Proposal for a future QLM revision — filed for the record after #62/#63/#64; no code attached.
The ledger's headline estimator is the median with seeded bootstrap CIs (ADR-020) — deliberately robust statistics. The publishability screen, however, is CV = stdev/mean (REQ-LEDGER-005), which is maximally non-robust: a single outlying repetition among 20 moves CV from ~0.005 to ~0.045, and one among 12 to ~0.060, while the median it is guarding moves by nothing. On intel-coffee-lake an irreducible ~1-in-35 per-process slow latch (+21%, survives C-state holds, uncore pinning, RT scheduling, hugetlb-pinned physical layout; SMI count zero during a caught run — see #62) makes this concrete: entries whose medians are rock-solid get excluded because the dispersion proxy is outlier-sensitive.
Proposal: gate on a robust dispersion measure matched to the estimator, e.g. bootstrap-CI width relative to the median ((ci95_hi − ci95_lo)/median, already computed and stored per entry) or MAD/median, with thresholds calibrated against the committed corpus so today's accept/reject boundary is approximately preserved for well-behaved entries.
This is a statistics-policy change: it bumps QLM (REQ-LEDGER-014) and needs the amendment process, threshold calibration against all committed runs, and a migration note for how QLM-1 and QLM-2 entries coexist in the append-only ledger. --rerun-noisy (#64) is the QLM-1-compatible operational answer; this issue is the principled successor when a methodology revision is next on the table.
Proposal for a future QLM revision — filed for the record after #62/#63/#64; no code attached.
The ledger's headline estimator is the median with seeded bootstrap CIs (ADR-020) — deliberately robust statistics. The publishability screen, however, is CV = stdev/mean (REQ-LEDGER-005), which is maximally non-robust: a single outlying repetition among 20 moves CV from ~0.005 to ~0.045, and one among 12 to ~0.060, while the median it is guarding moves by nothing. On intel-coffee-lake an irreducible ~1-in-35 per-process slow latch (+21%, survives C-state holds, uncore pinning, RT scheduling, hugetlb-pinned physical layout; SMI count zero during a caught run — see #62) makes this concrete: entries whose medians are rock-solid get excluded because the dispersion proxy is outlier-sensitive.
Proposal: gate on a robust dispersion measure matched to the estimator, e.g. bootstrap-CI width relative to the median (
(ci95_hi − ci95_lo)/median, already computed and stored per entry) or MAD/median, with thresholds calibrated against the committed corpus so today's accept/reject boundary is approximately preserved for well-behaved entries.This is a statistics-policy change: it bumps QLM (REQ-LEDGER-014) and needs the amendment process, threshold calibration against all committed runs, and a migration note for how QLM-1 and QLM-2 entries coexist in the append-only ledger.
--rerun-noisy(#64) is the QLM-1-compatible operational answer; this issue is the principled successor when a methodology revision is next on the table.