Sparsh/final submission pass - #146
Merged
Merged
Conversation
AAAI-27 main track allows 7 content pages plus up to 2 pages containing ONLY
references. Appendices bound into the main PDF count against the 7 content pages;
supplementary material must be uploaded as its own document. Our build was 11 pages:
1-7 main, 8 references, 9-11 appendices A-E, so the appendices were both over the
limit and in a position reserved for references.
* paper/tex/supplement.tex is new: a standalone document carrying appendices A-E,
numbered S1..S5 / Figure S1 so the main text can name them without a
cross-document reference
* main.tex no longer \input's the appendix; its nine Appendix/Table/Figure refs into
it are now textual ("supplementary Table S1", "Appendix C of the supplement")
* references that pointed the other way (appendix -> main paper floats, and the
best-of-K equation) are likewise textual, so both documents build standalone with
zero unresolved references
Also records the marc_aaai.tex -> main.tex rename done for Overleaf, and repoints
verify_paper_numbers.py, its messages, and paper/ABSTRACT.md at the new name; the
verifier was silently broken by the rename and now reads main.tex + the appendix.
Local proxy build (aaai2026 style, since aaai2027.sty needs pdfTeX): main 9 pages,
supplement 4, no unresolved refs, no overfull boxes. 52/52 number checks, 431 tests.
supplement.tex had the appendix body inlined, so the repo carried two copies of it. That is the same drift that put 0.717 in Table 4 while the artifact said 0.683, so supplement.tex now \input's marc_aaai_appendix.tex and there is one source of truth. That change surfaced a bug: the main-paper cross-references (Figure 1/2, Table 3) had been rewritten only in the inlined copy, never in marc_aaai_appendix.tex. With the \input in place they resurfaced as five unresolved references in the supplement. Fixed in the appendix file itself; every \ref left in it is now local. scripts/make_overleaf.sh assembles paper/tex/overleaf/ with exactly what the Overleaf project needs and nothing else: main.tex, supplement.tex, marc_aaai_appendix.tex, aaai2027.sty, aaai2027.bst, refs.bib, and only the three figures the two documents actually include (fig_repair_accuracy and fig_repair_kscaling are unused). The output is gitignored so a second copy of main.tex never lands in the repo. A README.txt in the bundle records the pdfLaTeX requirement and the pre-submission checks. Verified by building both documents from the bundle alone: main 9 pages, supplement 4, zero unresolved references in either, no appendix content in main.pdf. 52/52 checks.
Your build put the Conclusion tail and the whole "Use of AI Assistants" section on page 8
above the References heading. AAAI-27 allows 2 extra pages containing ONLY references, so
that spill is a violation on its own even though the appendix is now a separate document.
415 characters had to come out of the body.
* MATH-500: the admission stays in Limitations ("covers none of them"), the percentage
breakdown moves to a new supplement section. Two reviewers hit that passage, one as an
unmotivated tangent and one as a weapon, so it reads better scoped there anyway
* Limitations restated the Jensen-gap caveat that the factorization-law section already
makes -- and that section pointed *to* Limitations for it. Circular; kept one copy
* Limitations also restated the coupled-systems negative that Table 1 and its paragraph
already report
* conclusion, CircleLine sentence, and the distance-geometry paragraph tightened
supplement.tex gains \bibliography{refs} since the moved MATH-500 text cites Hendrycks.
Measured against your build rather than mine: the local proxy (aaai2026 style, since
aaai2027.sty needs pdfTeX) overstates the spill by ~2.4k characters, but the delta tracks,
and the proxy spill fell 2845 -> 2440, i.e. ~405 of the 415 characters needed. Rebuild and
confirm page 8 opens on "References".
…cle control
Presentation-only pass. No number changed, no limitation removed, no claim
strengthened; verified mechanically (all decimal literals identical before/after).
P1, volunteered novelty disclaimers. Four Related Work paragraphs now open with what
this work adds and carry the concession as a subordinate clause. Every citation,
attribution and scope limit is preserved:
* multistart: leads with the diagnostic, keeps "textbook" and "confirmation ... rather
than a finding"
* weak baselines: leads with the exact-bound point, keeps "neither it nor the
budget-matching remedy is ours"
* distance geometry: leads with the measurement, keeps "the structure is not ours"
* learned initialization: leads with what the negative establishes
"we claim no novelty for it" becomes the adjective "standard" in the Method's
architecture sentence; "MARC is the instrument of this study, not its contribution"
becomes "MARC is the instrument we measure with" in the framing.
P2, capacity-vs-regime confound. The oracle-marginal control was buried mid-paragraph in
the factorization-law section, which is not where a reviewer forms the objection. Moved
to the coupled section directly under the 0/5 result, with an explicit statement of what
the negative does and does not bound: it bounds product-of-marginals proposals, the class
the denoiser belongs to; an instance-conditional joint sampler is outside what the
control tests. The law section keeps a one-clause pointer.
P3, perceived scale. The 600-instance reachability sample now reads as the largest in the
study and separately drawn, rather than as an aside; the MATH-500 measurement is framed as
one the authors ran on themselves. Both were already true.
P4. The K=16 negatives were already in the main text but buried at the end of the cost
run-on; now a flagged sentence of their own.
P5. Introduction restructured so the positive arrives before the self-correction, with
four scannable contributions; each experimental subsection opens with its result.
Abstract untouched -- see the note in the PR. Net page impact: the pass added ~750
characters and gave back ~560, leaving roughly four lines to find; measured on the local
2026-style proxy, which overstates absolute spill but tracks deltas.
