system. Attributes tune the system.
\n
This separation gives each reward category a clear reason to exist.
\nThe important persistent object is not necessarily a backpack full of equipment.
\nIt is the build.
\nFor example:
\nbuild: Rift Farming\nproto:\n animal: Falcon\n power: Wind Purge\n\nkee:\n gadget: Jetpack\n\nalgorithm:\n set: A*\n\nlegendaries:\n - Momentum Loop\n - Chain Cleanse\n\nattributes:\n - Air Control\n - Energy Efficiency\n - Movement Speed\nAnother saved build might be designed for:
\nProgress therefore becomes a collection of solutions rather than a warehouse of junk.
\nThe campaign ultimately leads across the planetary system to the final planet.
\nThe last traversal should be a culmination of the multiverse's mechanics.
\nPlanet 7 contains the major boss or source of infection.
\nConceptually:
\nPlanet 1\n ↓\nDiscover / cleanse / traverse\n ↓\nPlanet 2\n ↓\nDiscover / cleanse / traverse\n ↓\n...\n ↓\nPlanet 6\n ↓\nFinal traversal\n ↓\nPlanet 7\n ↓\nBOSS\n ↓\nMULTIVERSE CLEANSED\nThe final world can potentially be visible much earlier in the campaign, giving the player a persistent destination to work toward.
\nAs the earlier planets are cleansed, the state of the multiverse can visibly improve.
\nDefeating the multiverse boss changes the role of that entire multiverse.
\nIt is no longer only campaign content.
\nIt becomes an endgame ecosystem.
\nCompletion unlocks:
\nSo every completed multiverse remains relevant after its story is finished.
\nRifts procedurally remix the capabilities and environment of the multiverse.
\nThey can reuse and recombine:
\nImportantly, because the campaign itself is built around travelling between planets, procedural Rifts can also generate increasingly unusual planet-bridging routes.
\nThe authored campaign teaches the player how the multiverse works.
\nRifts remix what was learned.
\nGreater Rifts are the scalable endgame frontier.
\nPlayers use them to:
\nThe game therefore retains an action-RPG farming loop without requiring traditional XP levelling.
\nEach multiverse has its own discoveries and endgame loot pool.
\nFor example, if an earlier multiverse contains the A* algorithm set, a player may return much later because A* interacts beautifully with an animal or gadget discovered several multiverses afterward.
\nNew content therefore expands the usefulness of old content.
\nThe ideal result is:
\n\n\nEvery new multiverse increases the number of meaningful combinations across the entire game.
\n
ALife includes a record-and-replay system for mastered content.
\nThe player first performs the activity manually.
\nThe system can record:
\nThat creates a reusable behavioural/data cache.
\nConceptually:
\nPlayer discovers challenge\n ↓\nPlayer solves challenge manually\n ↓\nPlayer records successful solution\n ↓\nALife replays learned solution\n ↓\nSolved content can be farmed\nThe long-term goal is not merely a dumb input macro.
\nALife is intended to support artificial-life behaviour that can run and learn at the same time, allowing mastered routines to potentially improve or adapt.
\nThe fundamental rule is:
\n\n\nThe player pushes the frontier. ALife farms what has been mastered.
\n
ENTER MULTIVERSE\n ↓\nARRIVE ON PLANET\n ↓\nDISCOVER ANIMAL / PROTO POWER\n ↓\nDISCOVER KEE GADGET\n ↓\nEXPERIENCE ALGORITHM\n ↓\nBUILD A COMBINATION\n ↓\nDEPART PLANET\n ↓\n3D RADIAL PLANET-BRIDGING PLATFORMING\n ↓\nCLEANSE INFECTION\n ↓\nARRIVE AT NEXT PLANET\n ↓\nREPEAT ACROSS THE SYSTEM\n ↓\nFINAL PLANET\n ↓\nBOSS\n ↓\nMULTIVERSE CLEANSED\n ↓\nRIFTS + GREATER RIFTS UNLOCK\n ↓\nFARM 3-PIECE ALGORITHM SETS\n ↓\nFARM LEGENDARIES\n ↓\nTUNE ATTRIBUTES\n ↓\nSAVE BUILDS\n ↓\nPUSH HARDER RIFTS\n ↓\nRECORD MASTERED ROUTES\n ↓\nAUTOPILOT FARMING\n ↓\nENTER THE NEXT MULTIVERSE\n\n\nCampaign content teaches capabilities, algorithm sets make algorithms permanently portable, attributes tune the build, and legendaries break its rules.
