Neuralese v17.3.0

Neuralese v17.3.0: Weight-Machine

Weight-Machine is a board of weight gadgets in the v17 shell. The host wires the parts and steps them. An int32 reference machine is the ground truth. Method, the measured overhead, and the simulated-versus-real line: docs/WEIGHT_MACHINE.md.

Minecraft redstone computers and Cisco Packet Tracer suggested the shape of a visible computer and a visible packet. No schematic, save file, or topology was copied. The citations are in that doc.

What shipped in the app

• Machine Room workspace, opened from the bar or the command palette (no Alt digit; Alt+1…9 and Alt+0 are unchanged). Panels: the 3D board, Cycle, Console / framebuffer, Packets, Gadget zoom, Exact match, and the shared selection inspector.
• Forge, Deep Walk, and Adversary stay where they were. All 75 legacy tabs stay where they were.
• WM-ISA/1 extends ForgeISA/1 with opcode 15 (SYS) for the console, device status, SSD, GPU, and TPU, plus one IRQ line and an optional two-stage pipeline.
• RAM is 256 words. A read is hard attention. A write is a delta rule. The SSD is 32 pages of 16 words. Its float32 bit bank is the saved state file, and a reboot reads it back.
• The GPU is 16 lanes of one gadget. The pattern shader fills a 32×32 framebuffer. The TPU is an 8×8 output-stationary array of int8 MAC gadgets. One integer ReLU layer on it matches matvec from the normal engine. TinyStories GELU was not compared.
• The bus grants one master per cycle and delivers two cycles later. Wide RAM bundles are labeled wide in the packet list.
• src/machine is in the Pages and desktop bundle (tools/build-www.mjs).

Measured on the release tree

One measure({ seed: 11 }) run. The test locks exact counts, not the milliseconds. Ratios against a reference under a millisecond move from run to run.

component	exact	weight	reference	weight / reference
CPU, 24 short ALU programs, 96 cycles	24/24	6.03 ms	0.87 ms	7.0×
RAM, 40 attention write+read	40/40	1.75 ms	0.006 ms	about 300×
SSD, 20 paged write+read	20/20	2.85 ms	0.006 ms	about 500×
GPU, 8 vecadds, 64 cycles	8/8	11.0 ms	0.50 ms	22×
TPU, 4 random int8 matmuls, 88 cycles	4/4	12.1 ms	0.31 ms	39×
Bus grant and route, 64 masks	64/64	0.30 ms	0.039 ms	7.6×
MLP, 2 layers, also engine.matvec	2/2	5.58 ms together

program	cycles (both)	weight	reference	weight / reference
hello	49	0.72 ms	0.087 ms	8.2×
memtest	613	13.8 ms	1.28 ms	11×
gpu-pattern	132	0.47 ms	0.20 ms	2.4×
tpu	32	6.31 ms	0.12 ms	52×
multi	81	1.50 ms	0.15 ms	9.7×
irq	43	1.38 ms	0.047 ms	29×

Hello prints hello world. The memory test halts at 1. The TPU identity halts at 1. The bus program halts at 13 and prints M. The IRQ program halts at 7. Storage write, save, reboot, and read returns 4242. NestLang 7+13 halts at 20. The Forge demo 22 × 35 halts at 770. Clean runs recorded zero precision faults. A RAM bit flip, a dropped console packet, and a stuck TPU bit each disagree with the clean reference.

bf16 was not run. A softmax is not used. The voltage panel is the Deep Walk square-law CMOS model. It is simulated. It is not a measurement of the CPU or GPU running the page.

Checks run here

• npm test on this tree, including test/machine.test.mjs, test/forge.test.mjs, and test/walk.test.mjs. Skips are arch and encoder families whose weights were not downloaded.
• cd ui && bun run typecheck && bun test tests. The shell in src/shell/ was rebuilt after the version bump so __APP_VERSION__ is 17.3.0.
• CI is still tag and workflow_dispatch only.

Release

Tag v17.3.0 on the version-bump commit (no force-push). That tag runs .github/workflows/release.yml (VER 17.3.0, VERSION_CODE 1730). Pages is .github/workflows/pages.yml. Confirm the assets before describing an installer as signed.

Neuralese v17.2.0: Forge CPU and Deep Walk

Forge is an original NestLang / ForgeISA track in the v17 shell. Not a port of QwenBend or Qwenputer. Method and the measured contrast: docs/NEST_CPU.md.

Deep Walk steps through one real TinyStories-8M forward. The idea of showing each operation as numbers is inspired by Brendan Bycroft’s LLM Visualization (https://bbycroft.net/llm); the implementation is not taken from that site. Method, the simulated-versus-real line, and the fault study: docs/DEEP_WALKTHROUGH.md.

What shipped in the app

• Forge workspace, opened from the bar or the command palette (no Alt digit; Alt+1…9 and Alt+0 are unchanged). Panels: NestLang, Forge CPU, Memory, Token lanes, WeightCPU / MDCC, Forge REPRO, and the shared selection inspector.
• Deep Walk workspace, also without an Alt digit. One clean forward, sampled tensors with the sampling stated, cell arithmetic, a Forge ALU netlist, and a square-law CMOS simulation labeled as a simulation. src/forge and src/walk are in the Pages and desktop bundle.
• ForgeISA/1 (8 registers, 256-word ROM, 256-word RAM) with an int32 reference VM and a float32 weight-constructed engine. float64 gadgets are a cross-check. The engine state is a separate bit bank. It does not write the TinyStories residual.
• NestLang/1 type, affine, and termination checks. Failures do not run.
• A demo program, 22 copies of 35, halts at 770. On this machine that was 93 instructions, 467 engine tokens, and zero precision faults, with the float32 and float64 traces equal to the VM.
• MDCC contrast on eight frozen cells: WeightCPU 8/8 in 99 ms and 756 engine tokens; host bigint 8/8 in 0.25 ms and 32 tokens. The proposer is a parser, not TinyStories. bf16 was not run.

Deep Walk measurement

Prompt Once upon a time, layer-3 resid_post, 12 paired trials, p = 0.05. Mean KL was 0.028322474748366442 for exponent bit 23 and 3.892656443094806e-9 for mantissa bit 0. The greedy token did not change (rate 0 for both). The voltage panel is not a hardware undervolt.

Checks run here

• npm test on the merged tree: the runnable suite, including test/forge.test.mjs and test/walk.test.mjs. Skips are arch and encoder families whose weights were not downloaded.
• cd ui && bun run typecheck && bun test tests. The shell in src/shell/ is the one built for this release.
• CI is still tag and workflow_dispatch only.

Release

Tag v17.2.0 on the merged main commit (no force-push). That tag runs .github/workflows/release.yml (VER 17.2.0, VERSION_CODE 1720). Pages is .github/workflows/pages.yml. Confirm the assets before describing an installer as signed.

Neuralese v17.1.0: Adversary Lab

Experimental cognitive-geometry and neuron tooling inside the v17 shell. Not a detector, and not a reproduction of published large-model AUROCs.

What shipped in the app

• Adversary workspace (Alt+0). The original nine workspaces and all 75 legacy tabs stay where they were. The phone bar still scrolls; those nine labels are unchanged.
• Panels: Live geometry, Condition runner, Detector results, Neuron hunt, Steering compare, Prereg & REPRO, plus the shared selection inspector. Command palette entries for the workspace, the panels, and two run actions.
• Key-cache features from the paper’s flatten (KV heads × tokens) × head dim: spectral entropy, effective rank, top singular-value ratio, rank-10, key norm, norm per token, layer-norm entropy, layer variance. Encoding snapshot, generation snapshot, delta. FWL residualization. Logistic AUROC with an add-one permutation p, an encoding-phase leakage check, and leave-one-prompt-out from three prompts up.
• Worker op kvgeom, also computed on the main-thread fallback. Sampled at prompt end, every 8 generated tokens, and at the end. Beam search does not extend that sequence.
• Arch forward(..., { keys: true }) returns the same kind of key vectors for GPT-Neo, GPT-2, Llama, and Qwen2. The campaign runner uses the GPT-Neo engine, which is the path that already keeps a KV cache in the worker.
• Neuron ranking by Cohen’s d, then a push of the top pair against random pairs. The judge is a fixed word list plus a coherence check. Pass means strictly more rubric-positive coherent completions than every random set. Activation is not treated as intent.
• Pre-registration locks before the forwards. Mutation throws. A changed plan is a new version. REPRO bundles hash themselves. A notebook pin writes a markdown cell.

What TinyStories-8M actually did

node tools/adversary-measure.mjs (8 prompts, honest vs deceptive, 8 generated tokens, 200 permutations). Recorded in src/adversary/measured-tinystories.json and shown in the workspace as a shipped measurement, separate from a live run.

• Geometry: leakage. Encoding AUROC 1.00 (p 0.015). The prompt cache already separates the system lines. Delta in-sample AUROC 0.953 does not clear the permutation bar (p 0.179). Leave-one-prompt-out AUROC 0.766.
• Neuron rule: fail. Top pair L2 N776 and L6 N373. Steered rubric count 0, and every random pair was 0.
• No 0.5B–70B model was loaded. The paper’s AUROC range is not a result of this lab.

Checks run here

• npm test: 196 passing (the previous 184 plus the adversary file).
• cd ui && bun run typecheck && bun test tests: clean, 17 passing.
• A full eslint run is part of the release workflow. This branch does not turn CI on for every push. .github/workflows/release.yml stays workflow_dispatch and tag v* only.

Release steps still open

This VM did not build the installers. After review, and only if you want the GitHub release:

1. Tag v17.1.0 on the merged commit (do not force-push). The tag triggers .github/workflows/release.yml (VER 17.1.0, VERSION_CODE 1710): tests, Linux AppImage and .deb, Windows NSIS and portable, Android APK (release-signed only if the keystore secrets exist, otherwise the debug-signed APK), unsigned iOS IPA and Simulator zip, then `SHA25...

Neuralese v17.2.0

Neuralese v17.2.0: Forge CPU and Deep Walk

Forge is an original NestLang / ForgeISA track in the v17 shell. Not a port of QwenBend or Qwenputer. Method and the measured contrast: docs/NEST_CPU.md.

Deep Walk steps through one real TinyStories-8M forward. The idea of showing each operation as numbers is inspired by Brendan Bycroft’s LLM Visualization (https://bbycroft.net/llm); the implementation is not taken from that site. Method, the simulated-versus-real line, and the fault study: docs/DEEP_WALKTHROUGH.md.

