Iteration 2 — Self-Test Seal 1/4: Planted Duplicate Caught (C1 PASS)

Iteration 2 — Self-Test Seal 1/4: Planted Duplicate Caught C1 PASS

What happened in this iteration, in plain words: before the lab is allowed to judge any real instrument, the judge itself gets tested with trick candidates whose correct answers we already know. Today was trick #1: we photocopied an already-trusted instrument, gave the copy an impressive new name ("Monotone Association Apex Index"), added a whisper of noise so it isn't byte-identical, and submitted it as if it were a new discovery. The lab caught it — it sailed through the paper interviews (as a good fake should) and was thrown out at exactly the right gate, the redundancy attack, with the exact reading a perfect duplicate must produce (correlation 1.000). Just as important: the same detector did not throw out things that genuinely differ. One trick down, three to go.

Preflight (A0)

2026-07-03 07:16 UTC — load1 2.85 / 32 cores (≤24, flat vs load15 3.26) · no competing nasimubd jobs · 21 GiB available, si/so=0 · clickhouse-server active · sidecar unit active + /health = healthy (checked via the repo's own heartbeat endpoint, scripts/heartbeat/checks/sidecar.py) → ALL PASS

First-firing bookkeeping (procedure G)

Ledger row	Content
0a (NOTE)	Pre-registration hash recorded: `f959424012d0b54a8984c7fbba00cdd4c2c31615` (frozen rulebook, PR #574 landing commit)
0b (NOTE)	Operator resource directive 2026-07-03 recorded: ClickHouse `readonly=2` SELECT-only · hard cap 5 CPU / 5 GB / swap banned via `systemd-run --user --scope` + `nice -n 19 ionice -c3` + single-thread BLAS · spare-only + watchdog (load1 > 24 / CH inactive / sidecar unhealthy ⇒ abort) · operator go required for any real-data compute run

Ledger row

Content

0a (NOTE)

Pre-registration hash recorded: f959424012d0b54a8984c7fbba00cdd4c2c31615 (frozen rulebook, PR #574 landing commit)

0b (NOTE)

Operator resource directive 2026-07-03 recorded: ClickHouse readonly=2 SELECT-only · hard cap 5 CPU / 5 GB / swap banned via systemd-run --user --scope + nice -n 19 ionice -c3 + single-thread BLAS · spare-only + watchdog (load1 > 24 / CH inactive / sidecar unhealthy ⇒ abort) · operator go required for any real-data compute run

The trick candidate and how it died

PLANT: "Monotone Association Apex Index" (MAAI) = admitted gate G0 (per-cell Spearman ρ), renamed, output + ε jitter cells = committed real-data artifact spike_2026-05-26/per_slice.jsonl (BTCUSDT@250 · 8 regime slices × 5 features = 40 cells · sha256-pinned) │ M0 status-linkage PASS ← claims the FAIL-line boundary (same as G0) — a fake presents well M1 trust-anchor PASS ← claims G0's own ground-truth anchor M2 substrate PASS ← same closed-form rank statistic as G0 │ M3 REDUNDANCY ATTACK (the gate that must kill it): L1 exact clone (ε=0) ρ vs G0 = 1.000000 ≥ 0.95 → FLAGGED ✓ ground truth: duplicate reads 1.0 L2 jittered clone (ε=1e-3) ρ vs G0 = 0.999625 ≥ 0.95 → FLAGGED ✓ rename+noise does not fool it L3 plant removed (ε=10) ρ vs G0 = −0.135835 < 0.95 → not flagged ✓ detection collapses w/o the plant L4 specificity ρ(granger_min_p, max|ρ|) = −0.023810 < 0.95 → not flagged ✓ doesn't kill everything ▼ VERDICT for the plant: EXCLUDE(DUPLICATE-GAUGE) at M3 VERDICT for the control: C1 PASS — the redundancy attack works

