v162: audited my own correction. Per-subset break-even bars (0.3304%-0.3939%) change NO verdict - path+relabel's own bar is stricter than the one I borrowed and it still clears on a zero bound. The retracted 8.685% never licensed a decision. v139's one-flip multiple was wrong the OTHER way (12x vs 1.69x). Formula error N*G/C vs N*G/(G+C) found by an internal contradiction. 1315 tests, 20/20 mutations fire

barsweep_report.json

@@ -0,0 +1,72 @@
+{
+  "question": "last turn's corrected break-even p* depends on g, which DIFFERS PER CONFIGURATION - yet I compared all four subsets against a single bar derived from path+relabel+ws. Does the licensing verdict survive per-subset bars?",
+  "prediction": "no verdict changes: zero sits below any positive bar, and the lossy subsets are 19-37x over, which no bar movement of a few percent can close. The error is real but harmless",
+  "prediction_outcome": "CONFIRMED - no verdict changes on any of the four subsets",
+  "per_subset_bars": {
+    "path+relabel": {
+      "component": 0.01955764752071971,
+      "g": 0.03911529504143942,
+      "own_bar": 0.0033041479473187095,
+      "bound": 0.0,
+      "clears_own_bar": true
+    },
+    "path+relabel+role": {
+      "component": 0.02264294055037343,
+      "g": 0.04528588110074686,
+      "own_bar": 0.0038214060567554886,
+      "bound": 0.12666666666666668,
+      "clears_own_bar": false
+    },
+    "path+relabel+ws": {
+      "component": 0.02025688625204707,
+      "g": 0.04051377250409414,
+      "own_bar": 0.0034214717900835703,
+      "bound": 0.06666666666666667,
+      "clears_own_bar": false
+    },
+    "path+relabel+ws+role": {
+      "component": 0.023345477164770423,
+      "g": 0.046690954329540846,
+      "own_bar": 0.003939037756312617,
+      "bound": 0.18666666666666668,
+      "clears_own_bar": false
+    }
+  },
+  "uniform_bar_applied_last_turn": 0.00342147,
+  "bar_spread": {
+    "min": 0.0033041479473187095,
+    "max": 0.003939037756312617,
+    "range_factor": 1.192,
+    "note": "p* is INCREASING in g, so a bigger gain tolerates a bigger flip rate. path+relabel's own bar (0.3304%) is STRICTER than the 0.3421% I borrowed, and it still clears because its bound is EXACTLY ZERO"
+  },
+  "impact_audit_of_the_retracted_constant": {
+    "records_mentioning_8.685": [
+      "findings",
+      "subset_robustness",
+      "flip_risk_structural_bound",
+      "flip_bound_and_corrected_breakeven"
+    ],
+    "decisions_it_licensed": "NONE. Every recorded decision in those entries was \"promote NONE\". The 25.4x error made the blocker look 25x less severe than it was WITHOUT ever unlocking anything",
+    "honest_scope": "the constant was wrong everywhere it appeared, but it never gated a promotion, so no prior decision needs reversing"
+  },
+  "v139_multiple_also_wrong_but_the_other_way": {
+    "v139_claim": "a single flip erases the stacked gain about 12 times over",
+    "corrected": 1.6924607173370732,
+    "factor": 7.09,
+    "direction": "v139 OVERSTATED the multiple by ~7x, while the same root cause UNDERSTATED the break-even rate by 25x. The two errors are independent and point OPPOSITE ways",
+    "my_own_slips": [
+      "I called the corrected 1.69x 'even harsher' than 12x - 1.69 is SMALLER, so it is LESS severe",
+      "I said v139 was '7.12x too LOW' when 12 vs 1.69 means too HIGH",
+      "I generalised that all my break-even calculations were 'too generous' when this one was too PESSIMISTIC"
+    ]
+  },
+  "formula_error_found_by_internal_contradiction": {
+    "symptom": "two of my own figures for the same quantity disagreed by 0.394% (150*p* = 0.590856 vs stack/one_flip = 0.593192)",
+    "diagnosis": "my back-of-envelope used N*G/C when the correct expression is N*G/(G+C) - it OMITS the gain the flipped challenge also stops earning, overstating tolerance",
+    "resolution": "the BAR was right and the shortcut was wrong; N*G/(G+C) reproduces 150*p* exactly",
+    "final": "one flip costs 1.69x the four-way stack's entire gain over 150 challenges, and the tolerable count is 0.5909 flips",
+    "lesson": "trusting a contradiction between two of my own numbers found a FORMULA error, not a typo - the third time this pattern has paid"
+  },
+  "conclusion": "the licensing verdict from last turn stands unchanged under per-subset bars: pathprefix+relabel is the only configuration that clears, and it clears on a bound of exactly zero. The corrections tighten the accounting without moving the decision",
+  "still_not_promoted": "shipped byte-identical to pristine; the comprehension question remains unanswerable by any bound"
+}