Reading Remembered Runs / V101–V101.1
A learned reader of the runs the system remembers for a configuration first failed: its accuracy stayed near 84% however many runs it saw, below a calibrated counting rule at 87.0%. A redesign that counts its evidence reached 89.1% on 1,800 fresh sealed runs, 4.6 points above the rule the demonstration had used and equal to the best calibrated rule, and passed every gate.
SCOPE + LIMITATIONS
- One task family. The reader’s value is expected where a fixed rule cannot be calibrated, such as several kinds of memory read together.
- This public edition reviews saved evidence; experiments were not rerun for publication.
CONTROL COMPARISON
For each configuration, give the reader 1 to 5 remembered runs and ask whether a new run fails. Train on earlier runs and score once on 1,800 fresh runs (300 configurations, 6 seeds) against the estimate alone and a counting rule at its original and best strengths. V101.1 is scored on a fresh sealed set.
V101 failed the accuracy and evidence-following gates; V101.1 passed all four. It equals, and does not exceed, the best single calibrated rule. In the whole demonstration the reader and the rule differed on 8 of 394 cases, within noise, so the simpler rule stays the default and the reader is kept (V102.5).
- V101 · reader
- 83.9% mean; failed
- V101 · best calibrated rule
- 87.0%
- V101.1 · reader
- 89.1%
- V101.1 · original rule / best rule
- 84.5% / 89.1%
- V101.1 · estimate alone
- 79.9%
- No remembered runs
- Estimate returned unchanged in both versions
Question and method
For each configuration, give the reader 1 to 5 remembered runs and ask whether a new run fails. Train on earlier runs and score once on 1,800 fresh runs (300 configurations, 6 seeds) against the estimate alone and a counting rule at its original and best strengths. V101.1 is scored on a fresh sealed set.
Recorded results
| Condition | Recorded result |
|---|---|
| V101 · reader | 83.9% mean; failed |
| V101 · best calibrated rule | 87.0% |
| V101.1 · reader | 89.1% |
| V101.1 · original rule / best rule | 84.5% / 89.1% |
| V101.1 · estimate alone | 79.9% |
| No remembered runs | Estimate returned unchanged in both versions |
Comparative standing and locked gates
V101 failed the accuracy and evidence-following gates; V101.1 passed all four. It equals, and does not exceed, the best single calibrated rule. In the whole demonstration the reader and the rule differed on 8 of 394 cases, within noise, so the simpler rule stays the default and the reader is kept (V102.5).
Interpretation and limitations
One task family. The reader’s value is expected where a fixed rule cannot be calibrated, such as several kinds of memory read together.
SOURCE PROVENANCE
Reading Remembered Runs / V101–V101.1
LABORATORY REPORT / 2026-10-25SOURCE CHECKSUM / SHA-256
3f4c6a13ddab0268897831a718d20f28ab2d28fafa45f731d5f356b9d18604d0Public journal edition reviewed 2026-10-28. Source documents and saved evidence were inspected; experiments were not rerun for this edition. Proprietary implementation code, model binaries, private infrastructure, and detailed machine records are not published here. Journal identifiers are editorial references. Catalog inclusion does not imply qualification or runtime promotion.