Skip to content

acceptance.md

Latest commit

 

History

History
50 lines (35 loc) · 4.81 KB

acceptance.md

File metadata and controls

50 lines (35 loc) · 4.81 KB
Raw

Pre-merge acceptance — September 5, 2026

Historical v0.1 record. Instructions and measurements below describe that completed pass; current work follows the v0.2.1 execution plan.

All required completion gates 0–6 pass, including one real Learn update. Optional second-producer coverage (M11) also passes. Ablation and poisoning spikes (M12) are deferred; no public security or model-quality claim is made.

Local boundary

  • Branch: model-lab (existing branch preserved; no codex/ prefix).
  • Starting SHA: cad66ada012d464f090a66da5f9bfcebc1a31f47.
  • Final local HEAD: the local acceptance commit containing this report; inspect with git rev-parse HEAD.
  • Pushed: no. GitHub remotes unchanged.
  • Changed paths confined to AGENTS.md, docs/**, and model-lab/**.
  • PRs #1, #7, and #8 were verified open during preflight. No outstanding-PR surfaces, infrastructure repositories, or licensing files were modified.

Reference and numerical acceptance

Reference revision 14fb038816c7aae0bb9342c2dbf1a51dd134a5ff; SHA-256 d47d88c2fd432c8ebdc1048beab7f7eb64ea7e0e664e11b812d72a6d95ebccee. Independent Python oracle agrees with the hash-verified original across 3,676 numerical values with maximum absolute difference 0.0. See provenance for the non-vendoring decision and reproducible comparison command.

The committed 896-parameter canonical state regenerates expected evidence byte-for-byte. TypeScript conforms for every captured forward vector, losses, all gradients, Adam moments/bias corrections/deltas, updated parameters, and post-update logits/probabilities. Tolerance: absolute 1e-10 plus relative 1e-9. Tests also cover connected causal K/V, shared/deep autograd graphs, exact observational invariance, immutable capture/replay, unavailable evidence, bounds, and continued training from complete serialized state.

Checks actually run

Check Result
python3 reference/compare_upstream.py /tmp/model-lab-microgpt.pinned.py 3,676 scalar comparisons; max error 0.0
npm run test:reference 9 Python tests; exact regeneration
npm test 29 TypeScript tests across model, trace, worker/client, and rendering
npm run build strict typecheck and Vite production bundle
npm run test:browser 2 Chromium scenarios: authentic Predict/Learn and desktop/mobile/reset/cancel
npm run test:isolation fresh tracked-source copy; npm ci, reference/tests/build/browser all pass
docker build -f model-lab/Dockerfile -t model-lab:acceptance model-lab isolated image builds
Container on localhost:4180, Chromium Predict + Learn real worker run/update pass; zero page errors and WAN requests
git diff --check and staged equivalent clean after normalizing attached plan's trailing whitespace

Browser tests compare full-precision probability values with the Python fixture, reconstruct Q×K products and scaled logit, verify masked future cells, verify actual displayed optimizer evidence and applied delta, and confirm the same input uses post-update state. Reset restores fixture predictions; immediate cancellation cannot render stale results. Mobile coverage includes the maximum eight-position context. Core browser requests are local; the production application needs no WAN model service or runtime downloads. The acceptance container was stopped and removed after validation.

Independent review and repairs

  • R1 numerical: no blockers; independently reran original-source comparator and reference/model tests.
  • R2 evidence/worker: three confirmed issues repaired and regression-tested: checkpoint identity collisions across histories, stale reset acceptance, and stale worker errors rejecting current-generation work.
  • R3 readability: no blockers; mathematical source has only model-local imports and direct curriculum vocabulary. The source tour matches implementation.
  • R4 build/boundary: no further actionable findings; subtree context and reset/cancel isolation reviewed.
  • Browser validation found mobile attention overflow. A bounded horizontally scrollable table fixes maximum-context rendering without representing masked cells as zero.

Handoff

No numerical, Learn, isolation, or container blocker remains. The upstream license status is unresolved, so upstream text is not distributed; its exact algorithm was validated through the pinned original and independently authored oracle without making a license decision. Stochastic sampling is outside this slice; deterministic teacher forcing has no active RNG state. The fixture tests one layer and two heads, not arbitrary production-scale model configurations.

After PR approvals and merges, a separate pass should reconcile contributor changes and integrate the catalog/launcher, then address existing-workshop hardening described in the execution plan. Keep this pre-merge deliverable isolated until then.