Iteration 3 — Self-Test Seal 2/4: Noise Gauge Rejected as Dead Weight (C2 PASS)

Iteration 3 — Self-Test Seal 2/4: Noise Gauge Rejected as Dead Weight C2 PASS

What happened in this iteration, in plain words: trick candidate #2. We built a fake instrument that emits pure random numbers — right shape, plausible look, zero information — and submitted it five times (five different random draws, so one lucky draw can't fool us). Each time, the lab measured whether adding the fake to the trusted instrument panel improved its forecasts of "which features stay orthogonal next regime." Verdict all five times: improves nothing — thrown out as dead weight. Crucially, we also proved the test isn't a rubber stamp: a secretly-perfect instrument fed through the same machinery WAS detected as genuinely valuable. And unlike trick #1 (which used a tiny 40-cell artifact), this one ran on the real campaign panel — 2,992 real regime transitions across 21 market/threshold combos, byte-verified to match the numbers of the original frozen campaign. Two tricks down, two to go.

Preflight (A0)

2026-07-03 07:55 UTC — load1 2.49 / 32 cores (≤24, flat vs load15 3.07) · no competing nasimubd jobs · 23 GiB available, si/so=0 · clickhouse-server active · sidecar unit active + /health = healthy → ALL PASS. Run executed inside the 5-core/5-GB systemd-run scope (operator directive; spikes pre-authorized per 2026-07-03b — no ClickHouse was needed anyway: the real panel is the on-disk ~/tierA_v3/cells cache of the committed campaign).

The trick candidate and how it died

PLANT: "noise gauge" = U(0,1) random output per (cell, feature) — correct shape, zero information x 5 independent seeds (a single lucky draw cannot pass) PANEL: real campaign walk-forward panel, ~/tierA_v3/cells (sha-pinned manifest) n = 2,992 crypto transitions · 803 among-orthogonal-at-t ← replicates the committed e123_v3 numbers EXACTLY (2992 / 803) — same data, same label (TAU=0.85) │ M3 passthrough check: max|rho|(noise, admitted) ≤ 0.110 < 0.95 │ (the plant legitimately reaches M4 — not killed early) M4 INCREMENTAL VALUE (the gate that must kill it): baseline = logistic(redundancy, adv_auc, granger_echo_p) [wold_R suspended] fit on cal epochs (first 60%), applied ONCE to test epochs conditional AUC among orth_t on test: 0.9753 (n=256) augmented = baseline + noise column, same protocol ΔAUC 95% cluster-bootstrap CI (by combo, B=400): seed 1 −0.0112 [−0.0242, +0.0009] contains 0 ✓ seed 2 −0.0013 [−0.0082, +0.0086] contains 0 ✓ seed 3 −0.0112 [−0.0272, +0.0023] contains 0 ✓ seed 4 +0.0000 [ 0.0000, 0.0000] contains 0 ✓ (degenerate: logistic gave the noise so little weight the test ranking never changed — purest dead weight) seed 5 −0.0036 [−0.0227, +0.0079] contains 0 ✓ shuffled-pairing null (refit per permutation): observed ΔAUC inside band, all seeds ✓ │ ANTI-VACUITY GUARD (is "CI contains 0" a rubber stamp?): planted PERFECT gauge (the true label + eps) through the SAME machinery → ΔAUC CI [+0.0008, +0.0512] — entirely > 0 → machinery CAN reject ✓ ▼ VERDICT for the plant: EXCLUDE(DEAD-WEIGHT) at M4 — all 5 seeds VERDICT for the control: C2 PASS — the incremental-value gate works

Technical record

Item	Value
Control	C2 noise gauge (SELF-TEST-CONTROLS.md) — matched-null anchor; must EXCLUDE at M4 (DEAD-WEIGHT), ΔAUC CI must contain 0
Harness	`c2_noise_gauge.py` (audit folder) — stdlib-only, deterministic; mirrors the frozen `e123_v3_walk_forward.py` machinery 1:1 (transition construction, TAU=0.85 label, rank AUC, cluster bootstrap by combo B=400, E2-convention temporal 60/40 split)
Panel	`~/tierA_v3/cells` (383 per-cell JSONs, the committed campaign's own compute cache; manifest sha256 `2ed074c9…` pinned in result). Panel replication verified against two committed invariants: crypto n_transitions 2992 = e123_v3's 2992; among-orth 547+256 = 803 = e123_v3's n_among 803. Committed leakage gate (T1–T4): PASS on record.
Baseline (admitted)	logistic over SIGN-adjusted redundancy + adv_auc + granger_echo_p; wold_R excluded (suspended, reported-not-gated per campaign verdict); cal-median imputation, cal standardization, apply-once to test. Conditional AUC 0.9753 (test among-orth, n=256)
Legs	L0 guard: committed gate PASS + label-shuffle null straddles 0.5 ✓ · L1 dead-weight: all 5 seeds CI ∋ 0 ✓ · L2 pairing-null: observed inside refit-null band, all seeds ✓ · L3 M3-passthrough: max\|ρ\| ≤ 0.110 ✓ · L4 anti-vacuity: perfect gauge CI [+0.0008,+0.0512] > 0 ✓
Honest notes	(1) seed 4's degenerate [0,0] CI is the strongest possible dead-weight reading, not a defect — its pairing-null band is non-degenerate, proving the pipeline moves when the pairing moves. (2) L4's lower bound (+0.0008) is thin because the baseline already sits at 0.9753 — headroom to 1.0 is small; the leg still separates signal from noise cleanly. (3) The noise ΔAUCs skew slightly negative (noise can only dilute a strong baseline) — "no measurable improvement" holds a fortiori.
Parameter tags (§0 guard)	TAU=0.85 FROZEN · 60/40 split DERIVED (frozen E2) · B=400/200 DERIVED (e123_v3) · admitted set DERIVED (certified minus suspended) · logistic LR/ITERS/L2 MAGIC-NUMBER (fit fixture, plateau-stable) · U(0,1)×5 seeds STRUCTURAL · perfect-gauge ε=0.05 MAGIC-NUMBER · seed base STRUCTURAL
Ledger row	2 (SELF-TEST C2 PASS)
Compute / prod impact	~2 min single-core inside the 5c/5G scope · zero ClickHouse access · zero writes outside the audit folder + dashboard

