Iteration 5 — Ruling A Recorded · C4 Replay 9/9

Iteration 5 — Ruling A Recorded · C4 Replay 9/9 SEAL COMPLETE LAB OPEN

What happened in this iteration, in plain words: the operator ruled on the one open question — the dead-weight trick counts as passed, because the gate did exactly what the frozen rulebook demanded (it refused the candidate), and the tiny whisker of real information we found in that candidate is now logged as a small discovery rather than a failure. That unlocked the fourth and final trick: we re-judged all nine historical candidates from June through the frozen six-gate ladder, deciding each gate from first principles rather than copying the old answers. Result: nine out of nine verdicts reproduced — the same seven rejections and the same two scoped conditional admissions, with one candidate dying at an earlier (cheaper) gate than before, for a reason we can state precisely. With all four controls green, the certification lab has now proven it can catch fakes, refuse dead weight, recognize the real thing, and reproduce its own history. The lab is open. Real candidates start next iteration.

Preflight (A0)

2026-07-03 08:46 UTC — load1 3.09 / 32 cores (≤24) · no competing nasimubd jobs · 21 GiB available · clickhouse-server active · sidecar active + /health healthy → ALL PASS. This iteration is paper/analysis only: zero ClickHouse, zero compute runs, zero synthetic data.

Part 1 — the operator ruling (LEDGER row 6)

Item	Value
Ruling (verbatim)	"My recommendation would be option A, as it passes the pre-registered citation. So yeah, I would say you can go with option A."
Effect	C2′ final verdict = PASS on the amendment-pre-registered criterion (gate EXCLUDEd `bootci_w` at M4, ΔAUC CI ∋ 0). The aux-leg whisker (+0.004, p=0.025 over 250 permutation refits) recorded as a first-class small discovery in BONEYARD: bootci_w remains INPUT-ONLY/excluded; do-not-re-propose unless +0.004 AUC ever becomes decision-relevant AND a native null is found.
Supersedes	LEDGER row 4 (the pending-ruling row)

Item

Value

Ruling (verbatim)

"My recommendation would be option A, as it passes the pre-registered citation. So yeah, I would say you can go with option A."

Effect

C2′ final verdict = PASS on the amendment-pre-registered criterion (gate EXCLUDEd bootci_w at M4, ΔAUC CI ∋ 0). The aux-leg whisker (+0.004, p=0.025 over 250 permutation refits) recorded as a first-class small discovery in BONEYARD: bootci_w remains INPUT-ONLY/excluded; do-not-re-propose unless +0.004 AUC ever becomes decision-relevant AND a native null is found.

Supersedes

LEDGER row 4 (the pending-ruling row)

Part 2 — C4: the nine-candidate replay (LEDGER row 7)

METHOD: each candidate walked through M0→M5 IN ORDER (early-exit), each gate decided from its own frozen criterion; June's answers consulted only afterward. Full walk: C4-REGRESSION-FIXTURE.md (audit folder) # candidate replay June 2026-06-16 verdict gate 1 RCIT / FastKCI EXCLUDE @ M3 EXCLUDE @ M3 ✓ ✓ 2 Anchor Regression / DRIG EXCLUDE @ M3 EXCLUDE @ M3 ✓ ✓ 3 CPSS stability selection EXCLUDE @ M3 EXCLUDE @ M3 ✓ ✓ 4 CPCV purged combinatorial EXCLUDE @ M1+M3 EXCLUDE @ M1+M3 ✓ ✓ 5 Deflated Sharpe (proper) EXCLUDE @ M0 EXCLUDE @ M3/inert ✓ deviation 6 Failing-Loudly EXCLUDE @ M3 EXCLUDE @ M3 ✓ ✓ 7 MP/permutation threshold EXCLUDE @ M2 EXCLUDE @ M2 ✓ ✓ 8 F014 nonlinear/relaxed ICP INCLUDE-IF INCLUDE-IF ✓ ✓ 9 e-BH + CTM (+e-detectors) INCLUDE-IF INCLUDE-IF ✓ ✓ SCORE: 9/9 verdicts · 7 EXCLUDE / 2 INCLUDE-IF reproduced exactly THE ONE DEVIATION (explained in writing, per section 6): DSR-proper now dies at M0 (OFF-GOAL) instead of June's M3 — the frozen label scope (section 1: no return/money axis) formalized what June applied ad hoc, so the cheapest gate now catches it first. Same verdict, same consequence, ladder working as designed.

Seal scoreboard — final

Control	Design (real-data-only, amendment 03c)	Result	LEDGER
C1′ duplicate detector	Exact renamed copy of admitted G0's real outputs	PASS — ρ=1.000000, EXCLUDEd at M3	row 3 (supersedes 1)
C2′ dead-weight detector	Real reject-listed gauge `bootci_w` through M4	PASS — refused, CI ∋ 0; whisker logged (ruling A)	rows 4+6 (supersede 2)
C3′ power	Leave-one-out: hidden `redundancy` resubmitted	PASS — ADMITted, ΔAUC +0.663, CI > 0	row 5
C4 replication	Nine 2026-06-16 candidates through the frozen ladder	PASS — 9/9, one gate deviation explained	row 7

Control

Design (real-data-only, amendment 03c)

Result

LEDGER

C1′ duplicate detector

Exact renamed copy of admitted G0's real outputs

PASS — ρ=1.000000, EXCLUDEd at M3

row 3 (supersedes 1)

C2′ dead-weight detector

Real reject-listed gauge bootci_w through M4

PASS — refused, CI ∋ 0; whisker logged (ruling A)

rows 4+6 (supersede 2)

C3′ power

Leave-one-out: hidden redundancy resubmitted

PASS — ADMITted, ΔAUC +0.663, CI > 0

row 5

C4 replication

Nine 2026-06-16 candidates through the frozen ladder

PASS — 9/9, one gate deviation explained

row 7

▶ Next iteration

Iteration 6 = Frontier #1: fix the forex blackout (first REAL work item). In plain words: half the goal — grading forex features — has been structurally impossible because the regime-invariance instrument silently skipped every forex environment (it was looking for five crypto column names on a forex table). The fix is fully planned: per-market anchor columns, thread the market through ICP and the granger echo, and make all-environments-skipped raise loudly instead of returning an empty page. Then the acceptance test: a capped, read-only Tier-A EURUSD@5 re-run (pre-authorized under directive 03b — readonly=2, 5-core/5-GB, spare-only) that must produce a non-empty forex ICP map, with nulls still collapsing and crypto results bit-identical. Bounded slice if the re-run is long; checkpoint ledger row; the campaign PR accumulates.

Iteration 5 · 2026-07-03 · operator ruling + C4 replay · paper/analysis only · zero compute · zero synthetic data · append-only