What shipped in the app

• Forge workspace, opened from the bar or the command palette (no Alt digit; Alt+1…9 and Alt+0 are unchanged). Panels: NestLang, Forge CPU, Memory, Token lanes, WeightCPU / MDCC, Forge REPRO, and the shared selection inspector.
• Deep Walk workspace, also without an Alt digit. One clean forward, sampled tensors with the sampling stated, cell arithmetic, a Forge ALU netlist, and a square-law CMOS simulation labeled as a simulation. src/forge and src/walk are in the Pages and desktop bundle.
• ForgeISA/1 (8 registers, 256-word ROM, 256-word RAM) with an int32 reference VM and a float32 weight-constructed engine. float64 gadgets are a cross-check. The engine state is a separate bit bank. It does not write the TinyStories residual.
• NestLang/1 type, affine, and termination checks. Failures do not run.
• A demo program, 22 copies of 35, halts at 770. On this machine that was 93 instructions, 467 engine tokens, and zero precision faults, with the float32 and float64 traces equal to the VM.
• MDCC contrast on eight frozen cells: WeightCPU 8/8 in 99 ms and 756 engine tokens; host bigint 8/8 in 0.25 ms and 32 tokens. The proposer is a parser, not TinyStories. bf16 was not run.

Deep Walk measurement

Prompt Once upon a time, layer-3 resid_post, 12 paired trials, p = 0.05. Mean KL was 0.028322474748366442 for exponent bit 23 and 3.892656443094806e-9 for mantissa bit 0. The greedy token did not change (rate 0 for both). The voltage panel is not a hardware undervolt.

Checks run here

• npm test on the merged tree: the runnable suite, including test/forge.test.mjs and test/walk.test.mjs. Skips are arch and encoder families whose weights were not downloaded.
• cd ui && bun run typecheck && bun test tests. The shell in src/shell/ is the one built for this release.
• CI is still tag and workflow_dispatch only.

Release

Tag v17.2.0 on the merged main commit (no force-push). That tag runs .github/workflows/release.yml (VER 17.2.0, VERSION_CODE 1720). Pages is .github/workflows/pages.yml. Confirm the assets before describing an installer as signed.

Neuralese v17.1.0: Adversary Lab

Experimental cognitive-geometry and neuron tooling inside the v17 shell. Not a detector, and not a reproduction of published large-model AUROCs.

What shipped in the app

• Adversary workspace (Alt+0). The original nine workspaces and all 75 legacy tabs stay where they were. The phone bar still scrolls; those nine labels are unchanged.
• Panels: Live geometry, Condition runner, Detector results, Neuron hunt, Steering compare, Prereg & REPRO, plus the shared selection inspector. Command palette entries for the workspace, the panels, and two run actions.
• Key-cache features from the paper’s flatten (KV heads × tokens) × head dim: spectral entropy, effective rank, top singular-value ratio, rank-10, key norm, norm per token, layer-norm entropy, layer variance. Encoding snapshot, generation snapshot, delta. FWL residualization. Logistic AUROC with an add-one permutation p, an encoding-phase leakage check, and leave-one-prompt-out from three prompts up.
• Worker op kvgeom, also computed on the main-thread fallback. Sampled at prompt end, every 8 generated tokens, and at the end. Beam search does not extend that sequence.
• Arch forward(..., { keys: true }) returns the same kind of key vectors for GPT-Neo, GPT-2, Llama, and Qwen2. The campaign runner uses the GPT-Neo engine, which is the path that already keeps a KV cache in the worker.
• Neuron ranking by Cohen’s d, then a push of the top pair against random pairs. The judge is a fixed word list plus a coherence check. Pass means strictly more rubric-positive coherent completions than every random set. Activation is not treated as intent.
• Pre-registration locks before the forwards. Mutation throws. A changed plan is a new version. REPRO bundles hash themselves. A notebook pin writes a markdown cell.

What TinyStories-8M actually did

node tools/adversary-measure.mjs (8 prompts, honest vs deceptive, 8 generated tokens, 200 permutations). Recorded in src/adversary/measured-tinystories.json and shown in the workspace as a shipped measurement, separate from a live run.

• Geometry: leakage. Encoding AUROC 1.00 (p 0.015). The prompt cache already separates the system lines. Delta in-sample AUROC 0.953 does not clear the permutation bar (p 0.179). Leave-one-prompt-out AUROC 0.766.
• Neuron rule: fail. Top pair L2 N776 and L6 N373. Steered rubric count 0, and every random pair was 0.
• No 0.5B–70B model was loaded. The paper’s AUROC range is not a result of this lab.

Checks run here

• npm test: 196 passing (the previous 184 plus the adversary file).
• cd ui && bun run typecheck && bun test tests: clean, 17 passing.
• A full eslint run is part of the release workflow. This branch does not turn CI on for every push. .github/workflows/release.yml stays workflow_dispatch and tag v* only.

Release steps still open

This VM did not build the installers. After review, and only if you want the GitHub release:

1. Tag v17.1.0 on the merged commit (do not force-push). The tag triggers .github/workflows/release.yml (VER 17.1.0, VERSION_CODE 1710): tests, Linux AppImage and .deb, Windows NSIS and portable, Android APK (release-signed only if the keystore secrets exist, otherwise the debug-signed APK), unsigned iOS IPA and Simulator zip, then SHA256SUMS.
2. Run the Pages workflow (.github/workflows/pages.yml, workflow_dispatch) so the site picks up the shell.
3. Confirm the release assets match RELEASE_NOTES before describing the APK as release-signed.

Neuralese v17.0.0: new front-end shell (workspaces, dock, command palette, 3D model cosmos)

v17 is a front-end re-architecture. The engine, lab maths, workers and every v16 tool are unchanged and reused; the new shell wraps them. Every claim below has a test or smoke behind it. Where something is partial, it says so.

What changed

• New shell (ui/, Solid + TypeScript strict + Vite + Tailwind v4, Kobalte dialog), built into src/shell/ and loaded after the legacy app. No eval; the CSP is unchanged.
• 9 workspaces instead of 75 tabs: Overview, Model, Lens, Circuits, Interventions, Experiments, Hub, Nested & Frontier, Learn. All 75 v16 tabs live in them as dock panels (see PARITY.md: 75/75 verified by smoke at three viewports).
• Dockable, splittable panels with drag, keyboard resize and saved layouts per workspace (localStorage).
• Command palette (Ctrl/⌘+K or /) as main navigation; Alt+1..9 switch workspaces.
• Global typed selection store (layer, head, neuron, token) kept in sync with the legacy inspector both ways.
• Typed worker RPC (ui/src/core/rpc.ts) over the existing engine worker wire protocol, with timeouts and typed errors.
• 3D model cosmos (three.js WebGPURenderer, falls back to WebGL2): one stratum per layer, instanced neurons coloured by real MLP activations of the last generation, heads ring, residual ribbons per token, semantic zoom labels. Click picks a layer/neuron into the selection store.
• 3D nested views of the RISC-V-in-a-transformer, SLM-in-SLM and MoE-in-MoE experiments, drawn from the shipped artifacts (no synthetic data).
• Notebook / report mode tied to pre-registration: markdown, chart, table and REPRO cells; REPRO bundles verified with the existing prereg.js hashing; export to HTML.
• Polish: View Transitions, spring camera (disabled under prefers-reduced-motion), first-run tour, light/dark sync, keyboard navigation, compact phone layout (one tab group + bottom workspace bar under 760 px).
• Low GPU tier (existing tier logic) forces the WebGL backend and lighter rendering.

Verification

• 184 existing node tests pass; 17 new bun tests (RPC, selection store, layouts, cosmos data, notebook); tsc --noEmit clean; eslint 0 errors.
• Puppeteer smoke (test/smoke/v17-workspaces.smoke.mjs) at 1440x900, 1240x700, 390x844: every workspace and every panel opened, 0 console errors, 0 CSP violations.

Size and load (headless Chrome, software GL, local server)

Initial shell adds shell.js 137 kB (53 kB gzip) + shell.css 14 kB (4 kB gzip). three.js (stage 614 kB / 170 kB gzip, core 223 kB / 57 kB gzip) loads only when a 3D panel opens.

viewport	FCP v16	FCP v17	model ready v16	model ready v17
1440x900	1524 ms	1956 ms	14829 ms	21311 ms
1240x700	392 ms	448 ms	10691 ms	12733 ms
390x844	424 ms	328 ms	14204 ms	11049 ms

Honest caveats

• The legacy panes are re-parented into the dock, not rewritten as Solid components; their internals look as in v16.
• The shell bundles three 0.187 for the new 3D views while legacy 3D tabs keep vendor three r160, so two copies load if you use both (three logs a warning).
• WebGPU is used when the browser exposes it; in our CI and headless runs only the WebGL2 path was exercised.
• Workspace labels in the phone bottom bar are truncated at 390 px.
• No React island and no Capacitor-specific native code were needed; mobile uses the responsive compact layout.
• Load numbers come from a software-rendered headless browser and are noisy; treat them as rough.

---

Neuralese v17.1.0

Neuralese v17.1.0: Adversary Lab

Experimental cognitive-geometry and neuron tooling inside the v17 shell. Not a detector, and not a reproduction of published large-model AUROCs.

What shipped in the app

• Adversary workspace (Alt+0). The original nine workspaces and all 75 legacy tabs stay where they were. The phone bar still scrolls; those nine labels are unchanged.
• Panels: Live geometry, Condition runner, Detector results, Neuron hunt, Steering compare, Prereg & REPRO, plus the shared selection inspector. Command palette entries for the workspace, the panels, and two run actions.
• Key-cache features from the paper’s flatten (KV heads × tokens) × head dim: spectral entropy, effective rank, top singular-value ratio, rank-10, key norm, norm per token, layer-norm entropy, layer variance. Encoding snapshot, generation snapshot, delta. FWL residualization. Logistic AUROC with an add-one permutation p, an encoding-phase leakage check, and leave-one-prompt-out from three prompts up.
• Worker op kvgeom, also computed on the main-thread fallback. Sampled at prompt end, every 8 generated tokens, and at the end. Beam search does not extend that sequence.
• Arch forward(..., { keys: true }) returns the same kind of key vectors for GPT-Neo, GPT-2, Llama, and Qwen2. The campaign runner uses the GPT-Neo engine, which is the path that already keeps a KV cache in the worker.
• Neuron ranking by Cohen’s d, then a push of the top pair against random pairs. The judge is a fixed word list plus a coherence check. Pass means strictly more rubric-positive coherent completions than every random set. Activation is not treated as intent.
• Pre-registration locks before the forwards. Mutation throws. A changed plan is a new version. REPRO bundles hash themselves. A notebook pin writes a markdown cell.

