NEURALESE: interpretability and weight-surgery lab (TinyStories-8M, runs in the browser; includes vendored Tsotchke WASM components)

Live: https://0thernes.github.io/neuralese/ · New here? Beginner's guide to steering · Audit: AUDIT.md · Tsotchke analysis: docs/TSOTCHKE.md · Claims & evidence: docs/CLAIMS.md · Stack and versions: docs/STACK.md

v17.3.0: Weight-Machine (4 Nov 2026)

A board of weight gadgets: a Forge CPU, attention RAM, a paged SSD, a 16-lane GPU, an 8×8 systolic TPU, and a bus. An int32 reference is the ground truth. The CMOS zoom is simulated. Machine Room is a workspace with no Alt digit. Write-up: docs/WEIGHT_MACHINE.md. Release notes: RELEASE_NOTES.md.

v17.2.0: Forge CPU and Deep Walk (3 Nov 2026)

Forge clocks a small original CPU from NestLang. An int32 VM is the ground truth. A float32 gadget engine matches it on the shipped suite. An MDCC lane answers the same arithmetic with host bigint and spends far fewer tokens. Exact is not the same as the right program, and bf16 parity is not claimed. Write-up: docs/NEST_CPU.md.

Deep Walk steps through one real TinyStories-8M forward as blocks of numbers, then a gate netlist and a simulated CMOS circuit. The voltage view does not undervolt the machine. A pre-registered activation bit-flip on layer 3 moved the next-token distribution more for an exponent bit than for a mantissa bit, and did not change the greedy token. Write-up: docs/DEEP_WALKTHROUGH.md.

v17.1.0: Adversary Lab (2 Nov 2026)

A tenth workspace for key-cache geometry (the Lyra Technique, on our small models) and an MLP-unit search with a steering-versus-random rule. It reports leakage, null, or above-chance from a locked pre-registration. On TinyStories-8M the shipped run is leakage, and the neuron rule fails. Method, citations, and that measurement: docs/ADVERSARY.md. Release steps still open are listed in RELEASE_NOTES.md.

v17.0.0: new front-end shell (1 Nov 2026)

Nine workspaces replace the 75-tab strip (all 75 tools kept, see PARITY.md), dockable/splittable panels with saved layouts, Ctrl/⌘+K command palette, a typed selection store, typed worker RPC, a 3D model cosmos and 3D nested-experiment views (three.js WebGPU with WebGL2 fallback), and a notebook/report mode tied to pre-registration. Shell source is in ui/ (Solid + TypeScript + Vite + Tailwind v4; cd ui && bun install && bun run build writes src/shell/). Details in RELEASE_NOTES.md.

v16.0.1: arch interface, WebGPU, HUB interpretability (2 Nov 2026)

• One forward for GPT-Neo/GPT-2/Llama/Qwen2 matching HF within 1e-4; an opt-in WGSL WebGPU path behind a logits-match gate (slower than JS on the software adapter we could test); a logit/embedding lens, rollout and probes on ViT, CLIP vision and the Whisper encoder (H-INTERP tab); enforced pre-registration with REPRO bundles; exact dependency pins and hash-pinned models. The APK is now release-signed, and the desktop builds are unsigned. Details and caveats: RELEASE_NOTES.md.

v15.3.0: engineering pass (1 Nov 2026)

• Prefix KV cache + independent KV equality test; generation in a Web Worker (main-thread fallback); CSP without unsafe-eval; SwiftShader only as a GPU-failure fallback; source moved to src/, feature-named tests, CI lint+test+smoke. Details and caveats in RELEASE_NOTES.md.

v15.2.0: refinement pass (31 Oct 2026)

• Sidebar labels wrap with full-title tooltips (no truncation); wider rail.
• Neuron Lens atlas: skeleton/shimmer placeholder with live load progress and a retry message on failure.
• 3D: GPU tiering. Software and weak GPUs start at a light quality level, and the adaptive level persists across sessions.
• Busy spinners and inline error notes on every action; the theme follows the OS on first run; mobile touch targets and inputs enlarged.

