traceability.md

Feature-to-evidence traceability

This matrix keeps the product, repository, video, and submission copy aligned. Every public claim maps to visible behavior, implementation, and automated proof.

Product matrix

Feature or claim	User-visible proof	Implementation	Automated proof
Synthetic flawed lesson and expert correction	Synthetic-data badge, correction, checked note, and false caption are visible	fixtures/inverse-sine-release.json, src/domain/fixture.ts, src/client/App.tsx	Engine, API, and UI suites
Correction binds to exact evidence	Source and target quotes plus the exact diff are visible; locations remain available under Technical details	src/domain/validation.ts, src/domain/engine.ts, src/client/snapshotAdapter.ts	Exact, ambiguous, stale, and unsafe evidence cases
GPT-5.6 proposes a bounded structured repair	Live mode shows the GPT-5.6 model badge and a structured proposal rather than chat prose	src/domain/planners.ts, src/domain/schema.ts	Seven planner-adapter tests plus the bounded provider smoke command
Built-in demo is honest and reproducible	Header reads Built-in demo · no AI call	FixtureRepairPlanner, src/client/App.tsx	UI and API mode tests
Unsafe plans fail closed	Error or blocked state appears; release hash and documents do not change	src/domain/validation.ts, src/domain/engine.ts, src/server/app.ts	Prompt injection, ambiguous evidence, unsafe path, malformed output, and provider-failure tests
Human approval is mandatory	Approve this exact change and Apply change & run 6 checks are separate controls	LessonProofEngine.approve, LessonProofEngine.apply, src/client/App.tsx	Engine, API, and UI state-transition tests
Six deterministic checks decide release	1 caption fixed. 2 dependency proofs recomputed. 6 of 6 checks passed. is visible after apply	LessonProofEngine.runChecks	Positive journey and negative safety tests
Verified release receives a proof hash	Plain outcome is visible; full new hash, previous hash, version, and verification time remain available under Technical details	src/domain/hash.ts, src/domain/engine.ts, src/client/App.tsx	Engine, API, and UI proof tests
Undo is hash-guarded	Undo remains a separate control and the proof-hash receipt is available under Technical details	LessonProofEngine.undo	Exact restore, stale hash, and journal-integrity tests
Browser workflows are isolated	Each browser can reset and progress independently	src/server/session.ts, src/server/app.ts	Cookie, isolation, expiry, eviction, and concurrent-mutation tests
Live model traffic is bounded	Rate-limit errors are presented without changing release state	LiveAnalysisGuard, stable API errors	Per-session, global, and concurrency limit tests
No learner data is required	Entire golden journey uses bundled synthetic content	Fixture inventory, SECURITY.md, public content scan	Fixture and public-package preflight tests
Judge can reproduce the app	README and testing guide provide a no-secret route	package.json, Dockerfile, Render blueprint, same-origin server	Clean commands, production build, API health, and public preflight

Verification inventory

Test surface	Count	Primary behavior
Domain engine	15	lifecycle, evidence, stale state, fail-closed analysis, atomic apply, proof, undo
GPT-5.6 planner adapter	7	request contract, strict schema, family check, full-body timeout, missing credentials, refusal, malformed output
HTTP API and browser sessions	11	routes, errors, golden journey, cookies, isolation, expiry, rate limits, concurrency, same-origin mutations
Interface	4	provenance/mode display, separate controls, live-mode continuity, error recovery
Snapshot adapter	3	stage projection, verified proof, guarded undo
Design guards	2	Flat CSS, specific copy, and emoji-free interface source
Total	42	End-to-end product and trust boundary

npm run check also runs TypeScript validation, the production build, and the public-package preflight.

Submission evidence map

Judge question	Evidence to show
What work does GPT-5.6 perform?	The live structured proposal turns reviewer intent and bounded evidence into exact evidence, patch, invalidation, and check selections
What prevents autonomous or stale mutation?	Separate approval/apply controls, allowlisted paths, current release hashes, six checks, and guarded undo
Is this a coherent product rather than a technical demo?	One continuous correction → proposal → approval → verification → proof journey in the evidence cockpit
Who benefits?	Teachers and educational-content teams maintaining a lesson plus derived captions, claims, and release packages
What is novel?	The product starts after an expert correction and turns that signal into a source-bound release contract
How did Codex contribute?	Candidate and rule audit, clean-room product boundary, parallel implementation, 42-test verification, red-team review, deployment and submission hardening
What decisions remained human?	Track and product selection, synthetic scope, model authority boundary, approval requirement, deterministic release gate, and final public claims

Recording proof order

The clearest under-three-minute evidence sequence is:

1. show the false synthetic caption and expert correction;
2. keep the live GPT-5.6 badge visible while generating the structured plan;
3. show exact source/target evidence and one before/after change; open Technical details briefly for paths and dependency proof records;
4. show that release remains blocked before approval;
5. approve, apply, and show Change verified plus the plain 1 caption fixed. 2 dependency proofs recomputed. 6 of 6 checks passed. result;
6. show the guarded undo condition;
7. close on the architecture, 42-test count, concrete Codex contribution, and division of authority.