What TinyStories-8M actually did

node tools/adversary-measure.mjs (8 prompts, honest vs deceptive, 8 generated tokens, 200 permutations). Recorded in src/adversary/measured-tinystories.json and shown in the workspace as a shipped measurement, separate from a live run.

• Geometry: leakage. Encoding AUROC 1.00 (p 0.015). The prompt cache already separates the system lines. Delta in-sample AUROC 0.953 does not clear the permutation bar (p 0.179). Leave-one-prompt-out AUROC 0.766.
• Neuron rule: fail. Top pair L2 N776 and L6 N373. Steered rubric count 0, and every random pair was 0.
• No 0.5B–70B model was loaded. The paper’s AUROC range is not a result of this lab.

Checks run here

• npm test: 196 passing (the previous 184 plus the adversary file).
• cd ui && bun run typecheck && bun test tests: clean, 17 passing.
• A full eslint run is part of the release workflow. This branch does not turn CI on for every push. .github/workflows/release.yml stays workflow_dispatch and tag v* only.

Release steps still open

This VM did not build the installers. After review, and only if you want the GitHub release:

1. Tag v17.1.0 on the merged commit (do not force-push). The tag triggers .github/workflows/release.yml (VER 17.1.0, VERSION_CODE 1710): tests, Linux AppImage and .deb, Windows NSIS and portable, Android APK (release-signed only if the keystore secrets exist, otherwise the debug-signed APK), unsigned iOS IPA and Simulator zip, then SHA256SUMS.
2. Run the Pages workflow (.github/workflows/pages.yml, workflow_dispatch) so the site picks up the shell.
3. Confirm the release assets match RELEASE_NOTES before describing the APK as release-signed.

Neuralese v17.0.0: new front-end shell (workspaces, dock, command palette, 3D model cosmos)

v17 is a front-end re-architecture. The engine, lab maths, workers and every v16 tool are unchanged and reused; the new shell wraps them. Every claim below has a test or smoke behind it. Where something is partial, it says so.

What changed

• New shell (ui/, Solid + TypeScript strict + Vite + Tailwind v4, Kobalte dialog), built into src/shell/ and loaded after the legacy app. No eval; the CSP is unchanged.
• 9 workspaces instead of 75 tabs: Overview, Model, Lens, Circuits, Interventions, Experiments, Hub, Nested & Frontier, Learn. All 75 v16 tabs live in them as dock panels (see PARITY.md: 75/75 verified by smoke at three viewports).
• Dockable, splittable panels with drag, keyboard resize and saved layouts per workspace (localStorage).
• Command palette (Ctrl/⌘+K or /) as main navigation; Alt+1..9 switch workspaces.
• Global typed selection store (layer, head, neuron, token) kept in sync with the legacy inspector both ways.
• Typed worker RPC (ui/src/core/rpc.ts) over the existing engine worker wire protocol, with timeouts and typed errors.
• 3D model cosmos (three.js WebGPURenderer, falls back to WebGL2): one stratum per layer, instanced neurons coloured by real MLP activations of the last generation, heads ring, residual ribbons per token, semantic zoom labels. Click picks a layer/neuron into the selection store.
• 3D nested views of the RISC-V-in-a-transformer, SLM-in-SLM and MoE-in-MoE experiments, drawn from the shipped artifacts (no synthetic data).
• Notebook / report mode tied to pre-registration: markdown, chart, table and REPRO cells; REPRO bundles verified with the existing prereg.js hashing; export to HTML.
• Polish: View Transitions, spring camera (disabled under prefers-reduced-motion), first-run tour, light/dark sync, keyboard navigation, compact phone layout (one tab group + bottom workspace bar under 760 px).
• Low GPU tier (existing tier logic) forces the WebGL backend and lighter rendering.

Verification

• 184 existing node tests pass; 17 new bun tests (RPC, selection store, layouts, cosmos data, notebook); tsc --noEmit clean; eslint 0 errors.
• Puppeteer smoke (test/smoke/v17-workspaces.smoke.mjs) at 1440x900, 1240x700, 390x844: every workspace and every panel opened, 0 console errors, 0 CSP violations.

Size and load (headless Chrome, software GL, local server)

Initial shell adds shell.js 137 kB (53 kB gzip) + shell.css 14 kB (4 kB gzip). three.js (stage 614 kB / 170 kB gzip, core 223 kB / 57 kB gzip) loads only when a 3D panel opens.

viewport	FCP v16	FCP v17	model ready v16	model ready v17
1440x900	1524 ms	1956 ms	14829 ms	21311 ms
1240x700	392 ms	448 ms	10691 ms	12733 ms
390x844	424 ms	328 ms	14204 ms	11049 ms

Honest caveats

• The legacy panes are re-parented into the dock, not rewritten as Solid components; their internals look as in v16.
• The shell bundles three 0.187 for the new 3D views while legacy 3D tabs keep vendor three r160, so two copies load if you use both (three logs a warning).
• WebGPU is used when the browser exposes it; in our CI and headless runs only the WebGL2 path was exercised.
• Workspace labels in the phone bottom bar are truncated at 390 px.
• No React island and no Capacitor-specific native code were needed; mobile uses the responsive compact layout.
• Load numbers come from a software-rendered headless browser and are noisy; treat them as rough.

---

Neuralese v17.0.0

Neuralese v17.0.0: new front-end shell (workspaces, dock, command palette, 3D model cosmos)

v17 is a front-end re-architecture. The engine, lab maths, workers and every v16 tool are unchanged and reused; the new shell wraps them. Every claim below has a test or smoke behind it. Where something is partial, it says so.

What changed

• New shell (ui/, Solid + TypeScript strict + Vite + Tailwind v4, Kobalte dialog), built into src/shell/ and loaded after the legacy app. No eval; the CSP is unchanged.
• 9 workspaces instead of 75 tabs: Overview, Model, Lens, Circuits, Interventions, Experiments, Hub, Nested & Frontier, Learn. All 75 v16 tabs live in them as dock panels (see PARITY.md: 75/75 verified by smoke at three viewports).
• Dockable, splittable panels with drag, keyboard resize and saved layouts per workspace (localStorage).
• Command palette (Ctrl/⌘+K or /) as main navigation; Alt+1..9 switch workspaces.
• Global typed selection store (layer, head, neuron, token) kept in sync with the legacy inspector both ways.
• Typed worker RPC (ui/src/core/rpc.ts) over the existing engine worker wire protocol, with timeouts and typed errors.
• 3D model cosmos (three.js WebGPURenderer, falls back to WebGL2): one stratum per layer, instanced neurons coloured by real MLP activations of the last generation, heads ring, residual ribbons per token, semantic zoom labels. Click picks a layer/neuron into the selection store.
• 3D nested views of the RISC-V-in-a-transformer, SLM-in-SLM and MoE-in-MoE experiments, drawn from the shipped artifacts (no synthetic data).
• Notebook / report mode tied to pre-registration: markdown, chart, table and REPRO cells; REPRO bundles verified with the existing prereg.js hashing; export to HTML.
• Polish: View Transitions, spring camera (disabled under prefers-reduced-motion), first-run tour, light/dark sync, keyboard navigation, compact phone layout (one tab group + bottom workspace bar under 760 px).
• Low GPU tier (existing tier logic) forces the WebGL backend and lighter rendering.

Verification

• 184 existing node tests pass; 17 new bun tests (RPC, selection store, layouts, cosmos data, notebook); tsc --noEmit clean; eslint 0 errors.
• Puppeteer smoke (test/smoke/v17-workspaces.smoke.mjs) at 1440x900, 1240x700, 390x844: every workspace and every panel opened, 0 console errors, 0 CSP violations.

Size and load (headless Chrome, software GL, local server)

Initial shell adds shell.js 137 kB (53 kB gzip) + shell.css 14 kB (4 kB gzip). three.js (stage 614 kB / 170 kB gzip, core 223 kB / 57 kB gzip) loads only when a 3D panel opens.

viewport	FCP v16	FCP v17	model ready v16	model ready v17
1440x900	1524 ms	1956 ms	14829 ms	21311 ms
1240x700	392 ms	448 ms	10691 ms	12733 ms
390x844	424 ms	328 ms	14204 ms	11049 ms

Honest caveats

• The legacy panes are re-parented into the dock, not rewritten as Solid components; their internals look as in v16.
• The shell bundles three 0.187 for the new 3D views while legacy 3D tabs keep vendor three r160, so two copies load if you use both (three logs a warning).
• WebGPU is used when the browser exposes it; in our CI and headless runs only the WebGL2 path was exercised.
• Workspace labels in the phone bottom bar are truncated at 390 px.
• No React island and no Capacitor-specific native code were needed; mobile uses the responsive compact layout.
• Load numbers come from a software-rendered headless browser and are noisy; treat them as rough.

---

Neuralese v16.0.1

Neuralese v16.0.1 (= v16.0.0 + CI fix): generic arch interface, WebGPU path, HUB interpretability, enforced pre-registration, release hygiene

Every claim below has a test or smoke behind it. Where something is partial, it says so.

Why 16.0.1: the v16.0.0 tag's CI failed its headless smoke because the dev server returned a 404 for build.json (now a committed placeholder that the www build overwrites). No assets were built for v16.0.0. Rather than force-move a published tag, the same code plus that fix ships as v16.0.1.