v15.1.0: polish pass (31 Oct 2026)

• Overview landing dashboard: a hero 3D network coloured by live head activity, live model stats and a last-generation surprisal strip.
• 3D: instanced heads, activation flow particles, hover info cards, fly-to camera, and bloom + depth of field with adaptive quality and an fps readout.
• Charts: animated transitions, brush zoom, cross-highlighting, an attention arc diagram and token heat strips.
• SVG icons, per-section accents, a shortcuts overlay (?, [/], H), AA contrast, focus rings, and empty/error states. Respects reduced motion.

v15.0.0: visual overhaul (31 Oct 2026)

• Navigation: a grouped, searchable sidebar (press /) replaces the wall of 72 tabs; on phones it is a drawer.
• Design system: self-hosted Inter and JetBrains Mono, card panels and restyled controls.
• Charts: axes, gridlines, legends, hover crosshairs, and viridis/magma or diverging heatmaps with colorbars.
• 3D: a tone-mapped scene with glowing heads, soft neuron points and layer labels.
• Behaviour: functionally identical to v14.9.0.

v14.9.0: Neuron Lens 14.1 – 14.9 (31 Oct 2026)

• More models: lenses for TinyStories-1M and 3M. The R.out lens beats the logit lens at every layer on all three models.
• New tests: SAE features, head attribution, calibrated auto-labels, cross-model matching and a Taylor check of RelP, all pre-registered in docs/PREREG-v14x.md.
• Failures: C3, C5, C6 and C8a failed and are listed in docs/REPORT-v14.md.
• New NESTED group:
  • a RISC-V CPU in transformer weights
  • a language model baked into a language model (float32)
  • a MoE inside a MoE
  • lens results on the nested models
• In-browser fit: the full TS-1M lens fit runs in-browser and matches the Node fit exactly.

v14.0.0: Neuron Lens (31 Oct 2026)

• Fitted R-, S- and mid-lenses for TinyStories-8M (lens/ts8m/, hash-pinned), with a neuron and attention-head database covering all 8×1024 neurons and 8×16 heads.
• Three new tabs under INTERPRET:
  • NEURON LENS: a searchable 3D atlas, neuron and head cards, and a live steer test.
  • LENS VALID: the pre-registered validation, with CIs and the failures included.
  • LENS CATALOG: an attributed CC BY 4.0 sample of the Qwen3-8B/Gemma data.
• An in-browser refit runs in a Web Worker.
• The method is prior art by @_a9lim and @treetowntree (stdlogits/neuron-lens), credited in docs/NEURON_LENS_METHOD.md. The code is an original implementation.
• Results are in docs/CLAIMS-v14.md: R.out predicts steering better than the logit and tuned lenses (P1/P2 supported), and the mid-lens advantage does not reproduce at 8M (P3/P4 not supported). The roadmap is in docs/ROADMAP-v14.md.

v13.0.x: native apps (26 Oct 2026; 13.0.1: Electron 42)

