## 1. REPAIR AUDIT

1. **C1 — PARTIAL.** It is now CP1, owned and versioned, and correctly says: “No uniqueness is claimed.” But it retains the rejected constraint: “Selected from admissible constructs — those with reconstructible referents (K14/K15) fixed per project-version (K13).” K14/K15 require reconstructible verdict grounds, not reconstructible construct referents.

2. **D1 — IMPLEMENTED.** The chosen underdetermine disposition exists in K14, and routing is correctly split: “semantic ambiguity → the owner as a K13 amendment question; record gaps → the K14 integrity process…; any question of worker responsibility for the gap → a K18 attribution proceeding.”

3. **D2 — PARTIAL.** The three required states and chosen abstention threshold exist. But r2 required I1 classification to be an operationalization; v0.3 instead labels D2 `[UNPACKED]` and claims necessity: “P4 is why the classifier must exist at intake.”

4. **D3 — IMPLEMENTED.** It derives only: “a negative verdict may claim no more than the completeness guarantee of its disclosed procedure and scope,” makes checklists a design response, and says planted recall merely “inform[s]” confidence until transport is validated.

5. **U1 — IMPLEMENTED.** Renamed D4 and correctly labeled `[DESIGN RESPONSE]`; it names “owner: program governance,” places granularity at I3, and inherits the document’s v0.3/date stamp.

6. **U2 — IMPLEMENTED.** It is split into a derived human/legal floor, a design-response validation gate, and CP4 residual prices “strictly after and subordinate to K17, K18, K21/K22, and K23.”

7. **U3 — IMPLEMENTED.** It is a CP3 governed choice, distinguishes “Authority-finality,” “epistemic accuracy,” and “non-identifiability,” and permits uncertainty to be “unknown or partially identified under named assumptions.”

8. **U4 — IMPLEMENTED.** The audit architecture is `[DESIGN RESPONSE]`, and continuation is “reviewed against measured dependence, gaming pressure, and frontier mass,” rather than declared permanent.

9. **U5 — PARTIAL.** U5 correctly moves sterilization, thresholds, stamps, and equating into an “Adopted implementation” and permits transparent recalibration. But CP9 reinstates the rejected entailment: “K16 forces the stamp.”

10. **D5 — PARTIAL.** It adds controls and I6 criteria, but not the complete prescribed threat/control matrix. Reviewers, collusion controls, and appeal-manipulation controls remain unnamed; the text only says: “a control matrix covering every actor” before listing an incomplete set.

11. **D6 — IMPLEMENTED.** The person/rank ban is explicitly “Adopted,” effects are measured at I6, and diversity monitoring feeds “the K20 escalation channel.”

12. **D7 — IMPLEMENTED.** The derivation is narrowed exactly to: “all verdicts owe the same admissibility protocol regardless of issuer.” Reliability and bias study design is moved to I5.

13. **D8 — IMPLEMENTED.** Only the applicable bar, dates, and grounds are derived; replay is a design response; reopening is separately governed by K24.

14. **K5 — PARTIAL.** The false reliability ceiling is removed, but the replacement protocol remains inside F1’s `[FACT]` tag: “no reliability advantage of evaluator over producer may be presumed; if claimed, it must be measured.” R2 expressly required that sentence to be protocol or `[COND]`, not mathematics.

15. **K6 — IMPLEMENTED.** F2 is restricted to “an explicitly existential claim with a valid witness relation” and caps negative results at their disclosed procedure and scope.

16. **K7 — PARTIAL.** F3 adds dependence, estimation-target, fallible-reference, and exchangeability qualifications. But calibration-to-deployment transport remains in `[FACT]`: “those assumptions are claims to be defended, not defaults,” rather than being placed in a deployment-transport I-value as prescribed.

17. **K8 — PARTIAL.** The corrected joint-error statement is present verbatim in substance. But F4 retains a normative measurement rule under `[FACT]`: “Whether given evaluators’ errors are dependent is an empirical property to be measured or structurally justified, never assumed away.”

18. **K9 — IMPLEMENTED.** F5 is conditional on actual underdetermination; material prevalence is separately presumed in P1 and measured at I6.

19. **K10 — IMPLEMENTED.** P2 now describes “non-negligible incentive and information” as a presumption and requires magnitudes and behavior to be measured at I6.

20. **K11 — IMPLEMENTED.** P3 keeps only the conditional shared-substrate presumption, moves ecology to I9, and states: “No trend about the future supply of independence is asserted.”

