Skip to content

fix(proof-synthesis): correct two miscited arXiv claims - #130

Merged
avrabe merged 1 commit into
mainfrom
fix/proof-synthesis-citations
Jul 15, 2026
Merged

fix(proof-synthesis): correct two miscited arXiv claims#130
avrabe merged 1 commit into
mainfrom
fix/proof-synthesis-citations

Conversation

@avrabe

@avrabe avrabe commented Jul 15, 2026

Copy link
Copy Markdown
Contributor

A claim-verification sweep of every external citation in the proof-synthesis skill (each arXiv page fetched and checked against the claim it backs).

Verified correct (no change): 2605.26457 (Verus-SpecGym '26% miss' — real, states the figure verbatim; a prior no-web honesty pass had wrongly flagged it), 2409.13082 (AutoVerus 90%/150), 2511.17330 (AutoRocq), miniF2F 100% (Seed-Prover).

Fixed:

  • 2511.11829 was the wrong paper for the Lean row (it's about NL-requirement autoformalization, not Lean-4 spec-equivalence proving) → re-pointed to CLEVER (2505.13938), which genuinely proves Lean spec-equivalence.
  • Softened the CLEVER overstatement — 'no SOTA agent verifies more than 1 / spec-equivalence the dominant wall' isn't supported by the source; reworded to the abstract-supported 'struggle to achieve full verification; the spec is the first-class hard task.'

Version: skill 0.4.0 → 0.4.1; plugin 0.14.0 → 0.14.1.

🤖 Generated with Claude Code

… sweep)

Verified every external citation in the skill against the actual sources
(WebFetch of each arXiv page). Findings:
- arXiv 2605.26457 (Verus-SpecGym, '26% LLM-as-judge miss'): CORRECT — a prior
  honesty pass wrongly flagged it as future-dated/unconfirmed; the paper is real
  (May 2026, Parno/Welleck et al.) and states the figure verbatim. No change.
- arXiv 2409.13082 (AutoVerus, 90%/150 tasks): CORRECT.
- arXiv 2511.17330 (AutoRocq): CORRECT.
- miniF2F '100%': CORRECT (Seed-Prover 100% valid / 99.6% test, 2507.23726).
Two fixes:
- arXiv 2511.11829 was MISCITED for the Lean row — it is about NL-requirement
  autoformalization, not Lean-4 spec-equivalence proving. Re-pointed to CLEVER
  (2505.13938), which actually does prove Lean spec-equivalence.
- Softened the CLEVER claim: 'no SOTA agent verifies more than 1 / spec-equivalence
  the dominant wall' is not supported by the source's abstract; reworded to the
  supported 'struggle to achieve full verification; the spec is the first-class
  hard task.'

Version: skill 0.4.0 -> 0.4.1; plugin 0.14.0 -> 0.14.1.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
@avrabe
avrabe merged commit 2f0eb54 into main Jul 15, 2026
1 check passed
@avrabe
avrabe deleted the fix/proof-synthesis-citations branch July 15, 2026 20:54
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.

1 participant