• Windows (x64): Neuralese-Setup-13.0.1-x64.exe is an NSIS installer that lets you choose the install folder. Neuralese-13.0.1-x64-portable.exe runs without installing. Both are Electron builds and both are unsigned, so SmartScreen will warn ("More info → Run anyway").
• Linux (x64): Neuralese-13.0.1-x86_64.AppImage (some distros need libfuse2) and Neuralese-13.0.1-amd64.deb.
• Android: Neuralese-13.0.1-debug.apk, a Capacitor debug build you can sideload.
• iOS: Neuralese-13.0.1-unsigned.ipa (re-sign with AltStore or Sideloadly) and Neuralese-13.0.1-simulator.app.zip (Xcode Simulator). Both are built on GitHub macOS runners with CODE_SIGNING_ALLOWED=NO. A signed .ipa is built only when the IOS_P12_BASE64, IOS_P12_PASSWORD, IOS_PROVISION_PROFILE_BASE64 and IOS_TEAM_ID secrets exist.
• Every app ships the same file set as the gh-pages site (tools/build-www.mjs). As on the web, the TinyStories weights (~112 MB) and the HUB ONNX models stream from Hugging Face on first run, so the first launch needs internet.
• The tests no longer depend on a box-specific path. node tools/fetch-model.mjs downloads the test weights into model/ and checks them against pinned SHA-256s, then node --test test/*.test.mjs runs the suite (128 tests, including the new test/v13.test.mjs packaging checks).
• CI (.github/workflows/release.yml) runs only on v* tags or by manual dispatch. It runs the tests, then the Windows, Linux, Android and iOS builds, then publishes one release.

v12.1.0: fixes from the 65-tab exploration audit (26 Oct 2026)

• TAYLOR: the derivative summary showed NaNNaN… because Float64Array.map coerced HTML strings to NaN. It now renders numbers (derivRows).
• INTERPRET → EMBED: the token-neighbour search was blank because HUB's out.embed overwrote it. HUB's function is now out.hubEmbed, and a test fails if two modules ever share an out.* name again.
• SWEEP: target and token arguments are resolved exactly as typed (no trimming), and the resolved ids and pieces are shown under the inputs. Multi-token input is flagged. Small effects print in exponent form instead of being rounded to 0.0000.
• LM SWAP: GPT-2 hung after config.json 100% because Xenova/gpt2 has no q8 ONNX file and the 404 was hidden. It now loads uint8, and load errors are shown. DistilGPT-2's newline-only output was real greedy behaviour. Generation now defaults to repetition penalty 1.3 with no-repeat 3-grams, a "plain greedy" checkbox turns that off, and a whitespace-only continuation is flagged. The GPU checkbox only enables WebGPU for the benchmark (generation stays CPU/JS), and the HUD now says so.
• 3D: a ⤢ fit-to-view button (3D and 3D MODES), and presets fit the camera to the area left visible by the dock and output panel. The 3D MODES text now matches the current mode.
• Charts: every chart carries a "computed HH:MM:SS · current model state" stamp. After any weight or steering change it switches to "last run at … · ⚠ model changed since".
• Provenance banners: FRONTIER, SCORECARD, SELF-IMPROVE and ATLAS are marked stored author results. MIND, GROKKING and CLAIMS are marked stored + live. F-CLAIMS is marked live consistency checks (not reruns). TSOTCHKE, QRNG and Q-GEOMETRY are marked classical simulation, no quantum hardware.
• STEER export/import is now bit-identical (the old toFixed(6) zeroed tiny components). The share-link make→load replay is verified in the smoke test.
• 🧭 guide: a 10-stop guided path through the 65 tabs.
• Layout: at a narrow dock (1240×700), charts were stretched to about 3× the pane width. Fieldsets now shrink and canvases are capped to the pane.
• Tests: test/v121.test.mjs (one per fix); browser smoke smoke-v121.mjs at 1240×700.

What's new in v12 (HUB group, after VIEW in the tab bar)

• Model hub: transformers.js 4.3.1 (pinned, SHA-256-checked) in a Web Worker; ONNX Runtime WASM pinned and SHA-256-checked. Eight public ONNX models, all local inference.
• EMBED: all-MiniLM-L6-v2 sentence embeddings, nearest neighbours, semantic search, rotatable 3D PCA map.
• SPEECH: Whisper tiny.en (mic or file), MMS-TTS with WAV export, speech-to-speech chain; round-trip self-test (WER 0 in our runs).
• VISION: TrOCR single-line OCR (self-test CER 0), ViT-GPT2 captioning, CLIP zero-shot (self-test 98.8%).
• LM SWAP: DistilGPT-2 / GPT-2, generation only. White-box swapping stays TinyStories 1M/3M/8M (TOOLS → ZOO).
• FINE-TUNE: Adam on 1,536 parameters of TinyStories-8M on your text, held-out CE with bootstrap CI, apply as an undoable edit.
• 100-item roadmap with status and acceptance tests: docs/ROADMAP-v12.md. Claims H1–H9: docs/CLAIMS.md.