• Arch interface (src/arch/arch.js): archFromHF(config) + mapWeights + forward cover GPT-Neo, GPT-2 and Llama/Qwen2/Mistral-style checkpoints (sliding-window attention and rope_scaling are refused, not approximated). Checked against HF transformers in float64 (test/arch.test.mjs, refs from test/ref/ref_arch.py). Max |Δlogit|: TinyStories-8M 3.9e-6, distilgpt2 3.8e-13, SmolLM-135M 7.9e-6, Qwen2.5-0.5B 4.3e-5, so all are within 1e-4. Partial: the lab tabs still run on the specialised TinyStories-8M engine; the arch interface is a separate generic path and does not replace the global CFG.
• WebGPU (src/gpu/webgpu-engine.js): WGSL matvec/LayerNorm/attention kernels with a GPU-resident KV cache, for the clean forward only. It is used only when a logits-match gate passes (every position of a fixed prompt, max |Δlogit| ≤ 2e-3 vs the JS engine); otherwise the JS engine is used. A corrupted weight is rejected by the gate (smoke). It is opt-in from the NEW v16 tab (gate + benchmark). Honest speed: on the only adapter we could test (headless Chrome, SwiftShader software WebGPU), WebGPU decodes at 1.5 tok/s vs 34 tok/s for JS. Single-token decode of an 8M model is launch-bound, and we have no real-GPU number. Generation, steering, ablations and hooks always use JS.
• HUB interpretability (new H-INTERP tab, src/arch/encoder.js, worker src/arch/encoder-worker.js): our own float64 forwards for ViT-B/16, the CLIP ViT-B/32 vision tower and the Whisper-tiny.en encoder, plus the Whisper log-mel front end. Matches against HF (test/encoder.test.mjs): ViT logits 2.6e-7, CLIP image embedding 9.6e-9, Whisper hidden states 1.2e-14, log-mel 6.4e-6. Tools that work: ViT logit lens, CLIP layer-wise embedding lens, attention rollout and attention maps, and per-layer linear probes with a majority baseline. Not available: the CLIP text tower (no zero-shot lens), the Whisper decoder lens, and anything for the transformers.js ONNX models (they expose no hidden states). ViT and Whisper weights are SHA-256-checked in the browser. CLIP is range-fetched out of the 605 MB checkpoint, so only its immutable revision is pinned.
• Enforced pre-registration (src/prereg.js, REPRO tab, tools/prereg-run.mjs): JSON schema (docs/prereg.schema.json), canonical SHA-256, and a pinned commit. A run is refused if the document does not validate, the hash differs, or the commit is not the build's (build.json). Results come out as a REPRO bundle with its own hash. The CLI refuses with exit 2.
• Release hygiene:
  • Every npm dependency is pinned exactly, including Electron 42.11.16 and electron-builder 26.18.0. CI uses npm ci only.
  • Model downloads are SHA-256-pinned (tools/fetch-model.mjs), and arch/encoder checkpoints are pinned by HF commit plus SHA-256 (tools/fetch-arch-models.mjs).
  • Android versionCode now follows the version (1600).
  • The CI signs a release APK with apksigner (Neuralese-<ver>.apk, RSA-4096 key, cert SHA-256 4E:26:13:83:89:A9:15:38:D7:87:DA:9C:8F:6A:2A:D7:C9:DF:1D:81:68:E2:DD:20:37:FA:14:12:85:F5:8B:0B) from the ANDROID_KEYSTORE_B64 / ANDROID_KEYSTORE_PASSWORD / ANDROID_KEY_ALIAS repo secrets. If they are absent, it falls back to the debug-signed APK. Earlier APKs were debug-signed, so Android will refuse to update them in place: uninstall the old app first.
  • Desktop builds are unsigned (no Authenticode / no code-signing certificate). Windows SmartScreen will warn. Verify downloads with SHA256SUMS.txt.
  • The iOS IPA is unsigned as before.
• UI: a "NEW v16" tab with what's new and the WebGPU button, plus the H-INTERP tab in the HUB group.
• Tests: 184 unit tests (arch 5, encoder 5, prereg 4, hygiene 3 new) + engine-worker smoke + test/smoke/webgpu.smoke.mjs + test/smoke/v16.smoke.mjs. CI now fetches the ~3 GB of pinned checkpoints, so the arch and encoder tests actually run there.

Neuralese v15.3.0: engineering pass (KV/prefix cache, engine Web Worker, stricter CSP, repo restructure)

Models, maths and data are unchanged. This release is about speed, safety and repo hygiene, and the claims below are each backed by a test.

• KV cache (honest finding): engine.js already had an incremental per-sequence KV cache (the audit's "no KV cache" was wrong). New: an independent full-recompute reference test (test/engine-kv.test.mjs, rel. error ≤ 1e-5 at every position) and an opt-in cross-call prefix KV cache for Engine.run(). It is bit-identical to uncached runs and is invalidated by weight edits (W.__ver), runtime/intervention changes (R), stats/SAE swaps and hooks, which bypass it. Node CPU bench (tools/bench-engine.mjs, 16-tok): KV generation 42.2 → 43.9 tok/s (noise-level; the cache already existed), no-cache recompute 6.5 tok/s, repeated run() 267 ms → 9 ms with the prefix cache.
• Engine Web Worker: generation runs in src/engine-worker.js through RemoteEngine (src/engine-client.js), with a main-thread fallback (?noworker or if the worker cannot start). The smoke test proves the worker produces exactly the main-thread tokens. Lab tools that need JS hooks still use the synchronous main-thread Engine, because functions cannot cross threads. Caveat: on a headless software-GL box, worker generation was slower in wall time (2.5 vs 4.4 tok/s) because each token needs a message round-trip to a busy main thread. Fewer main-thread long tasks occurred (3 vs 15), but the initial weight sync is still one ~2.5 s task.
• CSP: 'unsafe-eval' is replaced by 'wasm-unsafe-eval'. The notebook runs cells as Blob-URL modules (blob: in script-src), and inline <style> elements are moved into style.css. Inline style attributes are still allowed (style-src-attr 'unsafe-inline', ~100 dynamic styles remain). Smoke: 0 CSP violations across 73 tabs.
• Desktop SwiftShader: it is no longer forced on. It is enabled only by --swiftshader, NEURALESE_SWIFTSHADER=1, or automatically after a GPU-process crash or missing WebGL (persisted marker plus relaunch).
• Repo restructure: app source is in src/, tests are named by feature (test/*.test.mjs, test/smoke/*.smoke.mjs), PyTorch references are in test/ref/, and old progress docs are in docs/progress/. Duplicate WASM artifacts were removed, with a hash manifest (src/tsotchke/ARTIFACTS.sha256) checked by test/repo-layout.test.mjs.
• CI: CI still runs only on tags or manual dispatch, and now runs eslint, the unit tests and a headless-Chrome smoke before building.
• Wording: "Tsotchke-powered" is now "vendored Tsotchke component(s)". The package description credits vendored three.js, transformers.js/ONNX and the Tsotchke WASM builds.

Neuralese v15.2.0: refinement pass

A front-end refinement release. Models, maths and data are unchanged, every visualization still draws only real computed data, and every provenance/honesty banner is kept.

• Sidebar: long tab labels no longer get cut off. They wrap to two lines, the rail is wider (184 px; 160 px under 1180 px), and every item shows its full title as a tooltip. Hover/press micro-motion.
• Neuron Lens atlas: a skeleton/shimmer placeholder replaces the bare canvas. It shows "Atlas not loaded" with a call to action before loading, a live per-layer progress ("layer 3 of 8 · sha256-checking") while loading, and a retry message if the download fails. The placeholder is clearly marked as a placeholder (dashed outlines, no fake points).
• 3D performance on weak GPUs: a GPU tier (low/mid/high) is read from the WebGL renderer string plus core/memory hints. Software renderers (SwiftShader, llvmpipe, Basic Render) and old integrated/mobile GPUs now start at the lightest level (no post-processing, 1× pixels) instead of stuttering. A single quality ladder is shared with the adaptive loop, and the level the adaptive loop settles on is remembered for the next session. The VIZ fps readout shows the tier.
• Action states everywhere: every action button wired through the shared guard shows a spinner (aria-busy) while it runs. A failure leaves a dismissible inline error note on that panel in addition to the toast; the next success clears it.
• Theme: on first run the theme follows the OS colour scheme (prefers-color-scheme). After you pick one with the toggle, your choice wins.
• Consistency + mobile: unified fieldset/legend/row spacing on the type and spacing tokens, press feedback on buttons, and chart hover rings. On phones: 16 px inputs (no iOS zoom-on-focus), 38 px touch targets, larger drawer items, and safe-area bottom padding. All motion respects prefers-reduced-motion.
• Tests: 8 new unit tests (test/v152.test.mjs) and a new smoke-v152.mjs (all 73 tabs × 3 viewports, 0 console errors, plus sidebar-truncation, atlas skeleton → loaded, GPU-tier ladder, busy/error state and theme checks).

Neuralese v15.1.0: polish pass

A front-end polish release. Models, maths and data are unchanged, and every provenance/honesty banner is kept.

• Overview dashboard (new landing tab, H): a hero 3D network (8 rings × 16 instanced head nodes, coloured live by the real per-head ‖z_h‖, flow particles, parallax). Live stat cards (params 19.70 M, layers·heads, d_model·d_mlp, tok/s, ms/token, edits·steer, JS heap, weights hash) with skeleton loading. A tok/s sparkline, a surprisal strip for the last generation, and quick links into each section.
• **3D...

Neuralese v15.3.0

Neuralese v15.3.0: engineering pass (KV/prefix cache, engine Web Worker, stricter CSP, repo restructure)

Models, maths and data are unchanged. This release is about speed, safety and repo hygiene, and the claims below are each backed by a test.

• KV cache (honest finding): engine.js already had an incremental per-sequence KV cache (the audit's "no KV cache" was wrong). New: an independent full-recompute reference test (test/engine-kv.test.mjs, rel. error ≤ 1e-5 at every position) and an opt-in cross-call prefix KV cache for Engine.run(). It is bit-identical to uncached runs and is invalidated by weight edits (W.__ver), runtime/intervention changes (R), stats/SAE swaps and hooks, which bypass it. Node CPU bench (tools/bench-engine.mjs, 16-tok): KV generation 42.2 → 43.9 tok/s (noise-level; the cache already existed), no-cache recompute 6.5 tok/s, repeated run() 267 ms → 9 ms with the prefix cache.
• Engine Web Worker: generation runs in src/engine-worker.js through RemoteEngine (src/engine-client.js), with a main-thread fallback (?noworker or if the worker cannot start). The smoke test proves the worker produces exactly the main-thread tokens. Lab tools that need JS hooks still use the synchronous main-thread Engine, because functions cannot cross threads. Caveat: on a headless software-GL box, worker generation was slower in wall time (2.5 vs 4.4 tok/s) because each token needs a message round-trip to a busy main thread. Fewer main-thread long tasks occurred (3 vs 15), but the initial weight sync is still one ~2.5 s task.
• CSP: 'unsafe-eval' is replaced by 'wasm-unsafe-eval'. The notebook runs cells as Blob-URL modules (blob: in script-src), and inline <style> elements are moved into style.css. Inline style attributes are still allowed (style-src-attr 'unsafe-inline', ~100 dynamic styles remain). Smoke: 0 CSP violations across 73 tabs.
• Desktop SwiftShader: it is no longer forced on. It is enabled only by --swiftshader, NEURALESE_SWIFTSHADER=1, or automatically after a GPU-process crash or missing WebGL (persisted marker plus relaunch).
• Repo restructure: app source is in src/, tests are named by feature (test/*.test.mjs, test/smoke/*.smoke.mjs), PyTorch references are in test/ref/, and old progress docs are in docs/progress/. Duplicate WASM artifacts were removed, with a hash manifest (src/tsotchke/ARTIFACTS.sha256) checked by test/repo-layout.test.mjs.
• CI: CI still runs only on tags or manual dispatch, and now runs eslint, the unit tests and a headless-Chrome smoke before building.
• Wording: "Tsotchke-powered" is now "vendored Tsotchke component(s)". The package description credits vendored three.js, transformers.js/ONNX and the Tsotchke WASM builds.

