Formal Conjectures: Review, Verification, and Preservation Loop

One `conjectures` toolkit for contribution review, exact-target proof verification and retained evidence. Contributors use their existing agents. FC owns deterministic operations; maintainers retain mathematical and merge decisions.

## Delivery map — 11 September 2026

**9 active FC infrastructure PRs: 3 foundations, 4 consolidated deliveries and 2 independent CI improvements.** Mathematical corrections and superseded prototypes are separate. No additional umbrella issue or replacement PR is needed.

| PR | Delivery | Required predecessors | Current disposition |
| --- | --- | --- | --- |
| #4899 | Review foundation and contracts | None | Open for review |
| #5375 | Shared metadata and one native `conjectures.json` | None | Open; final review follow-up CI passed, upstream deployment awaits acceptance |
| #5337 | Native proof-workspace exporter | Generator #7 | Draft; fork qualification passed, accepted repin pending |
| #5386 | CLI **and Actions preparation** | #4899, #5375 | Active draft; absorbs #5356 |
| #5387 | Exact proof workspaces and verification | #5386, #5337, Comparator #87 | Active draft |
| #5388 | Evidence publication **and problem-page display** | #5386 | Active draft; absorbs #5389 |
| #4828 | Status **and proof-link metadata consumers** | #5375 | Active draft; absorbs #4749 |
| #5468 | Targeted problem-only PR validation | Maintainer acceptance of PR/queue coverage split | Draft; hosted targeted path qualified |
| #5461 | Reuse queue-built websites, retain Lean/cache refresh | #5435 | Draft; real queue-to-main qualification remains |

```mermaid
flowchart LR
R["#4899 Review"] --> C["#5386 CLI + Actions"]
M["#5375 Metadata"] --> C
M --> S["#4828 Status + links"]
C --> E["#5388 Publish + display evidence"]
C --> P["#5387 Proof verification"]
G["Generator #7"] --> X["#5337 Exporter"] --> P
V["Comparator #87"] --> P
```

Review evidence can ship without proof verification. #5387 supplies proof records to #5388 when qualified. CI changes are a separate lane; their overlapping workflow changes must be reconciled as they land. #5435 and #5460 are merged. The cache prerequisite for #5461 is satisfied; its real queue-to-main acceptance remains outstanding.

## PR and branch policy

- **Open draft:** active implementation with dependencies or qualification remaining.
- **Open and ready for review:** focused diff, prerequisites available and relevant checks passed.
- **Closed:** merged, superseded, consolidated or abandoned. Closure is not a waiting room.

#5356 → #5386, #5389 → #5388 and #4749 → #4828 are now consolidated. Keep the three absorbed PRs closed and continue on the retained PRs. Their old branches/commits remain history; do not develop separate deliveries there.

Cross-fork drafts still target upstream `main`, so inherited prerequisite changes can appear in their diffs. PR bodies identify dependencies and available parent comparisons. After predecessors merge, replay only the delivery's own changes onto current upstream `main`, verify the commit list and diff, rerun affected checks, then request review on the same PR. Draft status does not waive those requirements.

Use `codex/fc-toolkit-integration` for combined testing. Focused fork branches are `codex/fc-toolkit-cli`, `codex/fc-toolkit-proof-cli`, `codex/fc-toolkit-evidence-cli` and `erdos-status-from-json`. No upstream merge or automation activation without maintainer acceptance.

## Acceptance ownership

| Tracker | Owns | Close only when |
| --- | --- | --- |
| #5152 / #5375 | Native metadata and catalog deployment | Shared-consumer semantics and publication are accepted |
| #5376 | Review, proof, publication and package acceptance | Both contribution journeys pass on the supported release revision |
| #5377 | Configured feeds, site applicability and reader handoff | A reader can identify the exact statement, evidence and next action |

#5388 implements publication and display together; the two acceptance issues are different user journeys, not requests for more PRs. Keep detailed qualification checklists in these trackers, not duplicated roadmap status appendices.