\n
\n\nThe seven planets are the chapters; the journeys between them are the levels; the final planet is the boss destination; cleansing the system turns the whole multiverse into an endgame playground.
\n
ALife is ultimately about accumulating possibility.
\nProto discovers nature.
\nKee discovers machinery and computation.
\nThe player combines them.
\nThe campaign teaches those systems through compact radial-gravity platforming journeys between planets.
\nThe endgame takes the systems the player has learned and explodes them into buildcraft, procedural Rifts, Greater Rifts, legendary modifiers, algorithm sets, recorded solutions, and autonomous farming.
\nThere is no requirement for the player to become \"Level 100.\"
\nInstead, the artificial life becomes something much stranger:
\n\n\n","text":"# ALife\n\n\n> **A living AI world where biology, computation, platforming, learning, and buildcraft become one persistent artificial life.**\n\nALife is a 3D radial-gravity platforming game and on-device artificial-life system built around three characters:\n\n- **Proto** — primordial biological life that discovers animal forms and cleansing powers.\n- **Kee** — robotic/computational life that discovers gadgets and real computer-science algorithms.\n- **ALife / You** — the persistent synthesis of the two, able to build lasting combinations from what has been discovered.\n\nThe game takes inspiration from the compact, playful traversal of *Super Mario Galaxy*-style planetary platforming and the long-term buildcraft/endgame loop of action RPGs such as *Diablo*, but the progression model is built around **capabilities rather than character levels**.\n\nThe central loop is:\n\n> **Explore → discover → traverse → cleanse → unlock → build → master → automate → push further**\n\n---\n\n## Core World Structure\n\nALife is divided into **multiverses**.\n\nEach multiverse contains **seven planets**.\n\nThe important distinction is that the planets are **not simply seven conventional levels**.\n\n### The planets are destinations. The space between them is the platformer.\n\nA campaign progresses roughly like this:\n\n```text\nPlanet 1\n │\n ├── Planet-bridging platforming route\n ▼\nPlanet 2\n │\n ├── Planet-bridging platforming route\n ▼\nPlanet 3\n │\n ├── Planet-bridging platforming route\n ▼\nPlanet 4\n │\n ├── Planet-bridging platforming route\n ▼\nPlanet 5\n │\n ├── Planet-bridging platforming route\n ▼\nPlanet 6\n │\n ├── Final planet-bridging challenge\n ▼\nPlanet 7 / Boss\n```\n\nThe player can see and travel through a physical planetary system rather than selecting isolated stages from a menu.\n\n---\n\n## Planet-Bridging Levels\n\nThe actual authored platforming levels are the journeys **between planets**.\n\nThese routes use ALife's radial-gravity platforming system and can contain:\n\n- curved strips of terrain,\n- tiny planets,\n- large planets,\n- partial planetary surfaces,\n- walls that become floors,\n- changing curvature,\n- gravity transitions,\n- floating structures,\n- nested planetary spaces,\n- launch sequences,\n- traversal gadgets,\n- animal abilities,\n- infected machinery,\n- and other dynamically constructed geometry.\n\nA \"planet\" does not need a complete spherical walkable surface. A level can use only the curved land required for traversal while still producing the sensation of moving around a planetary body.\n\nThis allows the campaign to push radial platforming far beyond a single conventional spherical world.\n\n### Target length\n\nA normal planet-bridging challenge should generally be compact:\n\n> **Approximately three minutes for a successful run.**\n\nThe goal is dense mechanical variety rather than long stretches of filler.\n\n---\n\n## What Happens on a Planet?\n\nPlanets act as persistent campaign nodes.\n\nThey can provide:\n\n- discovery,\n- environmental storytelling,\n- animals,\n- Proto powers,\n- Kee technology,\n- algorithms,\n- infected robots,\n- resources,\n- campfires,\n- secrets,\n- build preparation,\n- and the setup for the next journey.\n\nThe rhythm becomes:\n\n```text\nARRIVE\n ↓\nEXPLORE\n ↓\nDISCOVER\n ↓\nCONFIGURE\n ↓\nDEPART\n ↓\nPLATFORMING ROUTE\n ↓\nARRIVE AT NEXT PLANET\n```\n\nThis makes reaching another planet feel like an actual journey rather than completing an arbitrary numbered stage.