review(p28): adversarial review against the challenge's own stated rules

Scored against the challenge's operative standard (section 4: a CAS-based
symbolic derivation is sufficient evidence) and its stated central risk
(section 1: retrieval vs reasoning), not a generic rigour bar.

Verdict: satisfies and exceeds the evidentiary standard. Findings:

  F1 CRITICAL  run_checks.sh exits 1 on any machine WITH SageMath installed;
               set -euo pipefail makes the declared-optional Sage cross-checks
               fatal, so the script passes without the optional tooling and
               fails with it
  F2 MEDIUM    false assertion at p28_lattice_hypotheses_certificate.sage:107 --
               component 1 is 216+154x as written, true value 216+262x
               (b2=(1,2,1,0) contributes 216x, not 108x). Non-load-bearing:
               the identity below it, Bcomb == 4x*Crow, is TRUE and Crow
               reproduces the manuscript's C(x) exactly
  F3 MEDIUM    Sage 10.9 parent-coercion TypeError at line 43 masks F2
  F4 MEDIUM    novel-vs-imported content not stated plainly, though the
               Chudnovsky import itself is cited precisely
  F5 PASS      proves the officially stated claim about the official object
  F6 PASS      evidence exceeds the section-4 standard, and is falsifiable
  F7 PASS      manuscript builds clean: 0 errors, 0 warnings, 17 pages

None of F1-F3 touches the mathematics; all are certificate plumbing.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WY6SfRYvm8zFKMX9GcjS8u
This commit is contained in:
allaun 2026-07-31 04:16:04 -05:00
parent 8d25a7b367
commit 3702a638ad