## Qualification — 11 September 2026

[RC3 qualification record](https://github.com/williamjblair/formal-conjectures/blob/e4dd405fc6803eb744922b3e527917e2cbd619d1/toolkit/qualification/rc3-final.md) retains exact revisions, runs and failures.

- Review preparation/completion, independent build, immutable publication, designated posting, cancellation/retry and historical inspection passed on the fork. No paid model calls.
- Fresh macOS initialization and Linux verification passed. Real controls distinguish accepted proofs, rejected submissions and unevaluated infrastructure failures. All ten exporter fixtures and 100/100 FC100 exports passed.
- William approved the five-case pilot packet qualitatively. Two original coverage gaps remain; no per-case scores or broad accuracy claim.
- Open Formal Workflows #2 is merged; its current deployment uses the shared reader. No replacement migration PR is needed.
- [RC3 is published](https://github.com/williamjblair/formal-conjectures/releases/tag/toolkit-v0.2.0rc3): clean macOS/Linux package and live fork browsing checks passed; published download commands, checksums, upgrade and uninstall passed. Final fork deployment, fresh desktop/mobile browser checks, CLI browsing and website-only snapshot reuse passed at `eee8f2ec`. The catalog and stylesheet match the qualified files byte-for-byte. Production default browsing and reader handoff still require upstream #5375/#5388 acceptance and deployment.

## Release gates

1. **Review:** clean install → exact PR/local snapshot → independent build and sources → existing-agent report → retained findings and coverage. William approved the retained five-case pilot packet qualitatively on 11 September; no per-case scores were supplied. Preserve its two coverage gaps and do not claim broad accuracy.
2. **Publication:** inspect → immutable archive → designated serialized publisher → receipt. Exercise competing/stale/cancelled requests and failures. A later head/base makes prior evidence historical.
3. **Proof:** macOS initialization → explicitly public committed workspace when using GitHub → qualified Linux verification → distinct success, rejection and infrastructure error. Retain exact tool/executor pins and sandbox controls.
4. **Reader handoff:** deployed catalog and configured feeds show statements, sources, evidence producer, applicability, maintainer status and next action.
5. **Release:** matching wheel/sdist/checksums, clean macOS/Linux installation, supported configurations and documentation all match the qualified commit. Upstream acceptance remains separate from fork delivery.

## Preserved decisions and history

The older #4884 account interpreted terminal axiom text as rejection. The [retained typed record](https://github.com/williamjblair/open-formal-workflows/blob/main/pilot/comparator-outcome.json) instead records an invocation error and unevaluated policy. Logs are diagnostics, not verification verdicts. Historical reports and failed model invocations remain preserved.

Mathematical corrections and proof-locator changes remain independent. #4951 is superseded by #5337; close it after the replacement lands. #5158 covers metadata/evidence design; #4930 covers downstream LeanEval intake and package policy; export success is not catalog admission. #4881's status interpretation belongs to #4828 and exact verification/evidence to #5387/#5388. #4876 covers guidance, #4825 queue maintenance and #4819 anonymous-declaration policy. Do not manufacture identities or reopen completed historical issues.

Runs are local by default; publication is explicit. No paid model calls are part of this qualification. Agents own authentication and model choice. Lean/Lake own dependency pins; the generator, Comparator, LeanEval and Palomar retain their roles. Autonomous proving/fixing, general MCP, daemons, new registries/services and Modal/Vela adapters remain deferred. Harbor export is an experimental follow-up on `codex/fc-toolkit-evals`, requiring a real qualified trial; it does not block the core CLI.

There is no single file or test for this umbrella issue; start by reading toolkit/qualification/rc3-final.md and the retained PR bodies (#4899, #5375, #5386, #5387, #5388, #4828, #5468 and #5461), then use codex/fc-toolkit-integration for combined testing. Completion depends on predecessor replay, affected checks, maintainer acceptance and the release gates passing.

Written by the indexing model from the issue text.