\n\n---\n\n# The Three-Part Capability System\n\nEach section of the campaign expands three major axes of the player's capability space.\n\n## 1. Proto: Biology\n\nProto is the primordial side of ALife.\n\nProto discovers animals and can copy their morphology, instincts, or biological properties.\n\nExamples:\n\n| Animal | Possible capability |\n|---|---|\n| Gecko | Wall adhesion |\n| Bird | Glide / aerial movement |\n| Electric eel | Electrical cleansing |\n| Mole | Burrowing |\n| Frog | Extreme jumping |\n| Octopus | Multi-surface gripping |\n\nProto is particularly important for **cleansing infection**.\n\nDifferent animals can provide fundamentally different ways of removing corruption from robots, machinery, or environments.\n\nProto answers:\n\n> **What biological capability can I use?**\n\n---\n\n## 2. Kee: Gadgets\n\nKee represents the robotic and mechanical side.\n\nKee can become or provide physical gadgets such as:\n\n- sword,\n- shield,\n- bow,\n- jetpack,\n- grappling system,\n- scanner,\n- launcher,\n- hover system,\n- drill,\n- drone,\n- energy projector,\n- and other mechanical forms.\n\nKee's gadget answers:\n\n> **What physical/mechanical interface can I use?**\n\nA gadget and an algorithm are deliberately separate concepts.\n\nThe gadget is the **body/tool**.\n\nThe algorithm is the **logic**.\n\n---\n\n## 3. Kee: Real Computer-Science Algorithms\n\nAlgorithms in ALife are intended to be based on **actual computer-science algorithms**, not merely fictional abilities with technical names.\n\nThe broader design has a large library of possible algorithms to draw from.\n\nExamples include:\n\n- A*\n- Dijkstra's algorithm\n- Breadth-First Search\n- Depth-First Search\n- Flood Fill\n- sorting algorithms\n- graph algorithms\n- scheduling algorithms\n- optimisation algorithms\n- pathfinding algorithms\n\nAlgorithms become gameplay logic.\n\nExamples:\n\n### A*\n\nFind or expose an optimal path through a traversal problem.\n\n### Breadth-First Search\n\nPropagate an effect outward through neighbouring nodes in layers.\n\n### Dijkstra\n\nFind a route according to minimum cost rather than simply minimum distance.\n\n### Flood Fill\n\nPropagate cleansing through a connected infected region.\n\n### Sorting / prioritisation\n\nDynamically reorder targets according to distance, danger, corruption, value, or another property.\n\nThe algorithm answers:\n\n> **How is the action organised, reasoned about, or applied?**\n\n---\n\n# The Combination\n\nThe fundamental build grammar is:\n\n```text\nProto Animal / Power\n +\nKee Gadget\n +\nAlgorithm\n +\nAttributes\n +\nLegendaries\n =\nPlayer Build\n```\n\nThis allows the same animal, gadget, or algorithm to behave very differently depending on what it is paired with.\n\nThe long-term objective is not simply to become numerically stronger.\n\nIt is to accumulate an increasingly large **vocabulary of possible behaviours**.\n\n---\n\n# Campaign Progression\n\nCampaign content introduces capabilities before the player permanently masters every possible combination.\n\nA campaign section can introduce:\n\n1. a new animal,\n2. a new Proto cleansing capability,\n3. a new Kee gadget,\n4. a computer-science algorithm,\n5. and a platforming challenge designed around their interaction.\n\nThe player experiences the intended combination in authored content.\n\nCompleting that content establishes that the capability has been **discovered**.\n\nFor algorithms in particular, discovery and permanent endgame ownership are intentionally different stages.\n\n---\n\n# Algorithm Discovery vs Mastery\n\nAn algorithm can first appear as part of its campaign content.\n\nFor example:\n\n```text\nDiscover A*\n ↓\nUse A* during its authored challenge\n ↓\nComplete the challenge\n ↓\nA* becomes eligible for that multiverse's endgame loot\n```\n\nThe campaign therefore says:\n\n> **I discovered and experienced this algorithm.**\n\nThe endgame says:\n\n> **I mastered it and can now build around it permanently.**\n\nThis creates a direct connection between authored campaign content and the later loot chase.\n\n---\n\n# Gear Philosophy\n\nALife does **not require traditional character levelling**.\n\nThere does not need to be a central progression of:\n\n```text\nLevel 1 → Level 20 → Level 50 → Level 100\n```\n\nInstead:\n\n```text\nSmall capability library\n ↓\nMore animals\n ↓\nMore gadgets\n ↓\nMore algorithms\n ↓\nMore attributes\n ↓\nMore legendaries\n ↓\nMore combinations\n ↓\nMassive build space\n```\n\nProgression is primarily **horizontal**.\n\nThe artificial life becomes more versatile and expressive rather than simply acquiring an ever-larger number next to its name.\n\n---\n\n# Normal Gear and Attributes\n\nNormal campaign content can produce random attribute rewards.\n\nAttributes are the tuning layer.\n\nPossible examples include:\n\n- movement speed,\n- cleansing radius,\n- cleansing strength,\n- energy efficiency,\n- cooldown reduction,\n- acceleration,\n- air control,\n- gadget recharge,\n- transformation speed,\n- attack speed,\n- defence,\n- resource collection.\n\nThe intention is to avoid turning ALife into inventory landfill.\n\nInstead of forcing the player to retain hundreds of nearly identical disposable items, useful attributes can become part of the player's persistent build library.\n\nConceptually:\n\n```text\nFind attribute\n ↓\nUnlock attribute\n ↓\nUse it when constructing builds\n```\n\nPlayers then save configurations they actually care about.\n\n---\n\n# Algorithm Sets\n\nPermanent portable algorithms are represented by **three-piece sets**.\n\nExample:\n\n```text\nA* Piece I\nA* Piece II\nA* Piece III\n ↓\nComplete A* Set\n ↓\nA* becomes available as part of the build\n```\n\nThe three-piece structure is intentional.\n\nAn algorithm should matter enough to occupy meaningful build space without consuming the entire equipment configuration.\n\nThis leaves room for legendary items and attribute customisation.\n\nThe core principle is:\n\n> **Algorithm sets give you the system.**\n\n---\n\n# Legendary Gear\n\nLegendary gear has a different purpose.\n\nIt should not simply duplicate the role of algorithm sets or exist only as a larger attribute number.\n\nLegendary items are **rule benders**.\n\nThey alter how an existing capability behaves.\n\nExamples:\n\n- A* can recalculate while airborne.\n- Cleansing one target automatically marks another.\n- Using a Kee gadget charges a Proto ability.\n- Using a Proto ability resets a Kee gadget.\n- An algorithm evaluates multiple destinations instead of one.\n- A traversal action gains an unusual interaction with another unlocked system.\n\nThe principle is:\n\n> **Algorithm sets give you the system. Legendaries mutate the system. Attributes tune the system.**\n\nThis separation gives each reward category a clear reason to exist.\n\n---\n\n# Saved Builds\n\nThe important persistent object is not necessarily a backpack full of equipment.\n\nIt is the **build**.\n\nFor example:\n\n```yaml\nbuild: Rift Farming\nproto:\n animal: Falcon\n power: Wind Purge\n\nkee:\n gadget: Jetpack\n\nalgorithm:\n set: A*\n\nlegendaries:\n - Momentum Loop\n - Chain Cleanse\n\nattributes:\n - Air Control\n - Energy Efficiency\n - Movement Speed\n```\n\nAnother saved build might be designed for:\n\n- bosses,\n- traversal,\n- cleansing,\n- multiplayer,\n- speed,\n- resource farming,\n- Greater Rifts,\n- or a particular Autopilot route.\n\nProgress therefore becomes a collection of **solutions** rather than a warehouse of junk.\n\n---\n\n# Multiverse Campaign Completion\n\nThe campaign ultimately leads across the planetary system to the final planet.\n\nThe last traversal should be a culmination of the multiverse's mechanics.\n\nPlanet 7 contains the major boss or source of infection.\n\nConceptually:\n\n```text\nPlanet 1\n ↓\nDiscover / cleanse / traverse\n ↓\nPlanet 2\n ↓\nDiscover / cleanse / traverse\n ↓\n...\n ↓\nPlanet 6\n ↓\nFinal traversal\n ↓\nPlanet 7\n ↓\nBOSS\n ↓\nMULTIVERSE CLEANSED\n```\n\nThe final world can potentially be visible much earlier in the campaign, giving the player a persistent destination to work toward.\n\nAs the earlier planets are cleansed, the state of the multiverse can visibly improve.\n\n---\n\n# Endgame Unlock\n\nDefeating the multiverse boss changes the role of that entire multiverse.\n\nIt is no longer only campaign content.\n\nIt becomes an **endgame ecosystem**.\n\nCompletion unlocks:\n\n- Rifts,\n- Greater Rifts,\n- algorithm-set farming,\n- legendary farming,\n- rarer attributes,\n- build experimentation,\n- recording,\n- Autopilot farming of mastered content.\n\nSo every completed multiverse remains relevant after its story is finished.\n\n---\n\n# Rifts\n\nRifts procedurally remix the capabilities and environment of the multiverse.