Neuralese v15.2.0: refinement pass

A front-end refinement release. Models, maths and data are unchanged, every visualization still draws only real computed data, and every provenance/honesty banner is kept.

• Sidebar: long tab labels no longer get cut off. They wrap to two lines, the rail is wider (184 px; 160 px under 1180 px), and every item shows its full title as a tooltip. Hover/press micro-motion.
• Neuron Lens atlas: a skeleton/shimmer placeholder replaces the bare canvas. It shows "Atlas not loaded" with a call to action before loading, a live per-layer progress ("layer 3 of 8 · sha256-checking") while loading, and a retry message if the download fails. The placeholder is clearly marked as a placeholder (dashed outlines, no fake points).
• 3D performance on weak GPUs: a GPU tier (low/mid/high) is read from the WebGL renderer string plus core/memory hints. Software renderers (SwiftShader, llvmpipe, Basic Render) and old integrated/mobile GPUs now start at the lightest level (no post-processing, 1× pixels) instead of stuttering. A single quality ladder is shared with the adaptive loop, and the level the adaptive loop settles on is remembered for the next session. The VIZ fps readout shows the tier.
• Action states everywhere: every action button wired through the shared guard shows a spinner (aria-busy) while it runs. A failure leaves a dismissible inline error note on that panel in addition to the toast; the next success clears it.
• Theme: on first run the theme follows the OS colour scheme (prefers-color-scheme). After you pick one with the toggle, your choice wins.
• Consistency + mobile: unified fieldset/legend/row spacing on the type and spacing tokens, press feedback on buttons, and chart hover rings. On phones: 16 px inputs (no iOS zoom-on-focus), 38 px touch targets, larger drawer items, and safe-area bottom padding. All motion respects prefers-reduced-motion.
• Tests: 8 new unit tests (test/v152.test.mjs) and a new smoke-v152.mjs (all 73 tabs × 3 viewports, 0 console errors, plus sidebar-truncation, atlas skeleton → loaded, GPU-tier ladder, busy/error state and theme checks).

Neuralese v15.1.0: polish pass

A front-end polish release. Models, maths and data are unchanged, and every provenance/honesty banner is kept.

• Overview dashboard (new landing tab, H): a hero 3D network (8 rings × 16 instanced head nodes, coloured live by the real per-head ‖z_h‖, flow particles, parallax). Live stat cards (params 19.70 M, layers·heads, d_model·d_mlp, tok/s, ms/token, edits·steer, JS heap, weights hash) with skeleton loading. A tok/s sparkline, a surprisal strip for the last generation, and quick links into each section.
• 3D scene:
  • Heads are drawn as one InstancedMesh. Animated activation flow runs along head→residual edges and up the residual column, with amplitude driven by the real activations.
  • Hover picking shows an info card (head ‖z‖ and rank, layer, residual norm).
  • Eased fly-to camera moves.
  • A custom post-processing chain: HDR MSAA target, a bloom pyramid, and depth of field from the depth texture. Toggle it under VIZ → rendering.
  • Adaptive quality drops DoF, then bloom, then pixel ratio when the 3-second average falls below 50 fps. A live fps/CPU-ms readout is shown.
• Data viz:
  • Charts animate between states (eased tweens).
  • Line charts support brush-to-zoom (double-click resets).
  • Charts are cross-highlighted through a shared token bus.
  • New attention arc diagram (hover a token to isolate its arcs, top-4 tooltip) and token heat strips (attention received; per-token surprisal under the Generate output, synced with the text).
• UI:
  • An inline SVG icon set and a per-section accent colour (sidebar, pane header, primary buttons).
  • Refined heading/legend hierarchy, visible :focus-visible rings, and AA-checked muted text (≥ 4.5:1 in both themes).
  • Empty-canvas states ("No data yet") and error toasts.
  • A keyboard shortcuts overlay (?). Use [ / ] for the previous/next tab.
  • Button ripples. Everything respects prefers-reduced-motion: no tweens, a static hero, frozen flow and instant camera moves.
• Tests: a new test/v151.test.mjs file, and a new smoke-v151.mjs that covers all 73 tabs at 3 sizes with 0 console errors. It also checks the overview stats, hero, icons/accents, shortcut overlay, [/]/H navigation, arcs + strip + cross-highlight, brush zoom, the 3D hover card + fly-to, contrast samples, reduced motion and fps.

Neuralese v15.0.0: visual overhaul

A front-end redesign. Models, maths and data are unchanged, and every provenance/honesty banner is kept.

• Design system: a dark "research lab" look built from CSS tokens: surfaces, text, accent palette, type scale, spacing, radii, elevation, focus rings and motion. Inter and JetBrains Mono are self-hosted (variable WOFF2, SIL OFL licences in vendor/fonts/), with no font CDN. There is a matching light theme, and prefers-reduced-motion turns off all animation.
• Navigation: the 72-tab wall is replaced by a grouped, collapsible sidebar.
  • Live search (press /); tabs are labelled with their descriptive titles.
  • A sticky pane header shows the group, the tab and its title.
  • On phones the sidebar is a drawer behind ☰. The ⌘K command palette is restyled.
• Components: panels are cards. Buttons, inputs, selects, checkboxes, radios and sliders, sticky-header tables, tooltips, toasts, chips and scrollbars are restyled. Charts that have not been rendered yet show an empty-state pattern, and there is a shimmer for busy elements.
• Charts (charts.js, rewritten, same API):
  • Crisp HiDPI rendering with nice-number ticks, gridlines, axes and legends.
  • Hover crosshairs and markers; area fills and soft glow on lines; rounded gradient bars with hover; scatter plots with gridlines and glow.
  • Heatmaps use perceptual colormaps (viridis and magma, plus a dark-centred blue→orange diverging map) and draw a colorbar.
• 3D:
  • ACES tone mapping with key, rim and fill lights and a gradient backdrop.
  • Additive glow halos on attention heads scale with ‖z_h‖, and MLP neurons are drawn as soft round points.
  • Glass layer slabs with edge outlines, layer labels and a ground grid.
  • Fit-to-view ignores the grid and accounts for the taller header. The heightfield uses the diverging palette.
• Tests:
  • test/v15.test.mjs adds 4 tests, for a total of 149.
  • smoke-v15.mjs opens every tab at 1440×900, 1240×700 and 390×844. It asserts each pane and header renders, the page never overflows horizontally, and there are 0 console errors. It also checks the sidebar search and collapse, the mobile drawer and that the fonts load.
• Cache busting: cache busters are now ?v=15.

Neuralese v14.9.0

All 14.x criteria were pre-registered in docs/PREREG-v14x.md (sha256 7f864197…93a1, pinned by test V14X-1) before any data. Results: docs/CLAIMS-v14x.md; failures: docs/REPORT-v14.md.

14.9.0

• Headless-Chrome fit of the full TS-1M lens through lens-worker.js: cosine 1.000000 with the Node fit for all 31 blocks (C9a supported). Scope: CPU Web Workers,...

Neuralese v15.2.0

Neuralese v15.2.0: refinement pass

A front-end refinement release. Models, maths and data are unchanged, every visualization still draws only real computed data, and every provenance/honesty banner is kept.

• Sidebar: long tab labels no longer get cut off. They wrap to two lines, the rail is wider (184 px; 160 px under 1180 px), and every item shows its full title as a tooltip. Hover/press micro-motion.
• Neuron Lens atlas: a skeleton/shimmer placeholder replaces the bare canvas. It shows "Atlas not loaded" with a call to action before loading, a live per-layer progress ("layer 3 of 8 · sha256-checking") while loading, and a retry message if the download fails. The placeholder is clearly marked as a placeholder (dashed outlines, no fake points).
• 3D performance on weak GPUs: a GPU tier (low/mid/high) is read from the WebGL renderer string plus core/memory hints. Software renderers (SwiftShader, llvmpipe, Basic Render) and old integrated/mobile GPUs now start at the lightest level (no post-processing, 1× pixels) instead of stuttering. A single quality ladder is shared with the adaptive loop, and the level the adaptive loop settles on is remembered for the next session. The VIZ fps readout shows the tier.
• Action states everywhere: every action button wired through the shared guard shows a spinner (aria-busy) while it runs. A failure leaves a dismissible inline error note on that panel in addition to the toast; the next success clears it.
• Theme: on first run the theme follows the OS colour scheme (prefers-color-scheme). After you pick one with the toggle, your choice wins.
• Consistency + mobile: unified fieldset/legend/row spacing on the type and spacing tokens, press feedback on buttons, and chart hover rings. On phones: 16 px inputs (no iOS zoom-on-focus), 38 px touch targets, larger drawer items, and safe-area bottom padding. All motion respects prefers-reduced-motion.
• Tests: 8 new unit tests (test/v152.test.mjs) and a new smoke-v152.mjs (all 73 tabs × 3 viewports, 0 console errors, plus sidebar-truncation, atlas skeleton → loaded, GPU-tier ladder, busy/error state and theme checks).

Neuralese v15.1.0: polish pass

A front-end polish release. Models, maths and data are unchanged, and every provenance/honesty banner is kept.

