โ€บNavigation

Iteration 100 โ€” H-055 SKIT NULL MACHINERY CERTIFIED INSTRUMENT EXCLUDED ยท KILL #43

In plain words: the batch finally produced a survivor โ€” and then the ladder did exactly what it exists to do. SKIT is a "betting" detector: it wagers imaginary wealth, step by step, on whether two series move together; its selling point is a safety guarantee that the wealth can't grow past a bound by luck alone, no matter when you stop watching. Every previous candidate's safety certificate broke on market data. SKIT's held โ€” twice over: on genuinely independent data the attack was proven fair (2.7% false alarms), and on the dependence-preserving null data that broke six other certificates it false-alarmed just 0.7%. That certification is banked permanently: it's only the second trustworthy noise floor the campaign has ever certified, and it is exactly the missing part the four killed cousins needed. The census then showed SKIT reads a channel nothing in the stack reads (worst overlap 0.59). And then the admission bar killed it anyway: added to the forward-prediction harness, it doesn't just fail to help โ€” it actively hurts, and its direction is significantly backwards: features whose trading intensity co-moves strongly with bar duration tend to LOSE their orthogonality in the next regime. That inversion is the eighth sighting of a mechanism the campaign has been tracking all month. Kill the instrument, keep the engine, log the physics.

Preflight held from iter 99 (same firing session; load re-checked by in-process watchdog, never tripped). Two capped runs (0.37 + 3.12 min), readonly=2 loader, permutations/shifts of real values (03c). Harness notes: log-domain wealth accumulator (float overflow at 1e308 on concatenated grids) + even-length pairing โ€” recorded, math identical.

The full ladder in one firing (skit_exam.py โ†’ skit_m3_m4_m5b.py)

GateReadingRuling
Entrance E1Wealth path bit-identical across duration twinsPASS
Entrance E2 โ€” two-arm certificate attackARM A (iid control โ€” proves the attack fair): FA 2.7% ยท ARM B (dependent truths): FA 0.7% โ€” the anytime guarantee HOLDS where all six certificate-class members brokePASS โ€” 2nd CERTIFIED NULL MACHINERY (standing datum)
Entrance E3sup-wealth 10ยฒยณโ€“10ยฒโธ vs threshold 20, 3/3 envs; G0-echo 0.18PASS
M3 census (130 subjects)Dial law 0.938 โœ“ ยท worst stack shadow 0.585 (hsicx) ยท k_v 0.068 ยท mmd 0.119 ยท w1 0.065 โ€” a genuinely novel channelCLEAR
M4 (sealed apparatus, frozen disjunction leg 1)ฮ”AUC โˆ’0.0089, CI95 [โˆ’0.0155, โˆ’0.0027] โ€” entirely negative: ACTIVELY HARMFUL (the H-060/RBO mode)FAIL
M5โ€ฒb (leg 2)ฯ = โˆ’0.154, CI [โˆ’0.183, โˆ’0.124], null_hi 0.039 โ€” significantly INVERTED: co-movement with duration โ‡’ orthogonality LOSS next regime โ€” mechanism datum #8FAIL
Verdict (row 114)EXCLUDE(no incremental value at the FAIL boundary) โ€” KILL #43, BONEYARD #58 ยท null machinery certification BANKED ยท re-entry: as the inference layer for the row-113 dcor/ฯ„* re-proposal

The constructive residue โ€” a matched pair on the operator's desk

Row 113 killed four real readings for lacking a valid null. Row 114 certified a valid null attached to a reading that failed admission. The proposal: dcor (the family's independent witness, shadow 0.59) inferenced by the SKIT betting null (certified, ARM B 0.7%) โ€” each supplies exactly what the other lacked. Filed as a proposal; not self-admitted; a future pre-registration would seal it.

โ–ถ Next iteration

Iteration 100 ยท 2026-07-14 ยท VERDICT (row 114) โ€” EXCLUDE ยท kill #43 ยท boneyard #58 ยท null machinery CERTIFIED (standing datum) ยท capped (3.5 min) ยท readonly=2 ยท zero generated values ยท append-only ยท evidence: skit_exam.py ยท skit_m3_m4_m5b.py + results JSONs