Operational Learning Controls / V61.2
Earlier V61 probes stopped before fitting because competent controls left insufficient headroom. V61.2 fitted a 20,737-parameter candidate, but on 160 sealed tasks it completed 102 versus 132 for Memory and 160 for exact and full-check controls. Useful generalized learning was not established.
CONTROL COMPARISON
160 sealed operational tasks; completion and optimal choice are separate outcomes.
The candidate did not establish a useful generalized-learning advantage and was rejected. Independent verification blocked incorrect accept decisions.
- Learned candidate
- 102 completed · 85 optimal
- Memory control
- 132 completed · 105 optimal
- Exact control
- 160 completed · 160 optimal
- Always full check
- 160 completed
Method
V61 and V61.1 compared actual task outcomes with competent controls and stopped before fitting because oracle headroom was below their required gates. V61.2 expanded the representation and fitted a separately owned 20,737-parameter verification-range candidate on 352 independently verified training and replay events.
Results
| Sealed condition | Completed tasks | Optimal decisions |
|---|---|---|
| Learned candidate | 102/160 | 85/160 |
| Memory control | 132/160 | 105/160 |
| Exact control | 160/160 | 160/160 |
The always-full-check control also completed 160 tasks. Independent verification blocked 58 incorrect accept decisions. Candidate utility remained below competent controls; the candidate was rejected after failed generalization and retention.
Interpretation and limitations
Actual execution and independent verification make feedback meaningful, but feedback alone does not establish useful learning. V62 and V63 investigate alternative isolated representations. These records do not qualify autonomous repair or deployment.
SOURCE PROVENANCE
V61–V61.2: verified operational learning, candidate rejected
LABORATORY REPORT / 2026-10-06SOURCE CHECKSUM / SHA-256
83e2a8fe7225f93b451a51b4172c401dac9556e6e428ff128ae0594be0cadb56Public journal edition reviewed 2026-10-08. 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.