inc-0010
degraded
citable URL: https://halobench.com/records/inc-0010/ — this address never moves; the anchor /records/#inc-0010 keeps resolving
Full-STRIX campaign retained corrected QSA evidence with bounded instrumentation gaps
detected
2026-08-30 · window 2026-08-29T22:30:57Z → 2026-08-30T10:29:33Z · by independent-package-review-and-raw-receipt-audit · lag Configuration drift was isolated before the corrected cohort; incomplete PLE mincore and engine scatter were contained during package review before publication.
signature: Earlier QSA attempts used default 32/8192 checkpoint semantics and thinking-on; corrected r1/r2 stopped before valid measurements. Corrected r3 produced admitted cache-false rows, but 30 of 40 before/after receipts had a six-layout/five-mincore mismatch with ValueError("mmap length is greater than file size"). Engine d0 and d32768 prompt-processing CVs were 3.731412% and 3.131244%.
should have caught it
The QSA launcher should have asserted thinking-off and checkpoint cadence before its first measurement; the PLE adapter should have validated every tensor range; the engine wrapper already caught the scatter through its CV trigger.
diagnosis
initial: One campaign could support capability, a comparable engine floor, served QSA, complete PLE fault tracking, cache traces and joined energy.
actual: Tau2 and the corrected thinking-off 512/2048 served cohort are interpretable; prefix reuse is separately interpretable. The earlier default-QSA attempts, incomplete per-PLE adapter and engine trigger rows cannot support their intended aggregate claims. An interrupted d240000 cell was repeated once, and an interrupted 131k cache warm-up was preserved before a distinct unchanged-config resume-r4 pair. Corrected r1/r2 launcher stops and every raw attempt remain in the immutable package.
false-path cost: Two pre-measurement corrected launch attempts, one interrupted engine cell, one interrupted cache warm-up, and package-only review rework; no valid model attempt was rerun for score.
falsified by: Corrected server argv/startup logs, all 40 memory receipts, raw engine samples, retained interrupted attempts, report schema v2, and the 517-member manifest.
actual: Tau2 and the corrected thinking-off 512/2048 served cohort are interpretable; prefix reuse is separately interpretable. The earlier default-QSA attempts, incomplete per-PLE adapter and engine trigger rows cannot support their intended aggregate claims. An interrupted d240000 cell was repeated once, and an interrupted 131k cache warm-up was preserved before a distinct unchanged-config resume-r4 pair. Corrected r1/r2 launcher stops and every raw attempt remain in the immutable package.
false-path cost: Two pre-measurement corrected launch attempts, one interrupted engine cell, one interrupted cache warm-up, and package-only review rework; no valid model attempt was rerun for score.
falsified by: Corrected server argv/startup logs, all 40 memory receipts, raw engine samples, retained interrupted attempts, report schema v2, and the 517-member manifest.
lesson
Corrected startup evidence can salvage a distinct served cohort without laundering earlier attempts. Incomplete per-tensor instrumentation and scatter above the house trigger must narrow claims even when whole-file memory receipts and individual raw rows remain useful.
state
resolved · class data
Cited by — computed at build time, never stored
claims clm-0127