›Navigation

Iteration 4 — Seal Re-validation (No-Synthetic) C1′ PASS C3′ PASS C2′ AWAITING RULING

What happened in this iteration, in plain words: under the operator's new rule — no synthetic data anywhere — all seal tricks were re-run using only real market-data-derived numbers. Trick 1 (the disguised duplicate) passed perfectly: an exact renamed copy of a trusted instrument reads correlation 1.000 and gets thrown out. Trick 3 (the "can the lab recognize the real thing" test, redesigned without synthetic signals) passed spectacularly: we hid the lab's strongest real instrument, resubmitted it as a fresh candidate, and the lab admitted it with an unmistakable skill jump (0.31 → 0.98). Trick 2 produced this campaign's first genuinely interesting surprise: the real reject-listed instrument we used as the "adds nothing" candidate WAS correctly refused by the gate — but a deeper diagnostic shows it carries a microscopic whisker of real information (four thousandths of a point, statistically detectable at p=0.025). The gate did its job; our assumption that the candidate was perfectly information-free was slightly wrong. That's the known cost of banning constructed test candidates: real ones can surprise you. Per the rulebook, the loop hard-stops this control and asks the operator to rule.

Preflight (A0)

2026-07-03 08:19 UTC — load1 3.79 / 32 cores (≤24, flat vs load15 4.06) · no competing nasimubd jobs · 22 GiB available, si/so≈0 · clickhouse-server active · sidecar active + /health healthy → ALL PASS. All runs inside the 5c/5G systemd-run scope; zero ClickHouse (real panel = the committed campaign's on-disk cell cache); zero synthetic values anywhere.

Results at a glance

ControlCandidate (all real data)ExpectedMeasuredVerdict
C1′ duplicate detectorExact renamed copy of admitted G0's real outputs (+ 6dp-rounding leg)EXCLUDE at M3, ρ=1.0ρ=1.000000 exact · 1.000000 rounded · specificity −0.0238 cleanPASS (supersedes LEDGER row 1)
C2′ dead-weight detectorReal reject-listed gauge bootci_w (boneyard 2026-06-16: rank-derived, INPUT-ONLY)EXCLUDE at M4, ΔAUC CI ∋ 0ΔAUC +0.0040, CI [−0.0026,+0.0115] ∋ 0 → gate refused it ✓ · BUT aux null-leg: obs above pairing-null band, post-hoc p=0.025 (200 draws)Primary criterion PASS · harness conjunction FAIL → OPERATOR RULING (supersedes row 2 pending ruling)
C3′ power (replaces synthetic planted signal)Leave-one-out: hide redundancy (committed AUC 0.945), resubmit as candidate vs {adv_auc, granger} baselineADMIT, ΔAUC CI > 0AUC 0.3126 → 0.9753; ΔAUC +0.6626, CI [+0.4275,+0.8358]; null band [−0.047,+0.194] — lift collapses when pairing destroyedPASS (LEDGER row 5)

The C2′ surprise, dissected

QUESTION the control asked : does the gate refuse a candidate that adds nothing?
WHAT THE GATE DID          : refused bootci_w — ΔAUC CI [−0.0026, +0.0115] contains 0
                             → no measurable improvement → EXCLUDE(DEAD-WEIGHT)  ✓ correct
WHAT THE AUX LEG FOUND     : the +0.0040 whisker is NOT pairing noise:
                             verdict run    (50 null refits): band [−0.0109, +0.0036], obs outside
                             post-hoc diag (200 disjoint refits): band [−0.0198, +0.0036], obs outside
                             one-sided permutation p = 0.025
WHAT THAT MEANS            : bootci_w carries a REAL but sub-bar sliver of label information
                             (+0.004 on a 0.9753 baseline). Its committed grounding said
                             NULL-INVARIANT → INPUT-ONLY ("cannot be a gate"), which remains
                             true — it never claimed literally zero label information.
READING                    : genuine small discovery, NOT a broken gate — precisely the
                             epistemic cost M0-AMENDMENT-03c pre-declared for real-data
                             controls (real candidates can surprise; constructed ones cannot).
ACTION PER HARD RULE 1     : hard-stop this control · publish this report · PARK C4 ·
                             operator rules which reading holds.
Operator ruling requested (the loop is parked on this):
Option A — accept the primary criterion (recommended): C2′ counts as PASS because the amendment pre-registered exactly one expectation ("EXCLUDE at M4, ΔAUC CI ∋ 0") and the gate delivered it; the aux-leg result is recorded as a first-class small discovery (bootci_w: weakly informative, still INPUT-ONLY, still excluded). Seal proceeds to C4.
Option B — strict conjunction: C2′ counts as FAIL; a superseding pre-registration must redesign the dead-weight control (knowing that no real candidate can be construction-certain information-free — the directive's inherent limit). Seal re-runs C2″.
Type your ruling into the loop session; it will be recorded verbatim in the LEDGER.

Technical record

ItemValue
Amendment in forceM0-AMENDMENT-2026-07-03c-NO-SYNTHETIC-DATA.md (LEDGER 0d) — no generated values; shuffles/bootstraps of real values permitted (§0 doctrine)
Harnessesc1_prime_exact_clone.py · c2c3_prime_real_data_m4.py · c2_prime_null_diagnostic.py (post-hoc, not verdict-bearing) — all stdlib-only, deterministic, capped scope
Panelreal campaign cells ~/tierA_v3/cells, manifest sha 2ed074c9…; 2992 crypto transitions / 803 among-orth (= committed e123_v3 exactly); committed T1–T4 leakage gate PASS
Sync notemain moved (PR #576 landed A0 into LOOP-PROMPT.md upstream); loop branch rebased cleanly, single A0 block verified
Ledger rows3 (C1′ PASS, supersedes 1) · 4 (C2′ pending ruling, supersedes 2) · 5 (C3′ PASS)
Synthetic data usedNONE — exact copies, deterministic rounding, real gauge columns, permutations/bootstraps of real values only
Compute / prod impact~8 min total single-core inside the 5c/5G scope · zero ClickHouse · zero writes outside audit folder + dashboard

▶ Next iteration

Iteration 4 · 2026-07-03 · seal re-validation under no-synthetic directive · real data only · capped · read-only · append-only