• Overview dashboard (new landing tab, H): a hero 3D network (8 rings × 16 instanced head nodes, coloured live by the real per-head ‖z_h‖, flow particles, parallax). Live stat cards (params 19.70 M, layers·heads, d_model·d_mlp, tok/s, ms/token, edits·steer, JS heap, weights hash) with skeleton loading. A tok/s sparkline, a surprisal strip for the last generation, and quick links into each section.
• 3D scene:
  • Heads are drawn as one InstancedMesh. Animated activation flow runs along head→residual edges and up the residual column, with amplitude driven by the real activations.
  • Hover picking shows an info card (head ‖z‖ and rank, layer, residual norm).
  • Eased fly-to camera moves.
  • A custom post-processing chain: HDR MSAA target, a bloom pyramid, and depth of field from the depth texture. Toggle it under VIZ → rendering.
  • Adaptive quality drops DoF, then bloom, then pixel ratio when the 3-second average falls below 50 fps. A live fps/CPU-ms readout is shown.
• Data viz:
  • Charts animate between states (eased tweens).
  • Line charts support brush-to-zoom (double-click resets).
  • Charts are cross-highlighted through a shared token bus.
  • New attention arc diagram (hover a token to isolate its arcs, top-4 tooltip) and token heat strips (attention received; per-token surprisal under the Generate output, synced with the text).
• UI:
  • An inline SVG icon set and a per-section accent colour (sidebar, pane header, primary buttons).
  • Refined heading/legend hierarchy, visible :focus-visible rings, and AA-checked muted text (≥ 4.5:1 in both themes).
  • Empty-canvas states ("No data yet") and error toasts.
  • A keyboard shortcuts overlay (?). Use [ / ] for the previous/next tab.
  • Button ripples. Everything respects prefers-reduced-motion: no tweens, a static hero, frozen flow and instant camera moves.
• Tests: a new test/v151.test.mjs file, and a new smoke-v151.mjs that covers all 73 tabs at 3 sizes with 0 console errors. It also checks the overview stats, hero, icons/accents, shortcut overlay, [/]/H navigation, arcs + strip + cross-highlight, brush zoom, the 3D hover card + fly-to, contrast samples, reduced motion and fps.

Neuralese v15.0.0: visual overhaul

A front-end redesign. Models, maths and data are unchanged, and every provenance/honesty banner is kept.

• Design system: a dark "research lab" look built from CSS tokens: surfaces, text, accent palette, type scale, spacing, radii, elevation, focus rings and motion. Inter and JetBrains Mono are self-hosted (variable WOFF2, SIL OFL licences in vendor/fonts/), with no font CDN. There is a matching light theme, and prefers-reduced-motion turns off all animation.
• Navigation: the 72-tab wall is replaced by a grouped, collapsible sidebar.
  • Live search (press /); tabs are labelled with their descriptive titles.
  • A sticky pane header shows the group, the tab and its title.
  • On phones the sidebar is a drawer behind ☰. The ⌘K command palette is restyled.
• Components: panels are cards. Buttons, inputs, selects, checkboxes, radios and sliders, sticky-header tables, tooltips, toasts, chips and scrollbars are restyled. Charts that have not been rendered yet show an empty-state pattern, and there is a shimmer for busy elements.
• Charts (charts.js, rewritten, same API):
  • Crisp HiDPI rendering with nice-number ticks, gridlines, axes and legends.
  • Hover crosshairs and markers; area fills and soft glow on lines; rounded gradient bars with hover; scatter plots with gridlines and glow.
  • Heatmaps use perceptual colormaps (viridis and magma, plus a dark-centred blue→orange diverging map) and draw a colorbar.
• 3D:
  • ACES tone mapping with key, rim and fill lights and a gradient backdrop.
  • Additive glow halos on attention heads scale with ‖z_h‖, and MLP neurons are drawn as soft round points.
  • Glass layer slabs with edge outlines, layer labels and a ground grid.
  • Fit-to-view ignores the grid and accounts for the taller header. The heightfield uses the diverging palette.
• Tests:
  • test/v15.test.mjs adds 4 tests, for a total of 149.
  • smoke-v15.mjs opens every tab at 1440×900, 1240×700 and 390×844. It asserts each pane and header renders, the page never overflows horizontally, and there are 0 console errors. It also checks the sidebar search and collapse, the mobile drawer and that the fonts load.
• Cache busting: cache busters are now ?v=15.

Neuralese v14.9.0

All 14.x criteria were pre-registered in docs/PREREG-v14x.md (sha256 7f864197…93a1, pinned by test V14X-1) before any data. Results: docs/CLAIMS-v14x.md; failures: docs/REPORT-v14.md.

14.9.0

• Headless-Chrome fit of the full TS-1M lens through lens-worker.js: cosine 1.000000 with the Node fit for all 31 blocks (C9a supported). Scope: CPU Web Workers, not WebGPU.
• docs/REPORT-v14.md lists every failed prediction from 14.0–14.8; test V14X-7 enforces it (C9b).
• test/v14x.test.mjs (8 tests); smoke-v14 covers the NESTED tabs. 145/145 tests, SMOKE OK.

14.8.0

• Host × nested cross-read. C8a not supported, as predicted: host-head reads of baked-guest units are exactly 0.
• C8b supported: the emulator's mode-0 and mode-1 reads of the same unit agree (median r 0.888 vs τ99 0.637).

14.7.0

• New NESTED group with four tabs:
  • RISC-V IN WEIGHTS, SLM IN SLM and MoE IN MoE, integrated from the neuralese-nested workstream (nested/).
  • NESTED LENS, with results and a unit reader.
• The nested SLM weights ship as float32 (nested/slm/models.f32, 1.5 MB, sha256-checked) instead of 6.2 MB of float64 JSON. Max KL against float64 is 2.4e-16.
• nestedlens.js adds a finite-difference R-lens. R.out passes the steering null on the composite (384/384) and the emulator (243/249). C7 supported.

14.6.0

• Taylor check of the RelP linearisation on TS-8M. C6 not supported at all 7 layers, as predicted.
• The pre-registered δ=1e-4 finite difference is float32-noise-limited. An exploratory δ=1e-2 rerun is included and labelled as such.

14.5.0

• Hungarian matching of TS-3M ↔ TS-8M neurons on out·R top-50 Jaccard. C5 not supported: 11/48 = 22.9% of pairs exceed τ99, against the ≥30% criterion.

14.4.0

• Auto-labels from S.mid reads.
• C4a supported: hit rate 12.95% vs 0.16% shuffled, McNemar p≈2e-295.
• C4b supported: logistic-z confidence is calibrated within 0.033 in every bin.

14.3.0

• Head attribution through the lens vs zero-ablation (node level; scope change). C3 not supported:
  • induction: recall 0.31, AUROC 0.63
  • IOI: recall 0.18, AUROC 0.48
  • the logit baseline is equally poor.
• forward gains ablate and zs options.

14.2.0

• SAE features (sae.js, m=1024, L3) through R.out pass 64/64, matching neurons at 64/64. C2 supported, but at ceiling, and the SAE is weakly sparse.

14.1.0

• TS-1M and TS-3M lenses (lens/ts1m, lens/ts3m; scope change from distilgpt2/GPT-2).
• C1a supported: R.out beats the logit lens at 7/7 layers on both models, null 0.011.
• C1b supported: split-half cosine ≥0.9927 for every block of 1M/3M/8M.
• C1c supported: the 8M rerun at 64 neurons per layer gives narrower CIs at 7/7 layers.
• Weak-spot fix: the NEURON LENS steer demo now averages Δ over positions and ranks only DB-vocabulary tokens.

14.0.0

• Neuron Lens: original JS implementation of R/S/mid lenses (RelP tape, forward/reverse doc Jacobians) for TinyStories-8M; prior-art credit in docs/NEURON_LENS_METHOD.md.
• lens/ts8m: lenses.f32 (sha256 6a668f3e…7636), validation.json, drivers.json, per-layer neuron/head DB with db-manifest sha256s.
• New tabs NEURON LENS / LENS VALID / LENS CATALOG; Web Wo...

Neuralese v15.1.0

Neuralese v15.1.0: polish pass

A front-end polish release. Models, maths and data are unchanged, and every provenance/honesty banner is kept.

• Overview dashboard (new landing tab, H): a hero 3D network (8 rings × 16 instanced head nodes, coloured live by the real per-head ‖z_h‖, flow particles, parallax). Live stat cards (params 19.70 M, layers·heads, d_model·d_mlp, tok/s, ms/token, edits·steer, JS heap, weights hash) with skeleton loading. A tok/s sparkline, a surprisal strip for the last generation, and quick links into each section.
• 3D scene:
  • Heads are drawn as one InstancedMesh. Animated activation flow runs along head→residual edges and up the residual column, with amplitude driven by the real activations.
  • Hover picking shows an info card (head ‖z‖ and rank, layer, residual norm).
  • Eased fly-to camera moves.
  • A custom post-processing chain: HDR MSAA target, a bloom pyramid, and depth of field from the depth texture. Toggle it under VIZ → rendering.
  • Adaptive quality drops DoF, then bloom, then pixel ratio when the 3-second average falls below 50 fps. A live fps/CPU-ms readout is shown.
• Data viz:
  • Charts animate between states (eased tweens).
  • Line charts support brush-to-zoom (double-click resets).
  • Charts are cross-highlighted through a shared token bus.
  • New attention arc diagram (hover a token to isolate its arcs, top-4 tooltip) and token heat strips (attention received; per-token surprisal under the Generate output, synced with the text).
• UI:
  • An inline SVG icon set and a per-section accent colour (sidebar, pane header, primary buttons).
  • Refined heading/legend hierarchy, visible :focus-visible rings, and AA-checked muted text (≥ 4.5:1 in both themes).
  • Empty-canvas states ("No data yet") and error toasts.
  • A keyboard shortcuts overlay (?). Use [ / ] for the previous/next tab.
  • Button ripples. Everything respects prefers-reduced-motion: no tweens, a static hero, frozen flow and instant camera moves.
• Tests: a new test/v151.test.mjs file, and a new smoke-v151.mjs that covers all 73 tabs at 3 sizes with 0 console errors. It also checks the overview stats, hero, icons/accents, shortcut overlay, [/]/H navigation, arcs + strip + cross-highlight, brush zoom, the 3D hover card + fly-to, contrast samples, reduced motion and fps.

Neuralese v15.0.0: visual overhaul

A front-end redesign. Models, maths and data are unchanged, and every provenance/honesty banner is kept.

