Memory Layers and Learned Estimates / V84
Combining a learned estimate with five remembered runs raised accuracy from 77.6% to 87.7% on 3,600 fresh outcomes from 600 configurations. Memory alone reached 87.3%. All five gates passed, including exact-recall and executor checks; repeated-seed observations are not independent cases.
CONTROL COMPARISON
Execute 600 configurations under six seeds each. For each question seed, Memory contains other seeds of the same configuration. The learned prior has a fixed strength equivalent to four observations. Compare one, three, and five remembered outcomes with the learned estimate and a remembered-outcome majority. This is same-configuration recall, not retrieval from merely similar configurations.
All five locked gates passed. Three remembered outcomes add 5.8 points and five add 10.1 points over the learned estimate. The five-outcome combination is approximately level with memory alone, so these results do not establish a large added value from the learned prior when substantial exact-configuration history exists. Executor parity passed 100/100 checks.
- Learned estimate alone
- 77.6%
- With one remembered run
- 79.5%
- With three remembered runs
- 83.4%
- With five remembered runs
- 87.7%
- Majority of five remembered runs alone
- 87.3%
- Exact-recall prototype check
- 100/100 stored outcomes; no execution
Question and method
Execute 600 configurations under six seeds each. For each question seed, Memory contains other seeds of the same configuration. The learned prior has a fixed strength equivalent to four observations. Compare one, three, and five remembered outcomes with the learned estimate and a remembered-outcome majority. This is same-configuration recall, not retrieval from merely similar configurations.
Recorded results
| Condition | Recorded result |
|---|---|
| Learned estimate alone | 77.6% |
| With one remembered run | 79.5% |
| With three remembered runs | 83.4% |
| With five remembered runs | 87.7% |
| Majority of five remembered runs alone | 87.3% |
| Exact-recall prototype check | 100/100 stored outcomes; no execution |
Comparative standing and locked gates
All five locked gates passed. Three remembered outcomes add 5.8 points and five add 10.1 points over the learned estimate. The five-outcome combination is approximately level with memory alone, so these results do not establish a large added value from the learned prior when substantial exact-configuration history exists. Executor parity passed 100/100 checks.
Interpretation and limitations
Three regimes of one synthetic task. The six seed questions for each configuration are correlated; the independent configuration count is 600. Memory was populated by the experiment rather than accumulated during use, and no promoted live unit was attached. Outcomes differed across seeds in 31.5% of cases, so only identical-seed recall is exact. Alternative prior strengths were diagnostic and remain unqualified.
SOURCE PROVENANCE
V84: the memory layers together against the learned estimate alone
LABORATORY REPORT / 2026-10-10SOURCE CHECKSUM / SHA-256
477e8b9cefdd2e78eb423c5a04038f2cd36de180a0a61349f13847b1d02ea284Public journal edition reviewed 2026-10-10. 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.