paper: final-submission pass — Table 4 corrected against the artifact, missing ablation landed, weak-baseline literature cited

[ImSpxrsh]

The blocking fix

Appendix Table 4 disagreed with the results file it reports. The paper had BilinearProduct at 0.717 / "random wins"; results/p_hard/hard_eval.json has 41/60 = 0.683, identical to the learned arm — a tie. The CI in the paper ([0.59, 0.81]) is the Wilson interval for 43/60, so both the rate and its interval came from a run we do not ship. PROVENANCE R11c carried the same stale value, so the drift predates this pass, and verify_paper_numbers.py never covered that table.

• Table 4 is now generated directly from the JSON
• main text: "no win on any of the four families: three ties and one complete failure on CircleLine" (was "ties on two and loses on two")
• added the LM arm the table omitted — 60/60 on all four families — with a caption line explaining why it belongs: the strong classical solver saturates these families, which is precisely why learned-vs-random is the informative contrast. Omitting it was a cherry-picking exposure.
• +7 verifier rows over the hybrid battery (52 checks total)
• R11c preserved as history, annotated rather than rewritten

The missing artifact

The operator-mask ablation was cited in the paper and recorded in RESULTS.md, but both JSONs had never been committed — they were missing from the .gitignore whitelist, so nothing could check them. Regenerated and committed:

• nonlinear reproduces: 0.978 [0.957, 0.989] vs 0.997 unmasked. It is not bit-reproducible at fixed seed (0.978 / 0.981 / 0.986 across identical reruns), so the paper cites the interval and says so.
• linear does not reproduce: 0.229 [0.206, 0.254] against a random arm of 0.249, where RESULTS.md recorded 0.379 [0.352, 0.407] — non-overlapping intervals, masked arm below chance. Consistent with the documented linear seed instability (R22: one seed at 0.227 below random's 0.248), not a new fault. The paper now cites the nonlinear ablation only; PROVENANCE says not to restore 0.379.

Shipping the original sentence would have meant a reviewer running the artifact finding a cited number that inverts.

Positioning

Added the weak-baseline critique literature this result sits inside: Nath & Kuhnle's max-cut benchmark and learned-local-search study, and Rodrigues et al.'s budget-matched LLM HPO study. We claim neither the observation nor the remedy as new. What is new is that one polish and one checker make best-of-$K$ the exact bound a learned proposal must clear, and the measured $\log q(n)$ slope predicts which families leave room above it.

Citations

Three entries had no author field and rendered as "arXiv:2502.00277. 2025." All three are real papers; two had paraphrased titles. Corrected against arXiv (Feng & Yang; Ding et al.; Wang et al.).

Presentation

• title/abstract tracked to the OpenReview record — 92% verbatim, three forced changes (false range; operator-aware → candidate-conditioned, which the ablation contradicts; and an abstract claim of reach that the Conclusion disclaims)
• both main-text figures redrawn at the width of the slot that holds them; they were being shrunk to ~41% and their labels were unreadable
• --replot added to run_crossover_theory.py so the figure can be redrawn without re-measuring
• prose pass: em dashes 58 → 0, colon sentences 33% → 21%, "The" openers 30% → 21%
• "Use of AI Assistants" statement added after the bibliography

State

7 content pages, 52/52 number checks, 431 tests, no overfull boxes, no unresolved references.

Still open for the team

• aaai2026.sty vs the AAAI-2027 submission record — page limits move between years
• the title still foregrounds the diffusion proposer, the half the paper explicitly does not claim