Every stored partition run rescored at penalty 1.0: H1 stays inconclusive and K falls¶
# research(2026-10). The reruns the decoder entry above called for. Every development partition
run and its planted variants on both organizations (76 runs, rq1-partition-*-rp1.0.json) and both
placebo runs the H1 simulation reads (rq1-placebo-*-rp1.0.json) were rescored with their adapters
reused (decoder_check.sbatch, pinned at ddb3d1b and 121d0c6). Partition 2's and partition 3's
rescorings, run a second time in new jobs, kept every prediction, so rescoring at 1.0 reproduces
across jobs in fp32. The retrieval and rules comparators were run again at 1.0 (44 jobs,
*-rp1.0.json, pinned at 3903b90) and read with retrieval_read.py, which takes the -rp1.0
adapter runs. Own minus sibling throughout, the pilots' at 97.5%:
| reading | at 1.1 | at 1.0 |
|---|---|---|
| Openstack H1 pilot, 22 runs, 97.5% | +0.005 [-0.009, +0.019], K 24 | +0.002 [-0.012, +0.016], K 19 |
| Wikimedia H1 pilot, 22 runs, 97.5% | +0.001 [-0.013, +0.015], K 16 | +0.006 [-0.005, +0.018], K 13 |
| Openstack retrieval, k = 1 | +0.022 [+0.002, +0.041], inconclusive | +0.012 [-0.005, +0.028], inconclusive |
| Openstack retrieval, k = 3 | +0.026 [+0.006, +0.047], inconclusive | +0.017 [+0.000, +0.034], inconclusive |
| Wikimedia retrieval, k = 1 | +0.005 [-0.012, +0.024], inconclusive | +0.009 [-0.004, +0.024], inconclusive |
| Wikimedia retrieval, k = 3 | +0.009 [-0.007, +0.026], inconclusive | +0.012 [-0.003, +0.027], inconclusive |
| Openstack distilled rules file | +0.002 [-0.003, +0.007], carries none as large as the SESOI | -0.003 [-0.009, +0.003], carries none as large as the SESOI |
| Wikimedia distilled rules file | -0.001 [-0.004, +0.002], carries none as large as the SESOI | -0.001 [-0.006, +0.004], carries none as large as the SESOI |
partition-pilot-<org>-rp1.0.json, -sensitivities-rp1.0.json (the AI-trailer and backport
sensitivities) and -k-rp1.0.json (the cell read at its new K) are the 1.0 counterparts. The run
spread falls at 1.0, so the sized K falls with it. No reading changes category: H1, retrieval and
the adapters stay inconclusive, and the distilled rules files still carry no contrast as large as
the SESOI.
The contamination battery at 1.0 (contamination-openstack-6mo-with_context-rp1.0.json) runs in
the dtypes of the stored -gapk run (membership in bf16, guided completion in fp32), on another
node. The membership gaps agree to three decimals (Min-K% -0.130, Min-K%++ -0.075, Gap-K% -0.056), and Min-K%++ separates the
windows at AUC 0.541 in both, so the blind-baseline reading stands. Guided completion's gap moves
from +0.003 to +0.008, still near its floor.
Not rebuilt, and why. The calibration ladder (calibration-marker-*) predates the fp32
registration and records no LoRA rank, so rescoring would change its dtype too and --reuse-from
refuses it; the partition-level planted runs, outcome-neutral test 5 for H1, are rescored above
instead. The -stopping pilot and simulation were written by an earlier partition_pilot.py and
are superseded by the fixed K. Still running: the H1 simulations on the 1.0 pilots and placebo
runs (SPORC 21829169 and 21829170), and the determinism check at 1.0 on two nodes.