What's new in v11 — FRONTIER, pre-registered

A FRONTIER tab group runs 14 hypotheses about consciousness-science proxies (Butlin/Long indicator scorecard, PCI-style complexity, ignition), introspection (injected-concept detection, privileged self-prediction vs TinyStories-3M/1M), a gated self-improvement loop, self-distillation, test-time training, grokking with a Fourier check, and an exhaustive exact causal atlas (14 pre-registered hypotheses: 6 supported, 8 not). Thresholds were frozen in docs/PREREG-v11.md and committed before any run; several hypotheses failed and are reported as failures. Nothing here claims the model is conscious, sentient or superintelligent. Write-up: docs/FRONTIER.md.

What's new in v10 — cross-checks against Tsotchke's own code

• Moonlab's native QGT (compiled to WASM) agrees with our Fubini–Study metric to ≈ 2e-8 (qgtcheck in the terminal).
• Eshkol's compiled Taylor kernels agree with taylor.js to 9.7e-16; Eshkol's AD tape reproduces our full backward pass (1.2e-7).
• Eshkol's soft-unify (neurosymbolic.esk) is ported and matches a real eshkol-run trace to 2.2e-16 (unify a b in the terminal).
• Negative result: natural gradient beats Adam on training loss but loses on held-out loss (95% CIs, 8 seeds).
• Known result: the QRNG byte difference between glibc and WASM comes from libm, not SIMD.
• 18 live claims. Details in docs/TSOTCHKE.md and docs/CLAIMS.md.

What's new in v9 — Eshkol types and QGTL run in the browser

• TYPES tab / typecheck: Eshkol's own type layer (lib/types/dependent.cpp, hott_types.cpp), compiled to WebAssembly, checks 73 judgments about the live model graph. The same checker rejects deliberately broken graphs.
• QGTL compiled to WebAssembly: QGTL's qgt_geometric_measure_fidelity now measures every Q-GEOMETRY position next to Moonlab; they agree to about 1e-7. QGTL's metric routine and its tensor-network QGT were tested too. Their limits are documented in docs/TSOTCHKE.md.
• Correction: v8 said Tsotchke's repos had no HoTT work. That was wrong, and TSOTCHKE.md now cites the type system with links.

v8 — the TSOTCHKE group (press 8)

v8 adds components built on the MIT-licensed open-source work of Tsotchke. Only these components carry the ⚛ Powered by Tsotchke badge. Each claim comes with a check that could refute it.

tab	Tsotchke project (commit)	what it does here	evidence
QRNG	quantum_rng 1.1.0 (1a77e77), compiled by us to WASM	research seeds with provenance, SP 800-90B health tests + SP 800-22 subset, simulated CHSH; qseed apply puts the seed into the manifest	deterministic and identical in Node and Chrome; health tests catch a stuck source; NIST worked examples reproduced
Q-GEOMETRY	Moonlab 1.2.1 (WASM) + QGTL metric (770994a)	Fubini–Study angle between clean and edited next-token distributions (ψ = √p, 16 qubits), and the QGT = Fisher/4 check	Moonlab fidelity = direct sum to 1e-16; s(ε)/ε → √Fisher/2; Fisher = jet g″ = PyTorch
TAYLOR	Eshkol 1.3.6 (341ac81), port of taylor_recurrences.def	exact Taylor jets of the loss along a weight direction, compared live with the true loss	closed forms to 5e-15; PyTorch float64 to ~1e-6
TYPES	Eshkol type layer (dependent.cpp, hott_types.cpp) compiled to WASM	73 dimension / path / bounds / universe-subtyping judgments over the live model graph, decided by Eshkol	73/73 hold; truncated, miswired and out-of-vocab graphs are rejected
CLAIMS	all of the above	the Claims & Evidence page: every claim, what would refute it, and the result computed now	test/tsotchke.test.mjs (15 tests incl. mutation tests)