21. **K12 — IMPLEMENTED.** Frontier mass and producer behavior are measured at I1/I6, the permanent-band claim is expressly retracted, and lanes persist only while measured mass exceeds declared thresholds.

22. **K17/K18/U2 precedence — IMPLEMENTED.** K17 protection is “not tradeable,” K18 cannot be collapsed by error prices, and CP4 is expressly subordinate to both.

23. **K18/K21 rider ambiguity — IMPLEMENTED.** K22 unambiguously elects the human-decision rule “globally, by choice,” and U2 uses that global floor.

24. **K20/H4 — IMPLEMENTED.** K20 is expressly “detection and escalation, not defense,” while CP5 names enterprise governance and owns the take/keep/fire decision.

25. **Fairness — IMPLEMENTED.** K23 creates the explicit `[CHOSEN]` commitment; CP6 names enterprise governance with counsel as owner of metrics and bounds.

26. **K17–K21 tag mixing — PARTIAL.** K17, K18, K19, and the K21/K22 split implement the repair. K20 still places a factual proposition inside `[CHOSEN]`: “Instructions can mis-capture the judgment the data exists to serve,” without separating it as a condition or marked motivation.

27. **Category (e) application — IMPLEMENTED.** The former H1–H5 choices, C1, I8 fairness choices, and I2 error prices now map to CP1–CP9 with named owners, version obligations, and effective dates.

28. **Dependency-table failures — IMPLEMENTED.** It is renamed a “Dependency and rationale map”; K1–K4 are added; the advisory-first, permanent-band, K18-human-floor, K20-defense, blinding, legacy-reference, and universal-justifiability claims are removed or corrected. A newly introduced assertion is addressed below.

29. **Tagging-rule self-violation — IMPLEMENTED.** The former Parts I–II universal tagging rule is gone, and Parts II/VI are explicitly declared untagged meta-analysis. A different new Part VI violation is addressed below.

## 2. NEW DEFECTS

1. **MAJOR — P4 and D2 form a new circular justification, while P4 mixes obligations into `[COND]`.**

   P4 says: “what the kernel retains is the *obligation to keep measuring it* and the *capability to route on it* (D2).”

   D2 then says: “P4 is why the classifier must exist at intake.”

   Thus P4 relies on D2 for the routing capability while D2 relies on P4 to establish that the classifier must exist. The obligation and architectural capability are also not empirical presumptions suitable for `[COND]`.

2. **MAJOR — CP8 introduces an unsupported derived aggregation rule.**

   > “The derivable part is not open: failure of a necessary (‘must’) clause fails the item, by K2’s construct-fixing contract.”

   K2 merely defines quality as whatever construct the contract fixes. It does not entail that every clause written as “must” is an item-level necessary condition or determine the aggregation consequence. This is precisely a governed aggregation choice being presented as derived.

3. **MINOR — D5’s perturbation inference is newly strengthened beyond entailment.**

   > “a support-removing evidence swap that leaves a verdict standing shows the verdict was not grounded in that support”

   A verdict may have redundant independent support; removing one support while the verdict survives does not show that the original verdict was ungrounded in it. At most, the result motivates re-examination.

4. **MINOR — K24 contradicts the changelog’s characterization.**

   K24 says: “Reopening a past leniency under the *original* bar is permitted only through the K18 proceedings.”

   The changelog calls K24 “exculpatory-only reopening.” It is therefore not exculpatory-only.

5. **MINOR — Part VI contains a substantive untagged obligation despite being exempted only as meta-analysis.**

   The document says: “Parts II and VI are meta-analysis and carry no tags by design.”

   Part VI then states: “the kernel is re-attacked, by at least one reviewer with no stake in its survival, at every version bump and at least once per model generation.” That is an operative governance obligation, not meta-analysis, and it has no discipline category, owner, or governed choice-point.

6. **MINOR — The repaired dependency map introduces a false universal dependency.**

   > “The vocabulary every element uses; K2 feeds CP1 and K20; K4 feeds D7”

   Not every element uses K1–K4’s vocabulary; K0, for example, uses none of the four defined terms. The row overstates the dependency instead of identifying K1–K4’s actual downstream consumers.

**REPAIRS: 21 implemented / 8 partial / 0 not implemented; NEW DEFECTS: 2 major / 4 minor**

**Single most important remaining item:** CP8’s unsupported claim that item-level failure is derivable from K2; it turns a consequential aggregation choice into a false theorem.