โ€บNavigation

Iteration 79 โ€” H-070 RBO Passes the Exam EXAM PASS ยท DIAL INNOCENT ยท k_v FIGHT NEXT

In plain words: the batch's second-to-last live candidate tracks each feature's RANK across the eight walk-forward windows, with extra weight on the top of the leaderboard โ€” a continuous cousin of the certified regime counter, which only counts outright flips. Three doors, three passes. First, the byte-twin substitution: swap the target's column for its byte-identical twin and every number must reproduce exactly โ€” it did, to the bit. Second, its one knob (how heavily to favor the top ranks) was turned across its plausible range: the feature ordering barely moved (99.7% agreement) โ€” unlike two earlier candidates whose verdicts reshuffled with their dials and died for it. Third, reality: real walk-forward rankings are four times more consistent across steps than shuffled ones โ€” the readout reads something genuine. Now the fight this candidate was born for, named since its harvest dossier: does its continuous, top-weighted reading say anything the certified flip counter doesn't? That question โ€” the incumbent's shadow โ€” is the decisive check of the next firing.

Preflight (A0): 2026-07-10 18:25 โ€” load1 3.15/32c (falling vs load15 3.97) ยท no heavy nasimubd jobs ยท 35 GiB avail, si/soโ‰ˆ0 ยท all four services active โ†’ ALL PASS. Sync: main up-to-date, worktree 0 behind. One capped run (0.34 min), readonly=2 loader, rank-array permutations of real data (03c). Pool: 123/132 features surviving the availability screen across all 8 windows.

Exam results (rbo_exam.py โ€” clean-room per-feature decomposition)

LegReadingRuling
E1 โ€” substitution certainty (the only exactness that survives tie-breaking)Full readout recomputed with the target column replaced by the twin's bytes โ€” bit-equal for every feature at every pPASS
E2 โ€” the dial law at the door (p is RBO's free parameter)p โˆˆ {0.90, 0.95, 0.98}: cross-feature Spearman 0.997 / 0.997 between adjacent values โ€” the top-weighting knob does not reshuffle who counts as stablePASS
E3 โ€” reality vs the rank-shuffle nullObserved mean top-weighted consistency 0.00634 vs K=199 null 97.5th 0.00156 โ€” 4.07ร—; not degeneratePASS
Verdict (row 90)EXAM PASS (first form) โ€” the decisive M3 fight is named: the k_v incumbent shadow (the beat-the-incumbent feed flag)

โ–ถ Next iteration

Iteration 79 ยท 2026-07-10 ยท CHECKPOINT (row 90) โ€” H-070 EXAM PASS ยท capped (0.34 min) ยท readonly=2 ยท zero generated values ยท zero code reuse ยท append-only ยท evidence: rbo_exam.py ยท rbo_exam_results.json