Honest limits:

• quantum_rng is a classical simulation. Its randomness comes only from the crypto.getRandomValues seed, and its CHSH value shows the simulator is correct, not that the bits are quantum.
• Eshkol's HoTT-style type layer is used as a dimension and subtype checker. It is gradual and has no univalence, so we claim no HoTT proofs. No tensor-logic artifact by Tsotchke was found; none is claimed.
• Eshkol's compiler is not run in the browser. Its type layer is, and so is QGTL's geometric-measure core (fidelity). QGTL's tensor-network QGT returned an all-zero state in our WASM build and is not used.
• Moonlab's ESM chunk carries a documented 4-line browser patch; its WASM is unmodified.

See docs/TSOTCHKE.md for the per-repo findings and what was not usable.

What's new in v7 (new RESEARCH group: press 7)

• REPRO: SHA-256 of the original and current weights (also in the top bar), a downloadable experiment manifest (version, model id + hash, edits, runtime, prompt, decoding, seed, environment), and replay that re-applies the edits from clean weights and checks the hash bit for bit.
• STATS: per-head zero-ablation significance on per-document CE (paired sign-flip permutation test, Cohen's d_z, bootstrap CI, Holm + Benjamini–Hochberg), and a general two-sample/paired test. Verified against scipy and statsmodels.
• ACDC: node-level ACDC-style circuit discovery (128 heads + 8 MLPs, τ slider, faithfulness) and causal scrubbing-lite for your own node hypotheses. Edge-level pruning is not implemented.
• PROBES: softmax-regression probes on the residual stream from word lists, split by word type, with a Hewitt–Liang control task and selectivity, swept over layers.
• DATA: upload or paste your own corpus. It stays in the browser, and the docs-based tools (drill, gradients, geometry, landscape, neuron atlas, stats, probes, eval) use it.
• EVAL: a clean-vs-current scoreboard covering perplexity, 12 cloze facts, distinct-1/2 and repeated 4-grams.
• NOTEBOOK: JavaScript cells against the lab API with inline plots, exportable as .md or .ipynb.
• REPORT: a self-contained HTML report with methods, results, figures and the manifest. Print it to get a PDF.
• LOG: JSONL export of the run log, a field-by-field diff of two runs, and full lab-state import/export.
• Hardening (see AUDIT.md):
  • Edit + undo is bitwise exact (fixed a drill restore bug).
  • three.js is vendored and pinned, and a Content-Security-Policy is set.
  • Weight downloads retry with backoff, and there is a retry button.
  • Long computations are chunked.
  • axe-core reports 0 WCAG 2 A/AA violations.
  • Terminal: hash, manifest, stats, acdc, scrub, probe, evalh, dataset, report, state.
• Learning: 19 new glossary terms, 5 Bible sections (including safety and ethics), and lessons 13–14 with quizzes.

Safety and ethics

TinyStories-8M is a tiny model trained on synthetic children's stories, so findings may not transfer to larger models. Edited models produce unfiltered text. Notebook cells execute arbitrary JavaScript in the page, so only run code you trust. Uploaded data never leaves the browser. The only network access is to huggingface.co, for the public weights and tokenizer.