• Design system: a dark "research lab" look built from CSS tokens: surfaces, text, accent palette, type scale, spacing, radii, elevation, focus rings and motion. Inter and JetBrains Mono are self-hosted (variable WOFF2, SIL OFL licences in vendor/fonts/), with no font CDN. There is a matching light theme, and prefers-reduced-motion turns off all animation.
• Navigation: the 72-tab wall is replaced by a grouped, collapsible sidebar.
  • Live search (press /); tabs are labelled with their descriptive titles.
  • A sticky pane header shows the group, the tab and its title.
  • On phones the sidebar is a drawer behind ☰. The ⌘K command palette is restyled.
• Components: panels are cards. Buttons, inputs, selects, checkboxes, radios and sliders, sticky-header tables, tooltips, toasts, chips and scrollbars are restyled. Charts that have not been rendered yet show an empty-state pattern, and there is a shimmer for busy elements.
• Charts (charts.js, rewritten, same API):
  • Crisp HiDPI rendering with nice-number ticks, gridlines, axes and legends.
  • Hover crosshairs and markers; area fills and soft glow on lines; rounded gradient bars with hover; scatter plots with gridlines and glow.
  • Heatmaps use perceptual colormaps (viridis and magma, plus a dark-centred blue→orange diverging map) and draw a colorbar.
• 3D:
  • ACES tone mapping with key, rim and fill lights and a gradient backdrop.
  • Additive glow halos on attention heads scale with ‖z_h‖, and MLP neurons are drawn as soft round points.
  • Glass layer slabs with edge outlines, layer labels and a ground grid.
  • Fit-to-view ignores the grid and accounts for the taller header. The heightfield uses the diverging palette.
• Tests:
  • test/v15.test.mjs adds 4 tests, for a total of 149.
  • smoke-v15.mjs opens every tab at 1440×900, 1240×700 and 390×844. It asserts each pane and header renders, the page never overflows horizontally, and there are 0 console errors. It also checks the sidebar search and collapse, the mobile drawer and that the fonts load.
• Cache busting: cache busters are now ?v=15.

Neuralese v14.9.0

All 14.x criteria were pre-registered in docs/PREREG-v14x.md (sha256 7f864197…93a1, pinned by test V14X-1) before any data. Results: docs/CLAIMS-v14x.md; failures: docs/REPORT-v14.md.

14.9.0

• Headless-Chrome fit of the full TS-1M lens through lens-worker.js: cosine 1.000000 with the Node fit for all 31 blocks (C9a supported). Scope: CPU Web Workers, not WebGPU.
• docs/REPORT-v14.md lists every failed prediction from 14.0–14.8; test V14X-7 enforces it (C9b).
• test/v14x.test.mjs (8 tests); smoke-v14 covers the NESTED tabs. 145/145 tests, SMOKE OK.

14.8.0

• Host × nested cross-read. C8a not supported, as predicted: host-head reads of baked-guest units are exactly 0.
• C8b supported: the emulator's mode-0 and mode-1 reads of the same unit agree (median r 0.888 vs τ99 0.637).

14.7.0

• New NESTED group with four tabs:
  • RISC-V IN WEIGHTS, SLM IN SLM and MoE IN MoE, integrated from the neuralese-nested workstream (nested/).
  • NESTED LENS, with results and a unit reader.
• The nested SLM weights ship as float32 (nested/slm/models.f32, 1.5 MB, sha256-checked) instead of 6.2 MB of float64 JSON. Max KL against float64 is 2.4e-16.
• nestedlens.js adds a finite-difference R-lens. R.out passes the steering null on the composite (384/384) and the emulator (243/249). C7 supported.

14.6.0

• Taylor check of the RelP linearisation on TS-8M. C6 not supported at all 7 layers, as predicted.
• The pre-registered δ=1e-4 finite difference is float32-noise-limited. An exploratory δ=1e-2 rerun is included and labelled as such.

14.5.0

• Hungarian matching of TS-3M ↔ TS-8M neurons on out·R top-50 Jaccard. C5 not supported: 11/48 = 22.9% of pairs exceed τ99, against the ≥30% criterion.

14.4.0

• Auto-labels from S.mid reads.
• C4a supported: hit rate 12.95% vs 0.16% shuffled, McNemar p≈2e-295.
• C4b supported: logistic-z confidence is calibrated within 0.033 in every bin.

14.3.0

• Head attribution through the lens vs zero-ablation (node level; scope change). C3 not supported:
  • induction: recall 0.31, AUROC 0.63
  • IOI: recall 0.18, AUROC 0.48
  • the logit baseline is equally poor.
• forward gains ablate and zs options.

14.2.0

• SAE features (sae.js, m=1024, L3) through R.out pass 64/64, matching neurons at 64/64. C2 supported, but at ceiling, and the SAE is weakly sparse.

14.1.0

• TS-1M and TS-3M lenses (lens/ts1m, lens/ts3m; scope change from distilgpt2/GPT-2).
• C1a supported: R.out beats the logit lens at 7/7 layers on both models, null 0.011.
• C1b supported: split-half cosine ≥0.9927 for every block of 1M/3M/8M.
• C1c supported: the 8M rerun at 64 neurons per layer gives narrower CIs at 7/7 layers.
• Weak-spot fix: the NEURON LENS steer demo now averages Δ over positions and ranks only DB-vocabulary tokens.

14.0.0

• Neuron Lens: original JS implementation of R/S/mid lenses (RelP tape, forward/reverse doc Jacobians) for TinyStories-8M; prior-art credit in docs/NEURON_LENS_METHOD.md.
• lens/ts8m: lenses.f32 (sha256 6a668f3e…7636), validation.json, drivers.json, per-layer neuron/head DB with db-manifest sha256s.
• New tabs NEURON LENS / LENS VALID / LENS CATALOG; Web Worker refit.
• Pre-registered validation (docs/PREREG-v14.md + Amendment 1): P1, P2, P5 supported; P3, P4 not supported (docs/CLAIMS-v14.md).
• Attributed CC BY 4.0 sample of stdlogits/neuron-lens dataset in lens/ext (no contexts).
• ROADMAP-v14 for v14.1–v14.9.

13.0.1

• Electron 38 → 42.11.16 and puppeteer-core 24 → 25.16.0 (Dependabot security updates, PR #2).
• Electron 42 no longer falls back to software WebGL by itself, so on machines without a usable GPU the lab stopped at "Error creating WebGL context". desktop/main.cjs now enables the SwiftShader fallback.

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

v13 packages the v12.1 lab as native apps. The lab itself is unchanged. It is the same in-browser interpretability and weight-surgery lab for TinyStories-8M.

Downloads

platform	file	notes
Windows x64	Neuralese-Setup-13.0.1-x64.exe	NSIS installer. Unsigned, so SmartScreen will warn: More info → Run anyway
Windows x64	Neuralese-13.0.1-x64-portable.exe	runs without installing. Unsigned
Linux x64	Neuralese-13.0.1-x86_64.AppImage	chmod +x, then run. Some distros need libfuse2
Linux x64	Neuralese-13.0.1-amd64.deb	sudo apt install ./Neuralese-13.0.1-amd64.deb
Android	Neuralese-13.0.1-debug.apk	debug build. Allow "install unknown apps"
iOS device	Neuralese-13.0.1-unsigned.ipa	unsigned. Re-sign with AltStore or Sideloadly
iOS Simulator	Neuralese-13.0.1-simulator.app.zip	unzip, then xcrun simctl install booted App.app

SHA256SUMS.txt lists the hash of every file.

Honest limitations

• None of the builds is code-signed: Windows SmartScreen, macOS-style Gatekeeper prompts on Linux desktops, and an unsigned iOS IPA. There is no auto-update.
• On first run the model (~112 MB) and any HUB models download from Hugging Face, so the apps need internet then. After that, the browser cache serves them.
• Mobile performance: the lab is heavy. Some tabs (3D, HUB ONNX models, long sweeps) are slow or memory-limited on phones.
• The Android build is a debug APK, not a Play Store release.

Neuralese v15.0.0

Neuralese v15.0.0: visual overhaul

A front-end redesign. Models, maths and data are unchanged, and every provenance/honesty banner is kept.

• Design system: a dark "research lab" look built from CSS tokens: surfaces, text, accent palette, type scale, spacing, radii, elevation, focus rings and motion. Inter and JetBrains Mono are self-hosted (variable WOFF2, SIL OFL licences in vendor/fonts/), with no font CDN. There is a matching light theme, and prefers-reduced-motion turns off all animation.
• Navigation: the 72-tab wall is replaced by a grouped, collapsible sidebar.
  • Live search (press /); tabs are labelled with their descriptive titles.
  • A sticky pane header shows the group, the tab and its title.
  • On phones the sidebar is a drawer behind ☰. The ⌘K command palette is restyled.
• Components: panels are cards. Buttons, inputs, selects, checkboxes, radios and sliders, sticky-header tables, tooltips, toasts, chips and scrollbars are restyled. Charts that have not been rendered yet show an empty-state pattern, and there is a shimmer for busy elements.
• Charts (charts.js, rewritten, same API):
  • Crisp HiDPI rendering with nice-number ticks, gridlines, axes and legends.
  • Hover crosshairs and markers; area fills and soft glow on lines; rounded gradient bars with hover; scatter plots with gridlines and glow.
  • Heatmaps use perceptual colormaps (viridis and magma, plus a dark-centred blue→orange diverging map) and draw a colorbar.
• 3D:
  • ACES tone mapping with key, rim and fill lights and a gradient backdrop.
  • Additive glow halos on attention heads scale with ‖z_h‖, and MLP neurons are drawn as soft round points.
  • Glass layer slabs with edge outlines, layer labels and a ground grid.
  • Fit-to-view ignores the grid and accounts for the taller header. The heightfield uses the diverging palette.
• Tests:
  • test/v15.test.mjs adds 4 tests, for a total of 149.
  • smoke-v15.mjs opens every tab at 1440×900, 1240×700 and 390×844. It asserts each pane and header renders, the page never overflows horizontally, and there are 0 console errors. It also checks the sidebar search and collapse, the mobile drawer and that the fonts load.
• Cache busting: cache busters are now ?v=15.

Neuralese v14.9.0

All 14.x criteria were pre-registered in docs/PREREG-v14x.md (sha256 7f864197…93a1, pinned by test V14X-1) before any data. Results: docs/CLAIMS-v14x.md; failures: docs/REPORT-v14.md.

14.9.0

• Headless-Chrome fit of the full TS-1M lens through lens-worker.js: cosine 1.000000 with the Node fit for all 31 blocks (C9a supported). Scope: CPU Web Workers, not WebGPU.
• docs/REPORT-v14.md lists every failed prediction from 14.0–14.8; test V14X-7 enforces it (C9b).
• test/v14x.test.mjs (8 tests); smoke-v14 covers the NESTED tabs. 145/145 tests, SMOKE OK.

14.8.0

• Host × nested cross-read. C8a not supported, as predicted: host-head reads of baked-guest units are exactly 0.
• C8b supported: the emulator's mode-0 and mode-1 reads of the same unit agree (median r 0.888 vs τ99 0.637).

14.7.0

• New NESTED group with four tabs:
  • RISC-V IN WEIGHTS, SLM IN SLM and MoE IN MoE, integrated from the neuralese-nested workstream (nested/).
  • NESTED LENS, with results and a unit reader.
• The nested SLM weights ship as float32 (nested/slm/models.f32, 1.5 MB, sha256-checked) instead of 6.2 MB of float64 JSON. Max KL against float64 is 2.4e-16.
• nestedlens.js adds a finite-difference R-lens. R.out passes the steering null on the composite (384/384) and the emulator (243/249). C7 supported.

