Iteration 90 โ H-046 RW EXCLUDED as an Instrument KILL #29 ยท THE BACKBONE STANDS ยท PIPELINE PROPOSAL FILED
In plain words: a kill that must not be misread. As a feature-RANKING instrument, the stepdown's per-feature score died at the dial law: its adjusted significance levels take only thirty distinct values (a consequence of the resampling count), and a coarse, steppy score inevitably reshuffles its ranking when the data volume changes slightly (80.3% agreement against the 90% bar). But the previous iteration's landmark result is untouched: as an ERROR-CONTROL engine, the procedure was flawless โ zero false discoveries in forty-nine worlds where the standard method failed forty-three percent of the time. The resolution of this apparent tension is the iteration's contribution: the declaration layer doesn't need another ranking instrument โ it needs a trustworthy inference backbone, and that is a PIPELINE decision, not a stack admission. A formal proposal now sits with the operator: adopt the rotation-based stepdown as the declaration pipeline's inference layer (consuming no gate slot), with the false-discovery backbone's re-entry and the wild-bootstrap gray-zone ruling folded into the same review.
Preflight (A0): 2026-07-10 22:26 โ load1 3.86/32c (steady) ยท no heavy nasimubd jobs ยท 30 GiB avail, si/soโ0 ยท all four services active โ ALL PASS. Sync: main up-to-date, worktree 0 behind. One capped run (0.38 min), readonly=2 loader, shifts of real values only.
M3 census (rw_m3_census.py, 124/124)
Check
Reading
Ruling
NSUB dial law (THE KILL)
ฯ(adj_p@1000, adj_p@1500) = 0.803 < 0.90 โ honest attribution: the readout carries only 30 distinct values (K=99 granularity + monotonicity coarsening); coarse discrete scores are inherently rank-unstable
FAIL
e-ICP evidence shadow (named decisive)
โ0.541 โ the evidence channels genuinely differ
CLEAR
Standard grid
worst |ฯ| = 0.611 ยท k_v 0.043
CLEAR
The distinction that matters
M3 kills the INSTRUMENT form (a per-feature ranking scalar); row 100's finding โ FWER 0/49 where HAC hit 43% โ concerns the CONTROL property and stands untouched
OPERATOR PROPOSAL (filed, row 101): adopt the shift-based RomanoโWolf/max-T as the declaration pipeline's inference layer โ a pipeline decision outside the gate grammar, consuming no stack slot. BY's auto-re-entry (row 99) rides on the same decision; the 03c wild-bootstrap gray zone (row 100) folds into the same review. The successor note for the instrument form (continuous margin or K=999) is on the boneyard, but the layer's actual need โ a valid inference backbone โ requires no admission.
โถ Next iteration
Next firing: B-06 advances to the H-094/H-095 SPAโMCS pair โ GROUPED exam (the feed's pair note: Hansen's Superior Predictive Ability + the Model Confidence Set โ snooping-robust superiority and the SET of regimes indistinguishable from best; stationary-bootstrap nulls in-box, though the 03c status of the stationary bootstrap (block resampling of real values โ clean) vs any wild component must be checked per candidate at the exam). Set-valued scoping is MCS's delta: it reads which REGIMES are indistinguishable from a feature's best โ directly the CONDITIONAL-scoping question. After the pair: H-090 exceedance correlation opens the characterization layer. Standing for the operator:the RW pipeline-adoption proposal (row 101), the 03c wild gray zone (row 100), the HAC flag (row 99), the M4-bar question, the H-062 proposal, the inverted-mechanism proposal (ready), the 73 cycle-1 drafts, the Frontier #4 review, the STABLE-declaration proposal, and PR #599 โ B-06 now 0 admits ยท 4 kills (one with a live pipeline proposal) ยท 7 queued.