Iteration 80 — H-070 RBO M3 CLEAR: the Campaign's Most Orthogonal Survivor

Iteration 80 — H-070 RBO M3 CLEAR k_v SHADOW 0.019 · MOST ORTHOGONAL SURVIVOR · FINALE NEXT

In plain words: the fight this candidate was born for ended in the most lopsided result of the campaign — in the candidate's favor. Its continuous, top-weighted reading of how features hold their leaderboard positions was expected to overlap heavily with the certified flip counter, since both nominally track walk-forward persistence. Measured across all 123 features: they share essentially NOTHING (correlation 0.019 — where every previous candidate measured at least a moderate ±0.35 against the same counter). And against the entire 21-instrument certified stack, its worst overlap is 0.63 — every previous survivor's worst was at least 0.84. This is the most genuinely NEW reading to survive the redundancy gate all campaign. The honest caveat is stated in the ledger: novelty is not value — an instrument can be perfectly original and perfectly useless. Whether this original reading actually improves held-out forecasts is now exactly the final test's question.

Preflight (A0): 2026-07-10 18:45 — load1 2.67/32c (falling vs load15 3.39) · no heavy nasimubd jobs · 36 GiB avail, si/so=0 · all four services active → ALL PASS. Sync: main up-to-date, worktree 0 behind. One capped run (0.35 min), readonly=2 loader, deterministic.

M3 census (rbo_m3_census.py, 123/123 subjects)

Check	Reading	Bar	Ruling
k_v incumbent shadow (DECISIVE — the beat-the-incumbent flag)	ρ = 0.019 — near-total orthogonality (prior candidates: −0.35 to −0.37 vs the same k_v; tie-attenuation noted honestly, but this is a genuine outlier in the orthogonal direction)	\|ρ\| < 0.95	PASS — decisively
p dial re-verification	0.997 / 0.997	≥ 0.90	PASS
Full stack grid (21 slots incl. MMD + W1) + cousins + corpses	worst \|ρ\| = 0.629 (vs eicp_n_acc) — every prior M3 survivor's worst was ≥ 0.84; MMD 0.439 · W1 0.375 · calgap 0.283 · redundancy 0.043	\|ρ\| < 0.95	PASS — cleanest clearance yet
Degeneracy	Readout spans three decades (1.7×10⁻⁵ / 4.4×10⁻⁴ / 1.3×10⁻²) — top-weighted by design, not degenerate	—	PASS
Verdict (row 91)	M3 CLEAR — the campaign's most orthogonal survivor heads to the finale. The ledger states the caveat out loud: novelty ≠ value; M4 decides.

Check

Reading

Bar

Ruling

k_v incumbent shadow (DECISIVE — the beat-the-incumbent flag)

ρ = 0.019 — near-total orthogonality (prior candidates: −0.35 to −0.37 vs the same k_v; tie-attenuation noted honestly, but this is a genuine outlier in the orthogonal direction)

|ρ| < 0.95

PASS — decisively

p dial re-verification

0.997 / 0.997

≥ 0.90

PASS

Full stack grid (21 slots incl. MMD + W1) + cousins + corpses

worst |ρ| = 0.629 (vs eicp_n_acc) — every prior M3 survivor's worst was ≥ 0.84; MMD 0.439 · W1 0.375 · calgap 0.283 · redundancy 0.043

|ρ| < 0.95

PASS — cleanest clearance yet

Degeneracy

Readout spans three decades (1.7×10⁻⁵ / 4.4×10⁻⁴ / 1.3×10⁻²) — top-weighted by design, not degenerate

—

PASS

Verdict (row 91)

M3 CLEAR — the campaign's most orthogonal survivor heads to the finale. The ledger states the caveat out loud: novelty ≠ value; M4 decides.

▶ Next iteration

Next firing: H-070 RBO M4/M5′ finale on the sealed forex apparatus: expanding-window RBO column (per feature, per epoch t: the RBO trajectory statistic over transitions within windows 0..t only — leakage-safe by construction, parameter p=0.95 census-verified), 60% epoch split, frozen disjunction (M4 ΔAUC CI-low > 0 OR M5′b), inline era attack (the row-86 procedure) and the row-80 disclosure standard (pairing-null-band comparison stated whatever it is). ADMIT → certificate #15 and the FRAGILE↔CONDITIONAL boundary gains its most orthogonal instrument; EXCLUDE → BONEYARD (novel-but-valueless, a new kill mechanism for the record) and B-04 advances to H-044, then H-072 closes the batch. Standing for the operator: the M4-bar question (rows 80/86), the H-062 proposal, the inverted-mechanism investigation (6 corroborations), the 73 cycle-1 drafts, the Frontier #4 review, the STABLE-declaration proposal, and PR #599 — B-04 now 2 admits · 7 kills · 1 finalist.

Iteration 80 · 2026-07-10 · CHECKPOINT (row 91) — H-070 M3 CLEAR · capped (0.35 min) · readonly=2 · zero generated values · append-only · evidence: rbo_m3_census.py · rbo_m3_census_results.json

Iteration 80 — H-070 RBO M3 CLEAR k_v SHADOW 0.019 · MOST ORTHOGONAL SURVIVOR · FINALE NEXT

M3 census (rbo_m3_census.py, 123/123 subjects)

▶ Next iteration

M3 census (`rbo_m3_census.py`, 123/123 subjects)