==================================================================================================== (b) CENSUS [as_ledgered] Opus 5.5 rows with an SE: 86; wins 26, nulls 38, losses 22 col_0 win null loss ctype carried_self 0 1 0 carried_w2s 0 7 5 layer_vs_bare 15 4 2 layer_vs_prior 6 3 0 layer_vs_rival 5 22 13 system_vs_single 0 1 0 trained_vs_frontier 0 0 2 [corrected] Opus 5.5 rows with an SE: 78; wins 19, nulls 41, losses 18 col_0 win null loss ctype carried_self 0 1 0 carried_w2s 0 8 4 layer_vs_bare 12 5 0 layer_vs_prior 6 3 0 layer_vs_rival 1 23 12 system_vs_single 0 1 0 trained_vs_frontier 0 0 2 removed: {'L112': 'mislabelled control (audit finding)', 'L158': 'threshold choice (audit finding)', 'L164': 'impaired Opus-built ACE playbook (audit finding)', 'L193': 'budget-contract voiding + no-thinking swapped turns (audit finding)', 'L194': 'budget-contract voiding (audit finding)', 'L052': 'model succession UP infra flip, withdrawn (program record)'} changed: L156 -> null (TIE); L165 -> corrected +7.1 [0, +16.7] (z 1.67); L014 -> corrected -0.54 (z -1.56); L184 -> corrected -0.223 (z -2.00); L174, L176 -> mixed [corrected_minus_rival_advantage_losses] Opus 5.5 rows with an SE: 74; wins 19, nulls 41, losses 14 col_0 win null loss ctype carried_self 0 1 0 carried_w2s 0 8 4 layer_vs_bare 12 5 0 layer_vs_prior 6 3 0 layer_vs_rival 1 23 8 system_vs_single 0 1 0 trained_vs_frontier 0 0 2 ==================================================================================================== (c) HEADROOM FIT, Opus 5.5 rows with a no-layer control (the record's model code, section 3A) id ctype contrast y s h flags L003 layer_vs_prior lineage - bare_notice 0.00 2.35 2.3 prereg,ceiling L004 layer_vs_bare lineage - bare 5.75 2.64 8.0 prereg L010 carried_w2s living - bare (AppWorld probes H0-H41, item-paired) -1.10 0.84 0.0 ceiling,prereg L011 carried_w2s living - bare 0.00 10.61 10.0 ceiling,prereg L012 carried_w2s living - bare 1.40 4.03 11.0 ceiling,prereg L016 carried_w2s living - bare -1.67 4.28 5.0 L057 layer_vs_bare a5 - fresh 15.20 6.59 25.8 L099 layer_vs_bare sensor - bare 10.80 4.46 11.0 L124 layer_vs_prior vp3 - prior 11.90 2.73 16.7 prereg L126 layer_vs_prior vp3 - prior 10.80 2.19 15.6 prereg L136 layer_vs_prior vp3.1 - prior 2.68 1.18 9.7 L147 layer_vs_prior vp2 - prior 9.30 3.62 13.1 L163 layer_vs_bare layer+sensor - bare 4.30 1.35 10.1 prereg L166 layer_vs_bare v3 - drifted bare 13.20 2.78 19.3 posthoc L174 layer_vs_bare Opus+layer - Opus alone 6.70 1.71 15.1 audited_cross -- all 15 (as published): n=15 head: k = 0.539 [0.427, 0.675] (profile 95%), tau 0.98 head+bias: k = 0.641 [0.49, 0.865], b = 1.59 [-0.05, 3.85] constant lift (bias model) = 5.71 pts AIC zero 107.1, constant 96.8, head 77.2, head+bias 75.6 -- drop carried-from-weaker rows L010 L011 L012 L016: n=11 head: k = 0.563 [0.443, 0.713] (profile 95%), tau 1.21 head+bias: k = 0.789 [0.498, 1.08], b = 3.38 [-0.2, 6.95] constant lift (bias model) = 7.25 pts AIC zero 81.8, constant 68.1, head 55.5, head+bias 55.0 -- drop L124 (subset of L126): n=14 head: k = 0.514 [0.4, 0.655] (profile 95%), tau 0.85 head+bias: k = 0.618 [0.463, 0.835], b = 1.52 [-0.1, 3.65] constant lift (bias model) = 5.2 pts AIC zero 98.7, constant 90.1, head 72.0, head+bias 70.6 -- drop L057 (Haiku->Opus