{"assessment":"Structural facts: 24 items split 12 per domain, four reader arms (M0_C, M0_D, MP_C, MP_D), and two rewrite conditions of 30 items each (24 core items plus 6 Z items). Several headline gate numbers are properties of the design rather than findings: in the correct baseline rewrite condition every wrong output is necessarily a status change, so delta_gate_wrong_admitted of 0 and correct_repairs of 0 are definitional; Gspan equals raw in every table, so the span or quote gate had no observable effect; and Gdelta in the correct condition equals raw because the 8 wrong outputs it routes to HOLD were already HOLD, so the gate changed nothing. In the wrong baseline condition the same delta gate admits all 8 unchanged wrong outputs and converts all 22 correct repairs to HOLD, dropping correct from 22 to 12 of 30, which contradicts the structural summary line delta_gate_wrong_admitted of 0 unless that line is scoped to one condition only. Empirical findings: zero unsafe verdicts anywhere; all errors are HOLD on gold PASS, concentrated on the P items (all six P items held by nearly every reader) and on A01R and A03R; reader accuracy ranges 16 to 18 of 22 to 24 with differences between M0 and MP, and between C and D, of one to two items on reduced n; pair topology scores 16 of 22, identical to singleC and one below singleD; the model readily moves baselines toward NOT ESTABLISHED and REFUTED but repairs NOT ESTABLISHED toward SATISFIED only on 4 of 12 PASS items and never on P items; blind review agrees 30 of 30 while simultaneously logging a Rule 1 instrument label mismatch covering 10 of the 12 P and R items. Interpretations that would be overclaims: that the delta gate catches wrong outputs (it has no net effect where wrong outputs are already HOLD and it destroys repairs where repairs occur); that pairing or any gate improves safety (no unsafe events exist in any arm, so there is nothing to catch and a floor effect prevents any demonstration); that MP outperforms M0 or D outperforms C (differences are within one or two items on n near 22, with missing data unevenly distributed); that both_wrong pairs are independent concurring errors (the both_wrong items are exactly the P items flagged by the defect, so shared item construction is the likelier common cause, and C and D sharing is not ruled out by separate calls); and that 30 of 30 blind agreement validates the gold (the review itself records a label mismatch that makes HOLD a defensible verdict on those items, so false_hold on P items may be a specification defect rather than reader error). Ledger inconsistencies also exist: item_outputs shows four reader slots MISSING (M0_C A01R, M0_D B03P, MP_D A01R, MP_D A02P) plus one rewrite output MISSING (correct condition A01R), but calls lists only 2 empties; missing is scored as a wrong HOLD in rewrites yet excluded from n in reader tables; the figures quote_gate_wrong_admitted 48 and joint_routed_unsafe 4 cannot be reconciled from any table shown.","caveats":["Zero unsafe outcomes in every arm means no safety mechanism (pairing, span gate, delta gate) has been tested against the failure mode it is meant to catch; absence of unsafe is a floor, not evidence of gate effectiveness.","delta_gate_wrong_admitted 0 is true only in the correct baseline condition by construction; in the wrong baseline condition the same gate admits 8 wrong outputs and blocks all 22 correct repairs (Gdelta 12 of 30 versus raw 22 of 30).","Gspan is identical to raw in all nine rewrite tables, so the quote or span gate contributed nothing observable and claims about it are uninformative.","The blind review defect (Rule 1 names a specific instrument while records are generic approvals) covers A01P, A01R, A02P, A02R, A03P, A03R, B02P, B02R, B03P, B03R; these overlap almost exactly with the false_hold set, so false_hold counts may measure a gold or specification problem rather than reader error.","B01P is held by all four readers and is not covered by the stated defect; its failure cause is unexplained.","both_wrong lists (M0: A01P, A02P, A03P, B01P, B02P; MP: A01P, A03P, A03R, B01P) are HOLD on HOLD agreements on items with a shared construction issue; they must not be read as independent concurring judgments, and C and D independence is not established by separate calls.","Blind review 30 of 30 agreement does not establish gold correctness, especially since the same review documents a label mismatch and then rationalises it.","Reader and topology comparisons differ by one to two items on n of 22 to 24; no comparison here supports a claim that MP beats M0, that D beats C, that pairing adds value, or that domain A is harder than B.","Missing outputs are not random: all missing reader slots are gold PASS items (A01R, A02P, B03P), so excluding them from n inflates reader accuracy relative to counting them as HOLD, and the two scoring conventions (exclude in readers, score as wrong HOLD in rewrites) are inconsistent.","Call ledger does not reconcile: 2 empties recorded against at least 5 MISSING outputs across readers and rewrites; 158 recorded calls versus 156 implied by 96 reader plus 60 rewrite calls is unexplained.","quote_gate_wrong_admitted 48 and joint_routed_unsafe 4 cannot be derived from the tables provided and should be treated as unverified.","Wrong baselines in the rewrite condition follow a fixed pattern by gold class (P and R to NOT ESTABLISHED, N and Z to REFUTED, F to SATISFIED); the 22 repairs may reflect pattern matching on this regular seeding rather than evidence based correction.","Rewrite correct condition shows the model corrupted 8 of 24 correct SATISFIED baselines into NOT ESTABLISHED; this is a real finding about rewrite instability on PASS items, independent of any gate."],"next_checks":["Reconcile the call ledger: list every MISSING slot (M0_C A01R, M0_D B03P, MP_D A01R, MP_D A02P, correct rewrite A01R) against recorded calls and explain why only 2 empties appear and where the extra 2 recorded calls came from.","Derive or correct quote_gate_wrong_admitted 48 and joint_routed_unsafe 4 from item level data, or remove them.","Re-score all reader and topology tables with MISSING counted as HOLD and report both conventions side by side.","Adjudicate the Rule 1 instrument label mismatch: either amend Rule 1 or the record descriptions and rerun the 10 affected items, or formally decide whether HOLD is an acceptable verdict on them; then recompute false_hold.","Investigate B01P separately since it is held by all readers and not covered by the documented defect.","Scope the delta gate claim explicitly per condition and report Gdelta cost (22 repairs lost) alongside any benefit; test the gate on seeded cases where a rewrite changes toward an unsafe verdict, since no such case exists in this run.","Add adversarial items that elicit unsafe verdicts in at least one arm so that pairing and gates can be evaluated against actual unsafe events rather than a zero floor.","Report exact binomial or paired confidence intervals for M0 versus MP and C versus D on matched items, and state that current differences are not distinguishable from noise.","Confirm whether C and D readers share model, prompt, retrieval, or context; measure agreement on items with randomised evidence to estimate baseline correlation before interpreting both_wrong or disagree counts.","Randomise wrong baseline assignments in the rewrite condition