Item

Value

Control

C2 noise gauge (SELF-TEST-CONTROLS.md) — matched-null anchor; must EXCLUDE at M4 (DEAD-WEIGHT), ΔAUC CI must contain 0

Harness

c2_noise_gauge.py (audit folder) — stdlib-only, deterministic; mirrors the frozen e123_v3_walk_forward.py machinery 1:1 (transition construction, TAU=0.85 label, rank AUC, cluster bootstrap by combo B=400, E2-convention temporal 60/40 split)

Panel

~/tierA_v3/cells (383 per-cell JSONs, the committed campaign's own compute cache; manifest sha256 2ed074c9… pinned in result). Panel replication verified against two committed invariants: crypto n_transitions 2992 = e123_v3's 2992; among-orth 547+256 = 803 = e123_v3's n_among 803. Committed leakage gate (T1–T4): PASS on record.

Baseline (admitted)

logistic over SIGN-adjusted redundancy + adv_auc + granger_echo_p; wold_R excluded (suspended, reported-not-gated per campaign verdict); cal-median imputation, cal standardization, apply-once to test. Conditional AUC 0.9753 (test among-orth, n=256)

Legs

L0 guard: committed gate PASS + label-shuffle null straddles 0.5 ✓ · L1 dead-weight: all 5 seeds CI ∋ 0 ✓ · L2 pairing-null: observed inside refit-null band, all seeds ✓ · L3 M3-passthrough: max|ρ| ≤ 0.110 ✓ · L4 anti-vacuity: perfect gauge CI [+0.0008,+0.0512] > 0 ✓

Honest notes

(1) seed 4's degenerate [0,0] CI is the strongest possible dead-weight reading, not a defect — its pairing-null band is non-degenerate, proving the pipeline moves when the pairing moves. (2) L4's lower bound (+0.0008) is thin because the baseline already sits at 0.9753 — headroom to 1.0 is small; the leg still separates signal from noise cleanly. (3) The noise ΔAUCs skew slightly negative (noise can only dilute a strong baseline) — "no measurable improvement" holds a fortiori.

Parameter tags (§0 guard)

TAU=0.85 FROZEN · 60/40 split DERIVED (frozen E2) · B=400/200 DERIVED (e123_v3) · admitted set DERIVED (certified minus suspended) · logistic LR/ITERS/L2 MAGIC-NUMBER (fit fixture, plateau-stable) · U(0,1)×5 seeds STRUCTURAL · perfect-gauge ε=0.05 MAGIC-NUMBER · seed base STRUCTURAL

Ledger row

2 (SELF-TEST C2 PASS)

Compute / prod impact

~2 min single-core inside the 5c/5G scope · zero ClickHouse access · zero writes outside the audit folder + dashboard

▶ Next iteration

Iteration 4 — self-test seal 3/4: the planted signal (C3, must be ADMITTED). In plain words: the first two tricks proved the lab can reject fakes. The third trick proves it can recognize the real thing: we secretly build a gauge that genuinely carries information about which features stay orthogonal — with a pre-registered signal strength, injected before anyone looks at outcomes — and submit it. The lab must admit it. If it gets rejected, the lab is blind, and every rejection it ever issues is meaningless. Technically: synthetic candidate carrying real information about the frozen label at pre-registered severity (severity written down and tagged BEFORE the run, per the §0 magic-number guard), driven through the same M4 machinery on the same real panel; must clear the ΔAUC CI > 0 bar and collapse under the shuffled-pairing null removal test. The severity grid will bracket the interesting regime (weak/medium/strong) with the ADMIT bar set at the pre-registered primary severity only. Then C4: replay of the nine 2026-06-16 candidates through the frozen ladder (must reproduce 7 EXCLUDE / 2 INCLUDE-IF). Seal completes → the lab opens; first real item = Frontier #1 forex active-set fix (Tier-A re-run pre-authorized, capped read-only).

Iteration 3 · 2026-07-03 · self-test seal 2/4 · real campaign panel, hermetic (no ClickHouse) · capped scope · read-only · append-only