14.6.0

• Taylor check of the RelP linearisation on TS-8M. C6 not supported at all 7 layers, as predicted.
• The pre-registered δ=1e-4 finite difference is float32-noise-limited. An exploratory δ=1e-2 rerun is included and labelled as such.

14.5.0

• Hungarian matching of TS-3M ↔ TS-8M neurons on out·R top-50 Jaccard. C5 not supported: 11/48 = 22.9% of pairs exceed τ99, against the ≥30% criterion.

14.4.0

• Auto-labels from S.mid reads.
• C4a supported: hit rate 12.95% vs 0.16% shuffled, McNemar p≈2e-295.
• C4b supported: logistic-z confidence is calibrated within 0.033 in every bin.

14.3.0

• Head attribution through the lens vs zero-ablation (node level; scope change). C3 not supported:
  • induction: recall 0.31, AUROC 0.63
  • IOI: recall 0.18, AUROC 0.48
  • the logit baseline is equally poor.
• forward gains ablate and zs options.

14.2.0

• SAE features (sae.js, m=1024, L3) through R.out pass 64/64, matching neurons at 64/64. C2 supported, but at ceiling, and the SAE is weakly sparse.

14.1.0

• TS-1M and TS-3M lenses (lens/ts1m, lens/ts3m; scope change from distilgpt2/GPT-2).
• C1a supported: R.out beats the logit lens at 7/7 layers on both models, null 0.011.
• C1b supported: split-half cosine ≥0.9927 for every block of 1M/3M/8M.
• C1c supported: the 8M rerun at 64 neurons per layer gives narrower CIs at 7/7 layers.
• Weak-spot fix: the NEURON LENS steer demo now averages Δ over positions and ranks only DB-vocabulary tokens.

14.0.0

• Neuron Lens: original JS implementation of R/S/mid lenses (RelP tape, forward/reverse doc Jacobians) for TinyStories-8M; prior-art credit in docs/NEURON_LENS_METHOD.md.
• lens/ts8m: lenses.f32 (sha256 6a668f3e…7636), validation.json, drivers.json, per-layer neuron/head DB with db-manifest sha256s.
• New tabs NEURON LENS / LENS VALID / LENS CATALOG; Web Worker refit.
• Pre-registered validation (docs/PREREG-v14.md + Amendment 1): P1, P2, P5 supported; P3, P4 not supported (docs/CLAIMS-v14.md).
• Attributed CC BY 4.0 sample of stdlogits/neuron-lens dataset in lens/ext (no contexts).
• ROADMAP-v14 for v14.1–v14.9.

13.0.1

• Electron 38 → 42.11.16 and puppeteer-core 24 → 25.16.0 (Dependabot security updates, PR #2).
• Electron 42 no longer falls back to software WebGL by itself, so on machines without a usable GPU the lab stopped at "Error creating WebGL context". desktop/main.cjs now enables the SwiftShader fallback.

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

v13 packages the v12.1 lab as native apps. The lab itself is unchanged. It is the same in-browser interpretability and weight-surgery lab for TinyStories-8M.

Downloads

platform	file	notes
Windows x64	Neuralese-Setup-13.0.1-x64.exe	NSIS installer. Unsigned, so SmartScreen will warn: More info → Run anyway
Windows x64	Neuralese-13.0.1-x64-portable.exe	runs without installing. Unsigned
Linux x64	Neuralese-13.0.1-x86_64.AppImage	chmod +x, then run. Some distros need libfuse2
Linux x64	Neuralese-13.0.1-amd64.deb	sudo apt install ./Neuralese-13.0.1-amd64.deb
Android	Neuralese-13.0.1-debug.apk	debug build. Allow "install unknown apps"
iOS device	Neuralese-13.0.1-unsigned.ipa	unsigned. Re-sign with AltStore or Sideloadly
iOS Simulator	Neuralese-13.0.1-simulator.app.zip	unzip, then xcrun simctl install booted App.app

SHA256SUMS.txt lists the hash of every file.

Honest limitations

• None of the builds is code-signed: Windows SmartScreen, macOS-style Gatekeeper prompts on Linux desktops, and an unsigned iOS IPA. There is no auto-update.
• On first run the model (~112 MB) and any HUB models download from Hugging Face, so the apps need internet then. After that, the browser cache serves them.
• Mobile performance: the lab is heavy. Some tabs (3D, HUB ONNX models, long sweeps) are slow or memory-limited on phones.
• The Android build is a debug APK, not a Play Store release.

Neuralese v14.9.0

Neuralese v14.9.0

All 14.x criteria were pre-registered in docs/PREREG-v14x.md (sha256 7f864197…93a1, pinned by test V14X-1) before any data. Results: docs/CLAIMS-v14x.md; failures: docs/REPORT-v14.md.

14.9.0

• Headless-Chrome fit of the full TS-1M lens through lens-worker.js: cosine 1.000000 with the Node fit for all 31 blocks (C9a supported). Scope: CPU Web Workers, not WebGPU.
• docs/REPORT-v14.md lists every failed prediction from 14.0–14.8; test V14X-7 enforces it (C9b).
• test/v14x.test.mjs (8 tests); smoke-v14 covers the NESTED tabs. 145/145 tests, SMOKE OK.

14.8.0

• Host × nested cross-read. C8a not supported, as predicted: host-head reads of baked-guest units are exactly 0.
• C8b supported: the emulator's mode-0 and mode-1 reads of the same unit agree (median r 0.888 vs τ99 0.637).

14.7.0

• New NESTED group with four tabs:
  • RISC-V IN WEIGHTS, SLM IN SLM and MoE IN MoE, integrated from the neuralese-nested workstream (nested/).
  • NESTED LENS, with results and a unit reader.
• The nested SLM weights ship as float32 (nested/slm/models.f32, 1.5 MB, sha256-checked) instead of 6.2 MB of float64 JSON. Max KL against float64 is 2.4e-16.
• nestedlens.js adds a finite-difference R-lens. R.out passes the steering null on the composite (384/384) and the emulator (243/249). C7 supported.

14.6.0

• Taylor check of the RelP linearisation on TS-8M. C6 not supported at all 7 layers, as predicted.
• The pre-registered δ=1e-4 finite difference is float32-noise-limited. An exploratory δ=1e-2 rerun is included and labelled as such.

14.5.0

• Hungarian matching of TS-3M ↔ TS-8M neurons on out·R top-50 Jaccard. C5 not supported: 11/48 = 22.9% of pairs exceed τ99, against the ≥30% criterion.

14.4.0

• Auto-labels from S.mid reads.
• C4a supported: hit rate 12.95% vs 0.16% shuffled, McNemar p≈2e-295.
• C4b supported: logistic-z confidence is calibrated within 0.033 in every bin.

14.3.0

• Head attribution through the lens vs zero-ablation (node level; scope change). C3 not supported:
  • induction: recall 0.31, AUROC 0.63
  • IOI: recall 0.18, AUROC 0.48
  • the logit baseline is equally poor.
• forward gains ablate and zs options.

14.2.0

• SAE features (sae.js, m=1024, L3) through R.out pass 64/64, matching neurons at 64/64. C2 supported, but at ceiling, and the SAE is weakly sparse.

14.1.0

• TS-1M and TS-3M lenses (lens/ts1m, lens/ts3m; scope change from distilgpt2/GPT-2).
• C1a supported: R.out beats the logit lens at 7/7 layers on both models, null 0.011.
• C1b supported: split-half cosine ≥0.9927 for every block of 1M/3M/8M.
• C1c supported: the 8M rerun at 64 neurons per layer gives narrower CIs at 7/7 layers.
• Weak-spot fix: the NEURON LENS steer demo now averages Δ over positions and ranks only DB-vocabulary tokens.

14.0.0

• Neuron Lens: original JS implementation of R/S/mid lenses (RelP tape, forward/reverse doc Jacobians) for TinyStories-8M; prior-art credit in docs/NEURON_LENS_METHOD.md.
• lens/ts8m: lenses.f32 (sha256 6a668f3e…7636), validation.json, drivers.json, per-layer neuron/head DB with db-manifest sha256s.
• New tabs NEURON LENS / LENS VALID / LENS CATALOG; Web Worker refit.
• Pre-registered validation (docs/PREREG-v14.md + Amendment 1): P1, P2, P5 supported; P3, P4 not supported (docs/CLAIMS-v14.md).
• Attributed CC BY 4.0 sample of stdlogits/neuron-lens dataset in lens/ext (no contexts).
• ROADMAP-v14 for v14.1–v14.9.

13.0.1

• Electron 38 → 42.11.16 and puppeteer-core 24 → 25.16.0 (Dependabot security updates, PR #2).
• Electron 42 no longer falls back to software WebGL by itself, so on machines without a usable GPU the lab stopped at "Error creating WebGL context". desktop/main.cjs now enables the SwiftShader fallback.

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

v13 packages the v12.1 lab as native apps. The lab itself is unchanged. It is the same in-browser interpretability and weight-surgery lab for TinyStories-8M.

Downloads

platform	file	notes
Windows x64	Neuralese-Setup-13.0.1-x64.exe	NSIS installer. Unsigned, so SmartScreen will warn: More info → Run anyway
Windows x64	Neuralese-13.0.1-x64-portable.exe	runs without installing. Unsigned
Linux x64	Neuralese-13.0.1-x86_64.AppImage	chmod +x, then run. Some distros need libfuse2
Linux x64	Neuralese-13.0.1-amd64.deb	sudo apt install ./Neuralese-13.0.1-amd64.deb
Android	Neuralese-13.0.1-debug.apk	debug build. Allow "install unknown apps"
iOS device	Neuralese-13.0.1-unsigned.ipa	unsigned. Re-sign with AltStore or Sideloadly
iOS Simulator	Neuralese-13.0.1-simulator.app.zip	unzip, then xcrun simctl install booted App.app

SHA256SUMS.txt lists the hash of every file.

Honest limitations

• None of the builds is code-signed: Windows SmartScreen, macOS-style Gatekeeper prompts on Linux desktops, and an unsigned iOS IPA. There is no auto-update.
• On first run the model (~112 MB) and any HUB models download from Hugging Face, so the apps need internet then. After that, the browser cache serves them.
• Mobile performance: the lab is heavy. Some tabs (3D, HUB ONNX models, long sweeps) are slow or memory-limited on phones.
• The Android build is a debug APK, not a Play Store release.