# Kernel review — Claude critic A (logical validity), on KERNEL.md v0.1
Review complete. All categories checked (per-derivation T1–T13, tag discipline, minimality, contradictions/holes, weakest inference). No FATALs — every derivation survives a charitable reconstruction. The recurring defect is over-claimed necessity: chosen/stipulated premises yielding conclusions voiced as theorems, plus citation hygiene.

## Per-derivation verdicts
- T1 — CIRCULAR/GAP. Cites K2 for a conclusion that IS K2; uses K1, K5 uncited.
- T2 — GAP. "Route to owner" needs K11; "never penalize attempter" needs K2/K3; none cited. Stated reason is a fairness principle not in the kernel.
- T3 — ENTAILED, phantom citation (K6 does no work).
- T4 — ENTAILED. Clean.
- T5 — ENTAILED from K7+K10. Cleanest fact-only derivation. Vulnerability is upstream in K7's tag, not the logic.
- T6 — ENTAILED. Cost-ranking is normative (K12/K15/K11), not arithmetic (K8); phrased so K8 looks like it delivers it.
- T7 — ENTAILED given K13, but largely restates K13's third clause; "theorem" overstates (K13 is CHOSEN).
- T8 — Substance ENTAILED; cites K6 (unused), uses K7/K3/T1 uncited; one loose sub-inference.
- T9 — GAP/over-reach. "Human-anchored" entailed only for meaning-adjudication (K11); K7/K10 permit tool-based independence for recompute-class — which T5 itself concedes. Universal claim exceeds premises.
- T10 — ENTAILED. Clean; honestly quarantines its empirical question to I5.
- T11 — ENTAILED. Clean.
- T12 — ENTAILED core, but SMUGGLES Kane's psychometric validity taxonomy into a closure-claiming derivation.
- T13 — ENTAILED. Clean.

## Findings ranked
### MAJOR
1. "Theorem" rhetoric re-smuggles the values the kernel honestly tags as chosen (T1 "only definition consistent"; T7 "theorem of the kernel"; T6 "what the arithmetic licenses"). Repair: re-voice as conditional on the chosen premises; reserve "theorem" for fact-only derivations.
2. T1 circular as cited; absolute necessity where it has conditional. Repair: drop K2 from premises; derive anti-consensus from K0+K5.
3. T9 "human-anchored" over-reaches K11; contradicts T5's tool-based independence concession. Repair: scope human-anchoring to meaning.
4. K7/K9 embed contingent "symmetric access" inside FACT tags; load-bearing for the permanent frontier band and T8 defenses. Repair: conditionalize or move to CHOSEN/censused premise.
5. K11's second sentence tags a normative claim ("right to punish") as FACT. Repair: descriptive rewrite.
6. T12 imports Kane's four-part validity taxonomy into a closure-claiming derivation. Repair: derive only "validity argument about an instrument"; move taxonomy to an interface.

### MINOR
7. Systematic citation incompleteness (T1, T2, T8, T9). 8. Phantom citations (K6 in T3/T8). 9. T2's desert-principle reason should be the K2/K3 grounding reason. 10. T8's evidence-perturbation inference valid only for support-removing swaps. 11. K8(i)'s illustrative numbers risk reading as kernel constants. 12. Minimality argued only for K5–K14; K12/K15 overlap within legal scope. 13. T3-corollary vs H1 seam — reconcile in one sentence.

## Weakest inference
T1's closing: "The construct is therefore not a preference among options; it is the only definition consistent with the kernel." The FACTs are overwhelmingly LIMITATIVE; the CONSTRUCTIVE architecture is driven by the telos and the choices. Until Part III's necessity-rhetoric is rewritten as conditional, the derivation layer commits the exact sin the kernel was written to avoid.