View file

@ -0,0 +1,185 @@
# Adversarial review of the Problem 2.8 submission, against the challenge's own rules
Reviewer: independent replay, 2026-07-31. Target: commit `492c8ab`
(tag `p28-submission-2026-08-01`). Environment: CachyOS, python3 3.14,
SageMath 10.9, TeX Live 2026.
## The standard actually being applied
The challenge states no prize, eligibility, or submission rules. Its operative
evidentiary standard is in §4, *Discussion — Proof in the Age of AI*:
> For the purpose of this challenge, we consider a derivation carried out using
> symbolic libraries within established computer algebra systems as sufficient
> evidence of a valid solution.
and §1 names the central risk the evaluators care about:
> if a problem or its solution appears in the AI training data, success may
> reflect retrieval rather than reasoning.
This review is scored against those two clauses, not against a generic notion of
rigour.
## Verdict
**The submission satisfies the challenge's evidentiary standard, and exceeds
it.** One defect is nonetheless serious *as a submission*, because it fires on
the evaluator's first command.
## Findings
### F1 — CRITICAL (process, not mathematics): `run_checks.sh` exits 1
The submission's single advertised replay command fails on any machine that has
SageMath installed.
```
$ bash run_checks.sh
... 43 PASS lines ...
TypeError: keys do not match self's parent
$ echo $?
1
```
Two independent causes compound:
1. **`set -euo pipefail` (line 2) makes the "optional" cross-checks fatal.**
Lines 4449 run three Sage certificates inside `if command -v sage`. Where
Sage is absent the script prints *"OPTIONAL: SageMath is not installed"* and
exits 0. Where Sage is present, any Sage failure aborts everything. The
script therefore **passes on machines without the optional tooling and fails
on machines with it** — precisely inverted. `ADVERSARIAL_AUDIT.md` claims
"Wolfram and Sage are now optional independent cross-checks"; the exit code
does not honour that claim.
2. **`p28_lattice_hypotheses_certificate.sage` fails**, for two reasons of very
different severity — see F2 and F3.
This matters disproportionately because the challenge is explicitly CAS-oriented.
An evaluator applying the §4 standard is *more* likely than average to have Sage
installed, and is therefore *more* likely to see the failure.
The other two Sage certificates (`p28_kernel_contiguity_certificate.sage`,
`all_four_columns_certificate.sage`) both exit 0.
### F2 — MEDIUM (real bug, non-load-bearing): a false assertion
`p28_lattice_hypotheses_certificate.sage` line 107 asserts
```
72*b3 + 108*x*b2 + 46*x*b1 + 5*x*b0
== vector(K, [72+108x+46x+5x, 216+108x+46x, 216+108x, 72])
```
Component 1 is **wrong**. Since `b2 = (1,2,1,0)`, the `108*x*b2` term contributes
`216x`, not `108x`:
| component | true value | as written | equal |
|---|---|---|---|
| 0 | `72 + 159x` | `72 + 159x` | yes |
| 1 | `216 + 262x` | `216 + 154x` | **no** |
| 2 | `216 + 108x` | `216 + 108x` | yes |
| 3 | `72` | `72` | yes |
**The mathematics is unaffected.** This is a redundant intermediate display
check. The load-bearing identity immediately below it,
```
Bcomb == 4*x*Crow, Crow = 18*b3/x + (5/4)b0 + (23/2)b1 + 27*b2
```
is **true**, and `Crow` reproduces the manuscript's compact denominator row
`C(x) = (18/x + 159/4, 54/x + 131/2, 54/x + 27, 18/x)` exactly. Verified
independently in sympy. So the defect is a transcription slip in a check that
proves nothing the next line does not prove correctly.
### F3 — MEDIUM (environment): Sage 10.9 parent-coercion failure masks F2
Line 43, `denominator.subs({x: 0})`, raises
`TypeError: keys do not match self's parent`. `K` is a fraction field, so
`.numerator()` / `.denominator()` return elements of the underlying polynomial
ring while `x` belongs to `K`. Sage 10.9 no longer coerces the substitution key.
Coercing the key (`denominator.parent()(x)`) clears the TypeError — and the
certificate then runs on to fail at F2. The version-drift bug was **hiding a
real one**, which is the more instructive fact: the certificate has evidently
not been executed to completion on current Sage.
### F4 — MEDIUM (framing): the contamination question is not addressed
§1 names retrieval-vs-reasoning as the central evaluation risk. Problem 2.8's
target is `sqrt(10005)/pi` — the Chudnovsky constant — and the submission
**imports the Chudnovsky identity** as its one external theorem.
The import itself is handled well: §"The CM function and its exact value" names
it as a single external theorem, cites Milla Theorem 0.1 with the modular/CM
proof located at Theorem 9.7 and Chapter 10, and displays the elementary
coefficient bridge. That is ordinary mathematical practice and is not a defect.
But the submission nowhere states plainly **what is novel versus what is
imported**. Given that the evaluators flagged contamination explicitly, a
reviewer could mistake the boxed `Phi(x_0) = sqrt(10005)/pi` for the result
rather than for a cited input. One short paragraph would remove the ambiguity:
the recurrence-to-`Phi` reduction is the new content; the CM evaluation of
`Phi(x_0)` is classical and cited.
### F5 — PASS: the submission proves the officially stated claim
Independently verified against the challenge text by exact rational arithmetic
(`p28_official_object_certificate.py`, 18 assertions, three negative controls):
- `R = 151931373056001`, `u = 2n+3`, `w = u(3u-2)(3u+2)` — match
- all 11 polynomial entries `a1..a4, b1..b3, c1..c4` at `r = R` — symbolically identical
- full 4x4 `M_N(x_0)` vs official `M(N)` — all 16 entries identical at `N = 0,1,2,3,5,8,17,40`
- product convention `M(0)M(1)...M(N-1)`, `M_0 = I` — match
- both integer seed rows, generated by `A_0 = AC - (5/4)H_0` and `A_1 = SC` — exact match
- deformed coefficient `(14R-567)/9 = 236337691420383` — the official value
- orientation: the challenge asks `P_{N,j}/Q_{N,j} -> sqrt(10005)/pi`; the manuscript proves that orientation, for all four columns
The proof is about the official object, not a surrogate.
### F6 — PASS (exceeds standard): evidence is stronger than §4 requires
§4 accepts a CAS derivation. The submission instead makes four standard-library,
dependency-free verifiers mandatory and demotes Wolfram/Sage to optional
cross-checks. All four exit 0 on plain `python3`, and the checks are falsifiable:
mutating `S[0][0]` from `64R-44` to `64R-43` yields exit 1 with an
`AssertionError` at `p28_dominant_product_algebra.py:268`.
Ironically, F1 is a direct consequence of this strength — the effort went into
the dependency-free path, and the demoted Sage path was left unexercised.
### F7 — PASS: the manuscript builds clean
`pdflatex` x3: 0 errors, 0 warnings on the final pass, 17 pages, all
cross-references resolved, zero `??` in the output. Note that pass 1 legitimately
reports ~58 undefined references; anyone grepping a combined `latexmk` log will
see a false alarm.
Cosmetic only (`chktex`/`lacheck`): ~43 missing non-breaking spaces
(`Theorem~3`), 5 wrong-length dashes (`Birkhoff--Poincare`), 3 spaces before
`\ref`, one whitespace-before-punctuation at line 316.
## Recommended actions, in priority order
1. **Fix F1.** Either drop `set -e` around the optional block, or guard each
optional invocation (`sage ... || echo "OPTIONAL: cross-check failed"`), so
that optional means optional. Highest value per unit effort in the whole list.
2. **Fix F2.** Correct `216 + 108*x + 46*x` to `216 + 216*x + 46*x`, or delete
the redundant assertion; the next line already proves the needed identity.
3. **Fix F3.** Coerce the substitution key into the polynomial parent so the
certificate runs on Sage 10.9.
4. **Add the F4 paragraph** separating novel content from cited input.
5. Optionally apply the F7 typography fixes.
None of 13 touches the mathematics. All are edits to certificate plumbing.
## What this review did not do
It did not audit the logical chaining of the manuscript prose — whether
`prop@257 -> lemma@451 -> prop@494 -> prop@539 -> cor@713 -> lemma@930 ->
theorem@1157` discharges each hypothesis without circularity. That remains the
one substantive unverified area. A prior audit (`reviewed PR 20 against the
KKT-F.md`) validated an earlier version of that chain and returned PASS, but 764
lines of `solution.tex` changed afterwards.