What's new in v6

• SCIENCE › DRILL: hierarchical inspector model → layer → sublayer → matrix → head/neuron → row/col → weight (paths like L3.attn.q.H5.r2.w17), with moments, sparsity, norms, full SVD metrics (effective/stable rank, condition number, Hill α), gradient norm, empirical Fisher, saliency, an analytic-vs-finite-difference check, zero-ablation ΔCE with token provenance, direct logit effect, and undoable edits at every level. The breadcrumb and the 3D selection stay in sync.
• SCIENCE › GRADIENTS: an exact backward pass (verified against PyTorch autograd), grad×input, integrated gradients with a completeness check, attribution patching validated against activation patching of all 128 heads (Pearson r ≈ 0.99), Fisher diagonal, and the top Hessian eigenvalue from HVPs.
• SCIENCE › CIRCUITS / GEOMETRY / LANDSCAPE:
  • OV/QK eigenvalue copying scores, Q/K/V composition, and induction/previous-token scores with bootstrap CIs.
  • CKA, TwoNN intrinsic dimension, anisotropy, logit-lens KL, and superposition interference.
  • 1-D/2-D loss slices.
• VIEW › 3D MODES: morphing layer stack, instanced voxel weight field (LOD), loss surface, residual manifold (PCA-3), attention graph, and circuit graph, with lighting, viridis/diverging metric colors, and camera presets with tweened transitions.
• TOOLS:
  • WebGPU unembedding matvec, verified against CPU (max relative error ~1e-7), with a tok/s benchmark for both.
  • Model zoo (TinyStories 1M/3M/8M, streamed with HTTP Range) with a comparison table.
  • Weight diff (orig/snapshot/current).
  • Share URLs (#x= compressed edits + runtime + prompt).
• CLI: node cli.mjs [--model DIR] <command> (or an interactive prompt) on the same core: gen, drill, attrib, atp, eigen, comp, induct, geom, superpos, land, hess, zoo, diff.
• Charts redraw on theme switch. 22 new glossary terms, 6 Bible sections, and lessons 11–12 with quizzes.
• Honest findings:
  • TinyStories-8M has only weak random-token induction (max ≈ 0.05), confirmed with HF.
  • On software (SwiftShader) WebGPU the GPU path is ~40× slower than CPU, so generation keeps the CPU path.

What's new in v5

• 28 tabs in 7 groups: RUN (gen, steer, control, audio) · MODEL (arch, params, weights) · INTERPRET (analyze, attention, residual, circuit, neurons, embed, SAE, SAE dash) · INTERVENE (causal, surgery, runtime, sweep, ops) · EVALUATE (timeline, calibration, impact, profiler, runs) · VIEW (3D) · TOOLS (terminal, learn). Every chart exports PNG, SVG or CSV.
• Advanced steering: add, norm-preserving add, directional ablation and clamp at any hook site (resid, attn_out, mlp_out, z, q, k, v), with head slices, token masks and time schedules. Also logit bias and h-steering.
• Sweeps and batch runs (edit strength α, steer coef, head temperature, neuron clamp, logit bias, window, h-steer), plus an experiment tracker (RUNS).
• Terminal with about 40 commands, Tab completion, history and scripts.
• Learning center: Bible manual, 129-term glossary, index and search, and a 10-lesson course with labs and quizzes. Also available as docs/BIBLE.md, docs/GLOSSARY.md and docs/COURSE.md, generated by node tools/gendocs.mjs.
• UI: light and dark themes, resizable dock, keyboard shortcuts (?), ARIA roles, and a visible version label. Assets are cache-busted with ?v=5.
• Fixes:
  • The first generation now records the timeline even when audio is off.
  • The root label shows the note and octave, matching the MIDI note.
  • Clamp no longer acts on unscheduled layers.
  • The head-ablation mode name now matches what the engine supports.

Live: https://0thernes.github.io/neuralese/

The app loads roneneldan/TinyStories-8M (GPT-Neo, 8 layers, 16 heads, d=256, 1024 MLP neurons) straight from Hugging Face. HTTP Range requests fetch only the 78.8 MB of fp32 tensors and skip the bool mask buffers. A from-scratch JS forward pass with hooks at resid_pre, z, attn_out, mlp_act, mlp_out, resid_post then runs everything on the CPU in f32.

