Iteration 40 — H-017 StabReg EXCLUDED: κ Resolved in Its Favor, Then Killed as an Oracle Shadow

Iteration 40 — H-017 StabReg EXCLUDED KILL — M3 REDUNDANT (ρ=0.977) κ FIGHT RESOLVED FIRST

In plain words: the team-judging candidate got the fairest possible trial — and the trial produced the cleanest kind of kill. First, the tuning-dial worry from last time dissolved: sweeping the dial across its whole sensible range showed the answers barely move (a plateau), and deriving the dial's value from first principles (how noisy the prediction-quality estimate itself is) lands almost exactly on the value the exam had guessed. The dial was innocent. Then came the killer question: does this candidate tell us anything our existing instruments don't? No. Its central reading — how many feature-teams survive — moves in near-lockstep (97.7% rank agreement) with the reading of an instrument we already certified. The reason is structural, and we had named it in advance: this candidate's "steadiness" test is literally built on the certified instrument's own engine, so its verdicts are that instrument's shadow. One instrument's shelf space, zero new information: rejected, with the reason on record so nobody re-proposes it unchanged. The queue advances to the next candidate.

Preflight (A0): 2026-07-10 00:25–00:34 — load1 2.14→3.37/32c (≤24) · no heavy nasimubd jobs · 33 GiB avail, si/so=0 · CH active · sidecar+kintsugi RESTORED by the operator (~00:07/00:23; kintsugi first pass "no shards found"; the 00:20 heartbeat CRITICAL predates the 00:23 restart) → GREEN — the loop un-parks. Firing 39 (row 50) was the pre-registered one-line PARKED checkpoint; this firing is iteration 40. One capped run: 0.98 min, 5 workers, systemd-run --scope 5-core/5-GB/no-swap, readonly=2 loader, watchdog clean (0 trips), permutations of real rows only.

One run, two fights (pre-registered rule in stabreg_kappa_sweep_m3.py)

Key insight that made it cheap: e_total and oos_mse per feature-set are κ-independent — the knob only enters at survivor selection. So one 130-subject × 8-set grid computation (identical machinery to the entrance exam, per-fold MSEs retained) supports the entire κ-sweep and the M3 census post-hoc.

Fight	Reading	Ruling
κ knob (§0: derive or show insensitive)	Derived κ_D = 0.010364 — the median (over 130 subjects) of the estimation-noise SE of the best set's fold MSEs. The 0.01 exam fixture sat almost exactly on the derived value. Plateau: adjacent-κ readout Spearman 0.948–1.0 across the ENTIRE sweep [0 → 0.5]; local insensitivity at κ_D/2..2κ_D: median Δn_surv = 0. Not degenerate: trivial-family rate ≤ 0.846 < 0.95 (at κ_D: 69 empty · 38 all-8 · 23 non-trivial).	RESOLVED — dial derived AND shown insensitive; the iter-37 red ink dissolves
M3 redundancy attack (kill bar \|ρ\| ≥ 0.95)	Readout 1, n_surv: vs e-ICP-fp n_acc ρ = 0.977 ✗ (vs eicp_log_min_e −0.877 · ipp_n_acc 0.641 · seqicp 0.595 · ias 0.118 · k_v 0.211 · mean-redundancy −0.315). Readout 2, w_sum: worst 0.804 — same instrument. Rule kills on worst across readouts.	FAIL — 0.977 ≥ 0.95
Verdict (row 51)	EXCLUDE(redundant — set-level readout repackages e-ICP-fp accepted-set count, ρ=0.977) → BONEYARD. Re-proposal admissible only with a stability screen NOT built on the admitted v5 oracle.

Fight

Reading

Ruling

κ knob (§0: derive or show insensitive)

Derived κ_D = 0.010364 — the median (over 130 subjects) of the estimation-noise SE of the best set's fold MSEs. The 0.01 exam fixture sat almost exactly on the derived value. Plateau: adjacent-κ readout Spearman 0.948–1.0 across the ENTIRE sweep [0 → 0.5]; local insensitivity at κ_D/2..2κ_D: median Δn_surv = 0. Not degenerate: trivial-family rate ≤ 0.846 < 0.95 (at κ_D: 69 empty · 38 all-8 · 23 non-trivial).

RESOLVED — dial derived AND shown insensitive; the iter-37 red ink dissolves

M3 redundancy attack (kill bar |ρ| ≥ 0.95)

Readout 1, n_surv: vs e-ICP-fp n_acc ρ = 0.977 ✗ (vs eicp_log_min_e −0.877 · ipp_n_acc 0.641 · seqicp 0.595 · ias 0.118 · k_v 0.211 · mean-redundancy −0.315). Readout 2, w_sum: worst 0.804 — same instrument. Rule kills on worst across readouts.

FAIL — 0.977 ≥ 0.95

Verdict (row 51)

EXCLUDE(redundant — set-level readout repackages e-ICP-fp accepted-set count, ρ=0.977) → BONEYARD. Re-proposal admissible only with a stability screen NOT built on the admitted v5 oracle.

WHY THE KILL WAS PREDICTED (iter 37, "shared-oracle question"): StabReg survivor = passes STABILITY screen AND passes PREDICTIVENESS screen └── the admitted v5 └── adds too little independent fold-product e-value variation at census scale oracle ITSELF │ ▼ count(surviving sets) ≈ count(e-value-accepted sets) = e-ICP-fp's n_acc → ρ = 0.977

▶ Next iteration

Next firing: B-01 advances to H-021 EILLS (environment-invariant linear least squares) — entrance exam on the byte-identical forex twin anchor per its CANDIDATE-CATALOG dossier (M0/M1 pre-screened by the harvest campaign; M2 risk flags noted there). Cheapest-kill-first; the ground-truth exam is the cheapest kill available. Standing for the operator: the 73 cycle-1 draft statuses (iters 11/13), the Frontier #4 expiry-monitor precondition review, and the ripening STABLE-declaration proposal (five certified forex instruments + validity-clean bedrock + shape-shift warning list).

Iteration 40 · 2026-07-10 · EXCLUDE (row 51, BONEYARD +1) — kills are wins · capped (0.98 min) · readonly=2 · zero synthetic data · append-only · evidence: stabreg_kappa_sweep_m3.py · stabreg_kappa_sweep_m3_results.json · stabreg_kappa_grid.json

Iteration 40 — H-017 StabReg EXCLUDED KILL — M3 REDUNDANT (ρ=0.977) κ FIGHT RESOLVED FIRST

One run, two fights (pre-registered rule in stabreg_kappa_sweep_m3.py)

▶ Next iteration

One run, two fights (pre-registered rule in `stabreg_kappa_sweep_m3.py`)