Evidence Lane — Governed State for Codex + ChatGPT

Immutable project evidence, lane-specific operators, and exact six-way human gates for Codex and ChatGPT.

Comment

Evidence Lane for Codex + ChatGPT

Live website · GitHub repository

The problem

A new AI task can inherit a polished summary and still resume from the wrong project state. The summary may omit the exact Git commit and tree, accepted version, source bytes, corrections, failed checks, pending candidate, or human decision that controls what happens next.

Evidence Lane treats this as an authority problem, not a prompt-length problem. It makes project state inspectable and hash-bound while keeping acceptance with the user.

What Evidence Lane does

Evidence Lane separates four states that agentic workflows often blur together:

immutable accepted project truth; the exact current source boundary; an unaccepted candidate with executable receipts; and the human decision about that candidate.

Every governed session records append-only Chat Lineage, binds work to one declared source boundary, rebuilds deterministic evidence, and stops at a six-way human gate. The model may inspect evidence, implement a bounded change, and propose a candidate. It cannot silently rewrite accepted history or approve its own output.

Six primary controls

Boot verifies runtime, locked ENV/UOP Flash, host class, and durable storage. Rollback moves only the accepted pointer across immutable accepted versions. Build produces a complete unaccepted candidate and stops at HIL. Refresh rebuilds changed sections and reuses only content-addressed unchanged evidence. Mode binds lane-specific governance, operators, schemas, and response law. Source Intake enrolls ordered source evidence without moving the accepted pointer.

State Travel is a conditional continuity event rather than a seventh everyday control. When the user requests it, it preserves the exact verified work boundary—including unfinished correction state, pending candidate when present, pointer base, Plan Lane, additive Deltas, resume row, and execution profile—and resumes that same row in a fresh task. Acceptance is not a prerequisite; an accepted-entry resume remains an explicit option.

Eighteen lanes, not one generic brain

Evidence Lane has 18 canonical lanes: Discussion, Analysis, Plan, Mode, Local Code, GitHub Code, Documents, Data/Excel, Presentations, PDF/OCR, Images/OCR, Artifacts, Custom, Brain Loader, Research, Project Engulf, Chat Lineage, and SQLite Brain.

Each lane has its own source-intake contract, parser/chunker settings, schema, retrieval surface, and exact proof package. The public Proof surface exposes each lane's downloadable SQLite, full generated MMD, reconciled DOT, refresh receipt, and derived 7680×4320 PNG plus lossless SVG. GitHub Code proves real commit-parent-file-change history; Local Code proves working-tree symbols, imports, routes, and dependencies without inventing Git history.

Codex and ChatGPT are separate host universes

Codex runs the full installed lifecycle over its governed local repository and accepted pointer. ChatGPT uses the same product code against its own mounted persistent PV/storage boundary under the same append-only lane, ENV/UOP, Exit-Slip, and HIL laws. Neither host's storage is treated as the other's authority.

The purchased website domain is a public product and evidence surface. Its Vercel MCP edge now verifies the exact deployed Git identity, but remains intentionally fail-closed until a Vercel-reachable durable HTTPS authority and authentication are configured. The persistent Windows OpenAI tunnel is a separate private stdio/control-plane transport and is not falsely presented as the public HTTPS origin.

Prompt Studio and public proof

The production website includes:

a 131-row additive Delta ledger in one collapsed-by-default table; a floating and full-page Prompt Studio backed by the same committed retrieval corpus; deterministic LlamaIndex-style chunks, SQLite FTS5, BM25, TF-IDF, reciprocal-rank fusion, JSON projection, and a hash manifest; strict refusal for project questions without supporting evidence; and an optional zero-cost general-question route with no paid fallback, disabled until separately configured.

The current committed corpus contains 243 public-safe source records and 1,773 chunks with corpus SHA-256 FC6DDEAC20443CFE0E612BE7DFCB8B82B93B18924B89316B35A6CA2F87F8D7F5. Its retrieval confidence canaries pass on the production domain.

Verified replacement candidate boundary

The exact production deployment was built from feature-branch commit:

commit: d919cbd0d73676c6e7c2a6b4189b614d1ec42144 tree: f1d80f829f0cffdfd962e637b5bc12a5e39d2e27 branch: agent/evi-v130-all-source-brain-workflow-hil-v1.3.0 Vercel production deployment: 7NTQQybuyJGZXdoiFhhx3MFhQWXr website: https://evidencelane.org

The complete Python suite passed 261 tests, TypeScript checks passed, the production Next.js build passed, all 17 public routes returned HTTP 200, and the production Prompt Studio routing checks passed. This is deployment evidence, not acceptance. Accepted PV5 and pointer generation 5 remain unchanged while replacement PV6 is resealed and presented at the next six-way HIL.

The six decisions remain:

APPROVE | APPROVE_WITH_DELTA | MORE_RESEARCH | ROLLBACK | REJECT | FAIL

Open-source provenance and rights boundary

Evidence Lane is independently authored. Pinned upstream sources are classified by exact role, adopted contract, refusal, maturity, license, and rights boundary. The public provenance includes GitHub Agentic Workflows, gh-aw-harness, the GitHub Agentic Workflows MCP gateway, Copilot SDK, GitHub Agentic Workflows threat detection, Open WebUI, Graphify, LlamaIndex, and OpenAI's Codex, MCP-server, ChatGPT UI, and app-review guidance.

Graphify informed stable-identity and explicit-edge graph concepts, but Evidence Lane does not adopt its Neo4j runtime, LLM extraction, server, installer, UI, or claim its version identity. GitHub Actions evidence is not represented as a GitHub Agent session. Third-party software, services, models, assets, and trademarks remain governed by their owners' licenses, terms, and rights.

Current limitations

The public MCP route remains explicitly blocked because no verified Vercel-reachable durable HTTPS origin is configured. Production release identity and deployed Git SHA match exactly at d919cbd0d73676c6e7c2a6b4189b614d1ec42144, so /healthz now reports only DURABLE_HTTPS_ORIGIN_REQUIRED; DNS and a website deployment still do not prove connector readiness. The optional external general-AI route is disabled and no paid fallback exists. AC12 remains post-seal and must bind the candidate engine, source, manifest, package, and Exit Slip to the same exact commit. Replacement PV6 remains unaccepted until the user makes the next six-way HIL decision. Evidence Lane does not claim production certification, universal provider parity, measured token savings, or measured performance gains.

What comes next

The immediate boundary is to seal replacement PV6 from the exact source above, execute AC12 post-seal, present the complete six-way HIL, and stop for the user's decision. Fuse, accepted-pointer movement, main-branch merge, and candidate installation remain outside this pre-HIL correction.

Built With

Share this project:

Updates