Modules

file	what
engine.js	GPT-Neo forward pass, KV cache, traces, hooks. Runtime controls: layer program (dup/delete/skip/reorder), per-head temperature and zero/mean/resample ablation, attention-mask modes, positional remap, steering stack (site, per-layer schedule, token mask), SAE clamps, neuron overrides
analysis.js	corpus stats, logit lens, tuned lens (ridge affine per layer), exact DLA (embed/heads/neurons/biases/other), residual and head patching, path patching head→logits, ablation sweeps, control vectors (mean-diff / PCA), vector arithmetic, token stats
surgery.js	replayable edit stack: elementwise ops, SVD (top-k/drop/scale/flatten), low-rank noise, int8/4/2 quantization, layer interpolation/transplant, LayerNorm γ/β, wpe tampering, ROME rank-one, chaos; safetensors export
rome.js	ROME: u = C⁻¹k*/(kᵀC⁻¹k) from the corpus MLP-activation second moment; v* optimised (direction ascent + SPSA/Adam) through the real forward pass
sae.js	ReLU SAE with L1, unit-norm decoder, Adam, trained in the browser
decode.js	temperature, top-k/p, min-p, typical, repetition/frequency/presence penalties, logit bias, beam search, contrastive decoding, CFG
linalg.js	one-sided Jacobi SVD, Cholesky solve, PCA

Verification

• v10: npm test runs 92 tests (+10: test/v10.test.mjs A1–A3, C1, F1, N1–N2; eshkol-taylor E1; eshkol-ad D1–D2). node smoke-v10.mjs <url> <prefix> checks 18 claims, the terminal commands, CSP and console errors.
• v9: npm test runs 82 tests (+7 in test/hott.test.mjs: H1–H4 Eshkol type checks with mutations, H5–H7 QGTL). node smoke-v9.mjs adds TYPES and QGTL checks.
• v8: npm test runs 75 tests (59 earlier + 15 in test/tsotchke.test.mjs). node smoke-v8.mjs runs the TSOTCHKE group in headless Chrome: every live claim, WASM loading under the CSP, golden QRNG bytes, and FS angle after an edit.
• v7: npm test runs 59 tests in one go (about 100 s). New tests cover bitwise undo exactness for every edit kind, stats against scipy/statsmodels, the weights hash, ACDC/scrubbing consistency, probe selectivity, and the eval harness. node smoke-v7.mjs <url> covers every tab, every RESEARCH tool and terminal command, CSP violations, a leak test (500 tokens plus 8 tab cycles), long tasks, and a 390 px layout. node tools/netfail.mjs checks retry and failure handling. node tools/a11y.mjs runs an axe WCAG 2 A/AA audit of every tab. npm run lint runs eslint.
• node --test test/ runs 32 tests (18 against PyTorch references produced by ref.py (HF transformers): forward logits, logit lens, DLA, resid/head/path patching, SVD vs numpy, ROME flip, quantization ordering, SAE, decoding, and the safetensors round-trip; 7 audio tests; 7 v5 lab tests covering steering ops, schedules, labcore, the terminal, and learning-data integrity). v6 adds grad/science/zoo/drill/cli tests (52 total, PyTorch/HF-verified). node smoke-v6.mjs <url> (superset of v5) and node smoke-v5.mjs <url> opens every tab, runs every terminal command, and answers every quiz in headless Chrome.
• node smoke.mjs runs a headless-Chrome end-to-end check of the UI.

Repo layout

main holds the source, tests, and PyTorch reference script. gh-pages holds the deployed static site (same JS files). To run the tests you need a local copy of the weights in model/ (model.safetensors, vocab.json, merges.txt). Then run python test/ref/ref.py (torch + transformers) followed by npm test. Serve the app locally with npm run serve (static server over src/) and run npm run smoke.