… filename #145 added a scrub for the staged camera-ready author block, guarded on `[ -f paper/tex/marc_aaai.tex ]`. This branch renames that file to main.tex and adds supplement.tex, so after both land the guard is false, the `&&` short-circuits, and the script exits 0 having scrubbed nothing -- the leak #145 fixed comes back silently. Now loops over main.tex, supplement.tex and the old marc_aaai.tex (so it works whichever order these merge), and fails loudly if no paper source is present at all rather than packaging an unscrubbed tree. Verified the sed range still matches the block in main.tex: 0 author-block lines survive. 52/52 number checks, 431 tests.
… verbatim) Task 3a: 'on coupled systems it never significantly beats random restart at any dimension tested' -- a reviewer who derives this from Table 1 unaided writes a harsher review than one handed it. Task 3b: the MATH-500 breakdown, 0 of 48 with the 20/31/48 split, restored from v3. 'covers none of them (breakdown in the supplement)' read evasive where the numbers read rigorous. Task 4: 'a bad training family actively hurts' restored as the characterization of the 0.70 -> 0.00 collapse. All three verbatim from the pre-page-fit source, +173 characters. No number altered; 52/52 checks. Paper is now over the 7-page limit -- reported, not silently offset.
…limit met
Abstract: pure reordering of the registered sentences so a Phase 1 reviewer meets
0.997 vs 0.236 in the first two sentences instead of the curtailment. Every registered
number retained and the curtailment retained verbatim ("this control sharply curtails our
headline claim"), now positioned as the method that produced the characterization rather
than the news. 254 words against the registered 253; no sentence rewritten.
Task 2, at a tenth of the cost. The three tie rates were visible only in supplementary
Table S1, and AAAI tells reviewers they need not read supplementary material. A float
would have cost 10-12 lines we do not have, so the rates are now inline in the body:
ties at 0.550, 0.683 and 0.683, each matching the random arm exactly, and 0.000 against
0.200 on CircleLine. Numbers copied from results/p_hard/hard_eval.json, not retyped.
Offsets, all removing duplication rather than content:
* weak-baselines paragraph keeps all three citations and the concession, compressed
* back half of the algorithm-selection paragraph
* "Langevin solves one, random solves four" -- Table 2's checkmarks already show it
* the finisher-not-the-proposal line appeared in prose and again in Table 2's caption
* fig:law caption restated slopes the body gives in full
Net -494 characters against the restores' +173. Proxy spill 2789 -> 2332, below the
2430 threshold calibrated against the last real build, so main content should now fit
7 pages. 52/52 checks, 431 tests, zero overfull boxes, zero unresolved references.
…ailment
The reorder in the previous commit moved the old sentence 20 to position 1 without its
connective tissue. Three real regressions, all fixed:
* the opening was a comparative claim about a referent never introduced -- no problem
class, no continuous/discrete split. Now three sentences establish the problem, the
gap, and the positive before any number appears
* "balanced nonlinear menu accuracy" landed before "menu" existed as a concept.
"choosing among K candidate augmentations" now gives it a referent one clause earlier
* worst of the three: "this control sharply curtails our headline claim" lost its
antecedent, and the nearest one in the new order was the ranker result -- so the
abstract read as though 0.997 had been walked back, the exact opposite of the edit's
purpose. Now "curtails what the learned proposal contributes on the value decision",
which names the claim and is the same disclosure
Also restored two glosses stripped for space: "candidate conditioning (the ranker encodes
each candidate-augmented graph)" in Related Work, where the term's first appearance was
bare although the whole positive rests on it, and the slopes in the Figure 1 caption,
which reviewers read standalone.
Paid for with duplication only -- the abstract's closing clause restated its new opening,
plus six body compressions. Every registered number retained; the curtailment retained.
52/52 checks, 431 tests, no overfull boxes, no unresolved references.
1. 'ties or loses at every dimension' was false: Table 1 row n=8 has learned 0.483 against random 0.467, so the learned arm is numerically higher there. The table's own verdict column already said 'no' correctly; only the prose overstated. Now 'records no significant win', which is what the p-values support (0/5 significant). 2. '3.7 to 600' are the MEASURED 1/q(n). A reader extrapolating v^-n at n=6 from v=0.27 gets ~2580, because the fitted slope (-1.03) is not exactly v^n with v=q(1). Now says 'from the measured q(n) rather than the v^n extrapolation'. 3. 0.933 (N=60 matched end-to-end) and 0.939 (N=360 probe comparison) sat two paragraphs apart with no signal they are different experiments; now labelled. No number altered. +146 chars, offset by -75 elsewhere. 52/52 checks.
31 questions answered against what the repo actually does, not aspirationally.
28 yes, two partial, one NA:
Q7 proofs of all novel claims -- PARTIAL. The factorization law is an identity
plus a measured validation (MAE 0.012 against a self-measured curve), not a
theorem carrying a proof obligation. 'Exactly one solvable option' is a CAS
certificate machine-checked per instance rather than proved on paper.
Q20 number and range of hyperparameter values tried -- PARTIAL. PROVENANCE records
the final command, seed and commit for every cited number, but not the full
sweep explored during development.
Q18 non-public datasets -- NA. Every family is procedurally generated from a seed;
MATH-500 is cited and public.
Everything answered yes is backed: MIT licence, requirements-lock.txt pins, seeds
recorded per result, Wilson intervals and z/McNemar tests throughout, 431-test CI,
and 52 automated checks tying each cited number to a committed JSON.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
No description provided.