b-cw9ii-2

Aion falsification window: the Christ chapter and the opening of the fishes. First batch mined BLIND — the miner was not told which vocabulary metrics are tracked. Mined 82 candidates + 13 attaches / 124 refs; deterministic pre-check 122/124 (2 exact self-duplicates dropped). Fable stage-1 structure gate: 89 SUPPORTED / 31 PARTIAL / 2 WRONG. The two WRONGs were a citation mismatch (evil/good equivalence anchored to §92, which is about Thomas on light and dark) and a reversed anticipation (the Antichrist figure does the anticipating, §114). Fable stage-2 voice gate: 19 flags on 16 refs, dominated by empty source fields on jung-quotes-source (4 of 5 in one chunk) and by Jung glosses absorbed into surrounding reportage. Cross-vendor (Gemini 3.1 Pro, PRE-merge, 6 chunks): 107 SUPPORTED / 12 PARTIAL / 1 WRONG. Adjudication: 1 XV WRONG upheld and dropped (the-devil/christ §78 — relation belongs to §77); 1 dropped for a quote that does not carry its relation (christ-and-church/coniunctio §72); 4 relation fixes; 4 claim_type corrections where XV reversed a stage-2 flag (quotes-source demoted where the quote was Jung glossing, not quoting); 2 sympathetic-reportage splits resolved to REPORTS per protocol default (n=22 Basil, n=31 the Fathers) against XV preferring asserts; 1 kept with a note and lowered confidence where the object is carried from the preceding paragraph (§125). Merged 92 new edges + 26 attaches, 16 new nodes, 118 refs. TRANSFER MEASUREMENT (post-merge, blind to the miner): string reuse against the pre-Aion set at b08b83f collapsed to 2/92 = 2.2%, minting 90 new relation strings, 66 of them sentence-shaped (>=6 hyphen tokens). b-cw9ii-1 measured 70.8% reuse but its miner knew the metric. Material and blinding changed together, so the two causes are confounded — but a 71->2 point drop is not plausibly all material. QA trend: rel_hapax_pct_vocab 86.9->89.0, sentence_edges_pct 21.5->27.2, cq_pass 13->12, cq_fail 2->3. The claims are individually verified; the relational layer got measurably worse. The pre-registered 70% family-fit threshold cannot be evaluated here: 66 of 90 minted strings are one-off clauses that do not sort into the proposed families without inventing one per string. Minted strings with edges, quotes and paragraphs captured in pipeline/work/b-cw9ii-2_minted_strings.json for Track B.