\n\nThey can reuse and recombine:\n\n- planetary geometry,\n- radial gravity,\n- animals,\n- Proto powers,\n- Kee gadgets,\n- algorithms,\n- infection mechanics,\n- enemies,\n- hazards,\n- and loot families.\n\nImportantly, because the campaign itself is built around travelling between planets, procedural Rifts can also generate increasingly unusual **planet-bridging routes**.\n\nThe authored campaign teaches the player how the multiverse works.\n\nRifts remix what was learned.\n\n---\n\n# Greater Rifts\n\nGreater Rifts are the scalable endgame frontier.\n\nPlayers use them to:\n\n- test builds,\n- hunt algorithm set pieces,\n- hunt legendaries,\n- find better attributes,\n- optimise combinations,\n- push increasingly difficult content,\n- record successful strategies,\n- and eventually automate content they have mastered.\n\nThe game therefore retains an action-RPG farming loop without requiring traditional XP levelling.\n\n---\n\n# Why Old Multiverses Stay Relevant\n\nEach multiverse has its own discoveries and endgame loot pool.\n\nFor example, if an earlier multiverse contains the A* algorithm set, a player may return much later because A* interacts beautifully with an animal or gadget discovered several multiverses afterward.\n\nNew content therefore expands the usefulness of old content.\n\nThe ideal result is:\n\n> **Every new multiverse increases the number of meaningful combinations across the entire game.**\n\n---\n\n# Recording and Autopilot\n\nALife includes a record-and-replay system for mastered content.\n\nThe player first performs the activity manually.\n\nThe system can record:\n\n- actions,\n- traversal,\n- ability use,\n- gadget use,\n- transformations,\n- interactions,\n- and the build used for the run.\n\nThat creates a reusable behavioural/data cache.\n\nConceptually:\n\n```text\nPlayer discovers challenge\n ↓\nPlayer solves challenge manually\n ↓\nPlayer records successful solution\n ↓\nALife replays learned solution\n ↓\nSolved content can be farmed\n```\n\nThe long-term goal is not merely a dumb input macro.\n\nALife is intended to support artificial-life behaviour that can **run and learn at the same time**, allowing mastered routines to potentially improve or adapt.\n\nThe fundamental rule is:\n\n> **The player pushes the frontier. ALife farms what has been mastered.**\n\n---\n\n# Full Gameplay Loop\n\n```text\nENTER MULTIVERSE\n ↓\nARRIVE ON PLANET\n ↓\nDISCOVER ANIMAL / PROTO POWER\n ↓\nDISCOVER KEE GADGET\n ↓\nEXPERIENCE ALGORITHM\n ↓\nBUILD A COMBINATION\n ↓\nDEPART PLANET\n ↓\n3D RADIAL PLANET-BRIDGING PLATFORMING\n ↓\nCLEANSE INFECTION\n ↓\nARRIVE AT NEXT PLANET\n ↓\nREPEAT ACROSS THE SYSTEM\n ↓\nFINAL PLANET\n ↓\nBOSS\n ↓\nMULTIVERSE CLEANSED\n ↓\nRIFTS + GREATER RIFTS UNLOCK\n ↓\nFARM 3-PIECE ALGORITHM SETS\n ↓\nFARM LEGENDARIES\n ↓\nTUNE ATTRIBUTES\n ↓\nSAVE BUILDS\n ↓\nPUSH HARDER RIFTS\n ↓\nRECORD MASTERED ROUTES\n ↓\nAUTOPILOT FARMING\n ↓\nENTER THE NEXT MULTIVERSE\n```\n\n---\n\n# Progression in One Sentence\n\n> **Campaign content teaches capabilities, algorithm sets make algorithms permanently portable, attributes tune the build, and legendaries break its rules.**\n\n---\n\n# World Structure in One Sentence\n\n> **The seven planets are the chapters; the journeys between them are the levels; the final planet is the boss destination; cleansing the system turns the whole multiverse into an endgame playground.**\n\n---\n\n# ALife's Design Identity\n\nALife is ultimately about accumulating **possibility**.\n\nProto discovers nature.\n\nKee discovers machinery and computation.\n\nThe player combines them.\n\nThe campaign teaches those systems through compact radial-gravity platforming journeys between planets.\n\nThe endgame takes the systems the player has learned and explodes them into buildcraft, procedural Rifts, Greater Rifts, legendary modifiers, algorithm sets, recorded solutions, and autonomous farming.\n\nThere is no requirement for the player to become \"Level 100.\"\n\nInstead, the artificial life becomes something much stranger:\n\n> **A growing library of forms, tools, algorithms, behaviours, builds, and learned solutions that belongs to the same persistent being across every multiverse.**\n"}}A growing library of forms, tools, algorithms, behaviours, builds, and learned solutions that belongs to the same persistent being across every multiverse.
\n