Every stage, with its own denominator

B Stage-specific numerical results

The tables below preserve their original study-specific denominators and primary weights. They are not pooled estimates of one population or a ranking across cohorts. OII-1–3 mean errors use their saved conservative score convention; the OII-4 endpoint uses its saved full-law numerical scores with separate event certificates. Dashes mean that the indicated control was not separately evaluated, not that its count was zero.

B.1 OII-1 and OII-2

Acquisition/predictor	Passes /420	Mean TV upper	All-core /28	Red /28
active/factorized	230	0.369042	9	–
active/memoryless	102	0.709150	3	–
active/selected	281	0.223891	12	15
passive/factorized	216	0.371787	10	–
passive/memoryless	103	0.713242	3	–
passive/selected	266	0.218708	13	13

Method	Passes /720	Mean TV upper	All-core /36	Red /36	Unsafe /4,990
refined_active	633	0.086537	21	14	43
refined_passive	630	0.078505	21	11	43
oii1_active	442	0.263312	14	21	85
no_confirmed_constraints	639	0.080784	21	–	–
memoryless	139	0.762637	3	–	–
factorized	502	0.256431	15	–	–

The OII-1 common core contains 420 slots per predictor. The expanded 705-slot panel included learner-selected probes and is not substituted for that common comparison. The two OII-1 unsafe-merge searches visited different sets of histories, so their 56 and 131 witnesses are not directly comparable population rates.

OII-2's no-confirmed-constraint ablation uses the same active dataset, including observations obtained during conditional counterexample replay. It tests the added fitting constraints, not the value of those replay observations themselves. Its red and common-merge outcomes were not separately searched. OII-2's memoryless and factorized controls are retained in their documented role; they should not be confused with the separate OII-1 controls.

B.2 All OII-3 V2 selectors

The labels identify the acquisition lane and selector. “CE” is the conditional-counterexample lane, “disagreement” omits that replay, and “passive” uses passive acquisition. INC is the likelihood-selected incumbent. L is a likelihood mixture; M is the empirical minimax selector; U optimizes a confidence upper bound; SAFE applies the measured-panel replacement rule. BASE, FREE and CON retain their saved candidate-bank/constraint meanings. The OII-2 controls have the same total call allowance but their own allocation rather than the new lanes' dedicated calibration design.

B.3 All OII-4 mechanism families

B.4 Deduplication sensitivity

Method	Literal passes /1,140	Behavior passes /1,135	All-core /48
R	1052	1048	37
S	1052	1048	37
E	902	898	28
G	556	553	15
B	1042	1038	36
P	958	954	24
O	858	854	23
H	1	1	0

Finite behavior means equality for all possible output strings at the matched horizon, not equality only on paths of one particular source. Calibration-matching contexts remain included. All-core counts are unchanged.

Removing duplicate within-fixture specifications or finite policy behaviors changes the primary weighting. This table is therefore a retrospective sensitivity only. It retains specification matches with calibration; deduplication is not a leave-calibration-out test. All-core fixture pass counts are unchanged because removal of an identical duplicated test cannot alter whether every distinct core context passed. The reported decimal law scores are those already stored; no models were refitted for the sensitivity.