Technical record

Item	Value
Control	C1 planted duplicate (SELF-TEST-CONTROLS.md) — ground-truth anchor; must EXCLUDE at M3 (DUPLICATE-GAUGE)
Harness	`c1_planted_duplicate.py` (audit folder) — stdlib-only, deterministic (seed 20260703), tie-safe average-rank Spearman
Input	`spike_2026-05-26/per_slice.jsonl` (sibling audit, committed) — sha256 pinned in `c1_result.json`; hermetic: no ClickHouse, no network, no new real-data run (re-analysis of a committed artifact ⇒ no operator-go required; flagged for operator review)
Resource cap	run inside `systemd-run --user --scope -p CPUQuota=500% -p MemoryMax=5G -p MemorySwapMax=0 nice -n 19 ionice -c3`, single-thread BLAS env — per operator directive (trivial compute, <1s)
Legs	L1 ρ=1.000000 flagged · L2 ρ=0.999625 flagged · L3 ρ=−0.135835 not flagged · L4 ρ=−0.023810 (n=8, independently hand-verified −0.0238) not flagged — all 4 as required
Parameter tags (§0 guard)	`0.95`=FROZEN (M3 spec, ground-truth-anchored production legacy) · `ε=1e-3`=MAGIC-NUMBER (test fixture; rank-preserving, plateau-stable vs inter-cell spread ~0.2) · `ε=10`=MAGIC-NUMBER (rank-destroying fixture) · seed=STRUCTURAL · n=40 cells=DERIVED (all cells in artifact)
Skeptic arm (M3)	MAAI is definitionally G0's formula under a new name: no new failure mode, identical (not tighter) guarantee ⇒ redundant by construction; empirical arm confirms at the frozen 0.95 line
Ledger rows appended	0a (NOTE, pre-reg hash) · 0b (NOTE, resource directive) · 1 (SELF-TEST C1 PASS)
Prompt SSoT reconciliation	The operator's launched prompt contains an A0 PREFLIGHT section that was missing from `LOOP-PROMPT.md`'s PROMPT block; folded in verbatim this firing (dated amendment note) so the SSoT matches the running loop. Strictly additive read-only safety checks.
Compute / prod impact	<1 s single-core inside the 5c/5G cap · zero ClickHouse access · zero writes outside the audit folder + dashboard

Item

Value

Control

C1 planted duplicate (SELF-TEST-CONTROLS.md) — ground-truth anchor; must EXCLUDE at M3 (DUPLICATE-GAUGE)

Harness

c1_planted_duplicate.py (audit folder) — stdlib-only, deterministic (seed 20260703), tie-safe average-rank Spearman

Input

spike_2026-05-26/per_slice.jsonl (sibling audit, committed) — sha256 pinned in c1_result.json; hermetic: no ClickHouse, no network, no new real-data run (re-analysis of a committed artifact ⇒ no operator-go required; flagged for operator review)

Resource cap

run inside systemd-run --user --scope -p CPUQuota=500% -p MemoryMax=5G -p MemorySwapMax=0 nice -n 19 ionice -c3, single-thread BLAS env — per operator directive (trivial compute, <1s)

Legs

L1 ρ=1.000000 flagged · L2 ρ=0.999625 flagged · L3 ρ=−0.135835 not flagged · L4 ρ=−0.023810 (n=8, independently hand-verified −0.0238) not flagged — all 4 as required

Parameter tags (§0 guard)

0.95=FROZEN (M3 spec, ground-truth-anchored production legacy) · ε=1e-3=MAGIC-NUMBER (test fixture; rank-preserving, plateau-stable vs inter-cell spread ~0.2) · ε=10=MAGIC-NUMBER (rank-destroying fixture) · seed=STRUCTURAL · n=40 cells=DERIVED (all cells in artifact)

Skeptic arm (M3)

MAAI is definitionally G0's formula under a new name: no new failure mode, identical (not tighter) guarantee ⇒ redundant by construction; empirical arm confirms at the frozen 0.95 line

Ledger rows appended

0a (NOTE, pre-reg hash) · 0b (NOTE, resource directive) · 1 (SELF-TEST C1 PASS)

Prompt SSoT reconciliation

The operator's launched prompt contains an A0 PREFLIGHT section that was missing from LOOP-PROMPT.md's PROMPT block; folded in verbatim this firing (dated amendment note) so the SSoT matches the running loop. Strictly additive read-only safety checks.

Compute / prod impact

<1 s single-core inside the 5c/5G cap · zero ClickHouse access · zero writes outside the audit folder + dashboard

▶ Next iteration

Iteration 3 — self-test seal 2/4: the noise gauge (C2). In plain words: the second trick candidate is a dial that emits pure random numbers dressed up as a metric. A trustworthy lab must reject it as dead weight — adding it to the trusted stack must measurably improve nothing (its incremental-value confidence interval must straddle zero). Technically: build the C2 harness (random per-cell outputs, correct shape, seeded) and drive it through M4's ΔAUC rule. If C2 can be computed hermetically from committed walk-forward artifacts it runs and its ledger row lands next firing; if it turns out to require a capped read-only run against real ClickHouse data, the exact command + preflight are published here as a PROPOSAL and the item is marked PARKED-AWAITING-GO instead (operator types the go into the loop session). After C2: C3 planted signal (must ADMIT), C4 historical replay (must reproduce 7 EXCLUDE / 2 INCLUDE-IF). The lab stays closed to real candidates until all four pass.

Iteration 2 · 2026-07-03 · self-test seal 1/4 · hermetic re-analysis of committed artifacts · read-only · append-only