mid-task handoff): n=14 head: k = 0.537 [0.422, 0.677] (profile 95%), tau 0.98 head+bias: k = 0.641 [0.485, 0.88], b = 1.59 [-0.1, 3.95] constant lift (bias model) = 5.41 pts AIC zero 98.2, constant 89.3, head 71.5, head+bias 70.0 -- drop L099 (unaudited) and L166 (post hoc): n=13 head: k = 0.499 [0.383, 0.643] (profile 95%), tau 0.78 head+bias: k = 0.608 [0.448, 0.825], b = 1.51 [-0.1, 3.6] constant lift (bias model) = 4.78 pts AIC zero 90.1, constant 82.9, head 65.9, head+bias 64.4 -- drop L174 (mixed-model stream): n=14 head: k = 0.568 [0.44, 0.718] (profile 95%), tau 1.15 head+bias: k = 0.674 [0.508, 0.93], b = 1.69 [-0.1, 4.2] constant lift (bias model) = 5.64 pts AIC zero 100.4, constant 91.9, head 73.3, head+bias 71.8 -- drop all of the above (R2 combined): n=6 head: k = 0.503 [0.338, 0.748] (profile 95%), tau 1.01 head+bias: k = 0.792 [0.338, 1.268], b = 3.44 [-1.95, 8.1] constant lift (bias model) = 5.02 pts AIC zero 41.4, constant 36.5, head 30.1, head+bias 30.4 ==================================================================================================== (a) STRENGTH GAP RULE: A row counts as a carried-knowledge case only if (i) the reader is given knowledge distilled from a source model's past episodes (a store, playbook, rules, recipes, minted checks or a compiled digest), not the in-progress work product of the current task; (ii) it is compared with the same reader without that knowledge; (iii) the reading is not WITHDRAWN or DO-NOT-USE in the program's audits; corrected values replace ledger values. Applied identically to both directions. included rows: id source_model reader gap est z sig source L001 opus-5.5 haiku -2 7.0000 3.25 True AppWorld: Opus 5.5-compiled AppWorld store read by Haiku L002 opus-5.5 haiku -2 5.1600 1.67 False AppWorld: Opus 5.5-compiled AppWorld store read by Haiku L017 opus-4.8 haiku -1 6.6667 0.84 False Long-running harness: Opus 4.8-built living store read by Haiku L015 opus-4.8 haiku -1 -9.0000 -1.88 False Long-running harness: Opus 4.8-built living store read by Haiku L026 opus-4.8 haiku -1 0.0940 NaN False Long-running harness: Opus 4.8-built living store read by Haiku L103 opus-5.5 qwen -2 0.4760 3.74 True Long-horizon: full_v2 capital minted by Opus 5.5 L111 opus-5.5 opus-4.8 -1 0.0000 NaN False Long-horizon: full_v2 capital minted by Opus 5.5 L018 opus-4.8 opus-4.8 0 -6.6667 -1.35 False self (same model) L058 haiku haiku 0 -17.9000 -3.64 True self (same model) L182 opus-4.8 opus-4.8 0 0.2100 NaN False self (same model) L191 opus-4.8 opus-4.8 0 0.1850 NaN False self (same model) L084 opus-5.5 opus-5.5 0 0.0200 1.12 False self (same model) L014 opus-4.8 opus-5.5 1 -0.5400 -1.56 False Long-running harness (horizon run): Opus 4.8 kernel store, offline L016 opus-4.8 opus-5.5 1 -1.6667 -0.39 False Long-running harness: Opus 4.8-built living store, AppWorld fresh tasks L183 opus-4.8 opus-5.5 1 -0.0300 NaN False Training: Opus 4.8 kernel store, served vs not (O-K arms and item-exposure run) L184 opus-4.8 opus-5.5 1 -0.2230 -2.00 True Training: Opus 4.8 kernel store, served vs not (O-K arms and item-exposure run) L185 opus-4.8 opus-5.5 1 -0.4830 -2.06 True Training: Opus 4.8 kernel store, served vs not (O-K arms and item-exposure run) L186 opus-4.8 opus-5.5 1 -0.2660 -2.58 True