โ€บNavigation

Iteration 81 โ€” H-070 RBO EXCLUDED KILL #23 ยท NOVELTY WITHOUT VALUE ยท MECHANISM DATUM #7

In plain words: the campaign's most original survivor died at the final gate โ€” and taught two lessons on the way down. First, the new kill mechanism: being genuinely NEW is not being USEFUL. Its reading overlapped with nothing in the certified stack, yet on held-out data adding it didn't merely fail to help โ€” it actively hurt, in both halves of the test period. Second, and more valuable than a certificate: its direction reading is significantly BACKWARDS โ€” features that persistently hold top leaderboard positions are LESS likely to keep their edge. That is the campaign's oldest discovery (iteration 9: regime-stable features lose orthogonality) reproduced from an instrument that shares essentially nothing with the one that found it first. The both-extremes-lose mechanism now has seven data points across three unrelated families of instruments โ€” persistence itself, however you measure it, predicts decay. That investigation is now formally proposal-ready for the operator. The boneyard entry leaves a precise door open: if an inverted-direction admission path is ever pre-registered, this readout is the first to revisit.

Preflight (A0): 2026-07-10 19:05 โ€” load1 3.65/32c (steady) ยท no heavy nasimubd jobs ยท 36 GiB avail, si/soโ‰ˆ0 ยท all four services active โ†’ ALL PASS. Sync: main up-to-date, worktree 0 behind. One capped run (9.05 min, era attack inline), readonly=2 loader.

Finale on the sealed apparatus (rbo_m4_m5b.py)

LegReadingRuling
M4 held-out liftฮ”AUC โˆ’0.00167, CI95 [โˆ’0.00336, โˆ’0.00041] โ€” entirely negative: actively harmful (the H-060 territory) ยท n_have 2,266FAIL
Era attack (inline)Early โˆ’0.0034 (CI entirely negative) ยท late โˆ’0.00066 (spans 0) โ€” consistently harmful, no flipHARM IS ERA-STABLE
M5โ€ฒb directionฯ = โˆ’0.184, CI [โˆ’0.219, โˆ’0.150], null band 0.038 โ€” significantly INVERTED: top-rank persistence predicts orthogonality LOSSFAIL (one-sided) ยท INFORMATIVE
Frozen disjunctionBoth legs failEXCLUDE
Verdict (row 92)EXCLUDE(novel but valueless โ€” actively harmful M4, significantly inverted M5โ€ฒ) โ€” KILL #23, BONEYARD #38. New kill mechanism for the taxonomy: NOVELTY WITHOUT VALUE (M3-orthogonal โ‰  M4-valuable)
Mechanism datum #7 โ€” the strongest cross-channel corroboration yet: iteration 9's original inverted discovery (regime-stable features lose orthogonality, from the k_v channel) is now reproduced from a channel sharing NOTHING with k_v (shadow 0.019): persistence-of-relevance itself predicts orthogonality loss, measured two independent ways โ€” plus the three distributional meters' negative directions (data #4โ€“6). Seven corroborations across three readout families: the inverted-mechanism investigation is proposal-ready, flagged for the operator with emphasis. Second operator-facing observation: the frozen one-sided M5โ€ฒ rule cannot admit inverted-direction instruments โ€” an inverted-admission path, if ever pre-registered, would revisit this readout first.

โ–ถ Next iteration

Iteration 81 ยท 2026-07-10 ยท VERDICT (row 92) โ€” EXCLUDE ยท kill #23 ยท boneyard #38 ยท capped (9.05 min) ยท readonly=2 ยท zero generated values ยท append-only ยท evidence: rbo_m4_m5b.py ยท rbo_m4_m5b_results.json