Review¶

Verification notes: I reviewed the staged diff, rechecked the original NSF-ACCESS proposal task/budget language in examples/proposals/NSF-ACCESS/<previous-proposal>.zip (main.tex lines 86, 105-112, 131-134, 179-188), and compiled progress_report.tex locally. The LaTeX source compiles cleanly on the second pass to a 2-page letter PDF with no overfull/underfull or undefined-reference warnings after rerun.

File/diff scope: staged changes from git diff --cached for proposals/registry.yaml, proposals/<project>/README.md, proposals/<project>/reason.txt, proposals/<project>/progress_report.tex, proposals/<project>/progress_report.pdf, and template/context/team/<pi>/nsf-access-acknowledgements.md.

Review lens: proposal/nsf plus common proposal dimensions, with extra focus on ACCESS Progress Report / Supplement expectations, remaining-month feasibility, Table 2 to narrative consistency, skim clarity, and formatting compliance.

New¶

None.

Previously raised¶

Fixed¶

T3 mislabeled as the largest compute line. This is fixed. proposals/<project>/progress_report.tex:122-127, proposals/<project>/progress_report.tex:143-145, and proposals/<project>/reason.txt:16-18 now describe T3 as the largest unfinished workstream rather than the largest original budget line, and Table 2 has been rebalanced to keep T3 at $X00k while shifting T2 to $Y00k at proposals/<project>/progress_report.tex:158-178.
Burn-rate mismatch unexplained. This is fixed. proposals/<project>/progress_report.tex:75-85 now gives two concrete causes, expanded open-weight model coverage and less favorable Delta exchange rates, and the short form carries the same explanation at proposals/<project>/reason.txt:4-10. For an ACCESS reviewer, this now reads as a real utilization explanation rather than an empty claim of "used everything."
Tracker acknowledgement text baked in both allocations verbatim. This is fixed. template/context/team/<pi>/nsf-access-acknowledgements.md:21-54 is now a real template with an explicit "insert only the allocations actually used" instruction, plus one-allocation and two-allocation examples. That rewrite materially lowers the chance of future over-claiming.

Still open¶

The publication-evidence issue is improved but not fully closed, because the README still overstates the AAAI paper's acknowledgement status and now contradicts the tracker's own stricter rule. proposals/<project>/README.md:23 tells the user to add three papers that "acknowledge compute," including the AAAI paper. But the tracker says the AAAI paper published with stale <ALLOC_B> boilerplate at template/context/team/<pi>/nsf-access-acknowledgements.md:82-91, and its process rule says ACCESS reports should list only papers that cite the target allocation in the camera-ready at template/context/team/<pi>/nsf-access-acknowledgements.md:129-131. This is now the only material inconsistency left in the submission instructions.

Recommended change: revise the README publications step so it matches the tracker exactly. Two clean options: - If the ACCESS form should list only papers whose published camera-ready cites <ALLOC_A>, tell the user to list only the two ACL 2026 papers and keep the AAAI paper only in the attached progress report with the stale-boilerplate caveat. - If the ACCESS form allows papers that used <ALLOC_A> compute even when the published acknowledgement is stale, say that explicitly: two papers cite <ALLOC_A> directly, while the AAAI paper used <ALLOC_A> compute but published with stale <ALLOC_B> boilerplate.

Deferred¶

None.

Reopened¶

None.

Additional notes¶

The burn-rate explanation is now concrete and credible. Section 5 and Table 2 are internally consistent after the T2/T3 rebalance. The reason.txt caveat reads honest rather than hedged.

There is now page headroom if you want to add one more sentence, but I would spend it only on clarifying the README/publications rule above, not on expanding the narrative elsewhere.