Training: Opus 4.8 kernel store, served vs not (O-K arms and item-exposure run) L187 opus-4.8 opus-5.5 1 -0.3470 -1.10 False Training: Opus 4.8 kernel store, served vs not (O-K arms and item-exposure run) L188 opus-4.8 opus-5.5 1 -0.3590 -2.06 True Training: Opus 4.8 kernel store, served vs not (O-K arms and item-exposure run) L028 opus-4.8 opus-5.5 1 -3.0000 -1.34 False Long-running harness: Opus 4.8-built living store, AppWorld fresh tasks excluded rows: L098 (iii) wave-2 lineage layer over tickets 41-50: DO-NOT-USE (audit of the wave-1/2 handover readings) L104 (i) guard: hard checks from the ticket's own examples, no distilled knowledge L105 (i) commit: commits the deployed tree after every ticket, no distilled knowledge L049 (i) mid-task handoff of the current task (DOWN, ours - Opus 4.8 alone) L054 (i) mid-task handoff of the current task (a5 - Opus 4.8 alone) L051 (i) mid-task handoff of the current task (Opus -> Haiku) L059 (i) kernel succession chain: the successor continues the predecessor's work product L057 (i) mid-task handoff of the current task (Haiku -> Opus 5.5, vs fresh) L056 (i) mid-task handoff of the current task (Haiku -> Opus 5.5, vs transcript) downhill (source stronger than reader): 5 of 6 nonzero helped (1 tie); significant: 2 helped, 0 hurt; independent sources 3: AppWorld: Opus 5.5-compiled AppWorld store read by Haiku: signs [1, 1], significant [True, False] Long-horizon: full_v2 capital minted by Opus 5.5: signs [1, 0], significant [True, False] Long-running harness: Opus 4.8-built living store read by Haiku: signs [1, -1, 1], significant [False, False, False] same model: 3 of 5 nonzero helped (0 tie); significant: 0 helped, 1 hurt; independent sources 1: self (same model): signs [-1, -1, 1, 1, 1], significant [False, True, False, False, False] uphill (source weaker or earlier than reader): 0 of 9 nonzero helped (0 tie); significant: 0 helped, 4 hurt; independent sources 3: Long-running harness (horizon run): Opus 4.8 kernel store, offline: signs [-1], significant [False] Long-running harness: Opus 4.8-built living store, AppWorld fresh tasks: signs [-1, -1], significant [False, False] Training: Opus 4.8 kernel store, served vs not (O-K arms and item-exposure run): signs [-1, -1, -1, -1, -1, -1], significant [False, True, True, True, False, True] row-level Spearman(gap, sign) under the rule = -0.764 (nominal p 5.5e-05, n 21; rows not independent) source-level table (net sign = sign of the sum of member signs): source side net_sign any_sig members AppWorld: Opus 5.5-compiled AppWorld store read by Haiku down 1 True [L001, L002] Long-horizon: full_v2 capital minted by Opus 5.5 down 1 True [L103, L111] Long-running harness (horizon run): Opus 4.8 kernel store, offline up -1 False [L014] Long-running harness: Opus 4.8-built living store read by Haiku down 1 False [L017, L015, L026] Long-running harness: Opus 4.8-built living store, AppWorld fresh tasks up -1 False [L016, L028] Training: Opus 4.8 kernel store, served vs not (O-K arms and item-exposure run) up -1 True [L183, L184, L185, L186, L187, L188] sources helped: downhill 3 of 3, uphill 0 of 3; Fisher exact p two-sided 0.100, one-sided 0.050 source-level Spearman(gap, net sign) = -0.933, exact permutation p = 0.100 (n 6 sources) wrote recount_output.json