Skip to content

docs(proofs): per-statement placement of the engineering proofs (split by property) - #518

Open
hyperpolymath wants to merge 2 commits into
claude/new-session-th4e9w-status-placementfrom
claude/new-session-th4e9w-placement-ledger
Open

hyperpolymath wants to merge 2 commits into
claude/new-session-th4e9w-status-placementfrom
claude/new-session-th4e9w-placement-ledger

Conversation

@hyperpolymath

@hyperpolymath hyperpolymath commented Nov 16, 2026 •

Copy link
Copy Markdown
Owner

Stacked on #514, which records the ruling. This PR adds the per-statement ledger.

Summary

Owner ruling 2026-11-16: split by property. docs/PROOF-PLACEMENT.adoc gives every top-level statement in R9, R10, R4's engineering modules, the R3 Lean corpus and the Idris 2 verification corpus exactly one home. A row is one statement or a named group of closely related declarations. A file covered by two areas appears in only one, so no statement has two homes.

Home Rows
rust (Creusot/Pearlite and Verus; target function and contract named) 92
idris2 (interface or wire fact) 4
agda-theory 31
retire (successor named) 106
Total 233

Each area also lists model gaps (places where a proof models behaviour the Rust lacks) and open questions (points only the owner can settle).

How it was produced

Three passes per area, all reading the source:

  1. a classifier;
  2. an adversarial reviewer told to refute each verdict;
  3. a judge that reconciled the two against the source.

No canonical check ran and no receipt was produced. One judge rechecked a few Agda probes on a scratch Agda 2.6.3. The document is rendered mechanically from the judges' JSON with jq. It is a placement, not a receipt, and no PROOF-STATUS token changes.

Source readings that look like live holes

These are not yet confirmed by a test or repro, and none is fixed here.

  • Integrity checking never runs in the shipped binary.
    • with_integrity_checker (src/rust/dispatch.rs:317) has no caller.
    • All 17 entries of config/solver-manifest.toml hold PLACEHOLDER_COMPUTE_ON_FIRST_RUN.
    • So solver_integrity_ok is always true.
    • src/rust/integrity/mod.rs:5-8 claims a mismatched solver "will not be used".
  • The cross-checked dispatch loses the axiom cap. verify_proof keeps only the first axiom usage (dispatch.rs:643-645), and the cross-checked path recomputes the danger from it (:851-855, :890).
  • A disagreement can still verify. Under the Bayesian and DempsterShafer policies, a conclusive disagreement can still yield a verdict (result_arbiter.rs:374-431). Dispatch then returns verified = true (dispatch.rs:899-900).
  • The integrity cache is refreshed on Tampered and on placeholders (solver_integrity.rs:250-253, :272-275).
  • The manifest key mapping breaks for some provers. The ProverKind-to-manifest key mapping fails for lean4 and hol_light, and collides MizAR with Mizar.
  • Unused functions carry proofs. clamp_trust, max_level, combine_trust, min_level, is_solver_safe and quick_reverify have no caller outside tests.

🤖 Generated with Claude Code

https://claude.ai/code/session_019a95kk88TZ1iegsBK3kox3

@hyperpolymath
hyperpolymath marked this pull request as ready for review November 16, 2026 23:37
@chatgpt-codex-connector

chatgpt-codex-connector Bot commented Nov 16, 2026 •

Copy link
Copy Markdown
Codex Review Summary: ❌ Failed

This comment shows the latest Codex review activity on this pull request.

Review Status Commit Review trigger
📝 Code Review ✅ Completed 2026-11-16T23:51:04.481608Z 9a22dd8 New commits
🔒 Security Review ❌ Failed 2026-11-16T23:47:52.230349Z
We couldn’t complete this review. Comment @codex security review to try again.
9a22dd8 New commits
ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

  • Open a pull request for review
  • Mark a draft as ready
  • Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

@chatgpt-codex-connector chatgpt-codex-connector Bot left a comment

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 99941b9162

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

  • Open a pull request for review
  • Mark a draft as ready
  • Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

Comment thread docs/PROOF-PLACEMENT.adoc Resolved
Comment thread docs/PROOF-PLACEMENT.adoc Resolved
Comment thread docs/PROOF-PLACEMENT.adoc Resolved
hyperpolymath pushed a commit that referenced this pull request Nov 16, 2026
Codex on #518, three findings, all confirmed:

- The census double-counted statements and gave some two homes. Four
  files were covered by two or three areas (ClampTrustBounds.idr and
  NatLte.idr by R10 and the Idris corpus; DispatchOrdering.idr by R9
  and the Idris corpus; the key-* aliases in MetaChecker.agda by R9, R10
  and R4). The 15 order lemmas were retire in one area and rust in the
  other, with the same successor contract. Each file now sits in one
  area (the finer-grained rows kept; the 23 aliases stay one retire
  row, since an alias proves nothing beyond its target); the order
  lemmas are retire, as the corpus row's own reason ("closed facts
  about literals") says. Census: 233 rows (92 rust, 4 idris2, 31
  agda-theory, 106 retire), defined as one statement or a named group.
- The method note said nothing was run, but one judge rechecked Agda
  probes on a scratch Agda 2.6.3. It now says no canonical check ran,
  names that recheck, and says it is not a receipt.
- Two model gaps counted a verify_all error as an integrity fail-open
  path; verify_all only returns Ok (solver_integrity.rs:164-209), as
  the R9 analysis says. Both now say the case is unreachable.

The ledger is re-rendered from the corrected JSON by the same jq
generator; PROOF-STATUS quotes the new counts.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019a95kk88TZ1iegsBK3kox3
@hyperpolymath
hyperpolymath added this pull request to stack #522 November 18, 2026 18:43
Repository owner deleted a comment from chatgpt-codex-connector Bot Nov 19, 2026
Repository owner deleted a comment from chatgpt-codex-connector Bot Nov 19, 2026
Repository owner deleted a comment from coderabbitai Bot Nov 19, 2026
Repository owner deleted a comment from chatgpt-codex-connector Bot Nov 19, 2026
…t by property)

Owner ruling 2026-11-16: each existing engineering proof statement goes
where its property lives. docs/PROOF-PLACEMENT.adoc records the home of
all 247 top-level statements in R9, R10, R4's engineering modules, the
R3 Lean corpus and the Idris 2 verification corpus:

  rust (Creusot/Pearlite and Verus)  96
  idris2 (interface / wire)           4
  agda-theory                        31
  retire (with successor named)     116

Each rust row names its target function and the contract to discharge.
Per area it lists the places where a proof models behaviour the Rust
lacks, and the questions only the owner can settle.

Produced in three passes, all read-only: a classifier per area, an
adversarial reviewer per area, and a judge that reconciled the two
against the source. Nothing was built or run (no Agda, Lean, Idris 2,
Creusot, Verus or cargo), so this is a placement, not a receipt; it
changes no PROOF-STATUS token. The defects it names are source
readings that still need a failing test or a repro.

The ledger was rendered mechanically from the reconciled JSON (jq), not
hand-transcribed. PROOF-STATUS links it from the 2026-11-16 ruling
paragraph.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019a95kk88TZ1iegsBK3kox3
Codex on #518, three findings, all confirmed:

- The census double-counted statements and gave some two homes. Four
  files were covered by two or three areas (ClampTrustBounds.idr and
  NatLte.idr by R10 and the Idris corpus; DispatchOrdering.idr by R9
  and the Idris corpus; the key-* aliases in MetaChecker.agda by R9, R10
  and R4). The 15 order lemmas were retire in one area and rust in the
  other, with the same successor contract. Each file now sits in one
  area (the finer-grained rows kept; the 23 aliases stay one retire
  row, since an alias proves nothing beyond its target); the order
  lemmas are retire, as the corpus row's own reason ("closed facts
  about literals") says. Census: 233 rows (92 rust, 4 idris2, 31
  agda-theory, 106 retire), defined as one statement or a named group.
- The method note said nothing was run, but one judge rechecked Agda
  probes on a scratch Agda 2.6.3. It now says no canonical check ran,
  names that recheck, and says it is not a receipt.
- Two model gaps counted a verify_all error as an integrity fail-open
  path; verify_all only returns Ok (solver_integrity.rs:164-209), as
  the R9 analysis says. Both now say the case is unreachable.

The ledger is re-rendered from the corrected JSON by the same jq
generator; PROOF-STATUS quotes the new counts.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019a95kk88TZ1iegsBK3kox3
@hyperpolymath
hyperpolymath force-pushed the claude/new-session-th4e9w-placement-ledger branch from 9a22dd8 to 3334c26 Compare November 19, 2026 15:32
@sonarqubecloud

Copy link
Copy Markdown

Repository owner deleted a comment from coderabbitai Bot Nov 19, 2026

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants