Decision-Head Optimization Audit / V78.1
A training-only audit corrected the earlier underfit explanation: the Applies head reached 99.6% training accuracy after fifteen epochs, but Match remained near 64–65% after twenty epochs at two rates. No development or sealed qualification was performed.
CONTROL COMPARISON
Study 14.8-million-parameter marker heads over a frozen encoder with 3,000 examples per link, three learning rates, and up to twenty epochs. Measure training curves and a fit holdout, stopping early only at 99.5% training accuracy. One seed was executed; the planned second seed was not.
This is optimization diagnosis, not a gated qualification. Longer training explains Applies fitting but does not establish the cause of Match failure. Prior claims that both heads failed solely because of the learning-rate and epoch budget are corrected. Open trained weights remain the stronger Match reference in the saved comparison.
- Applies · rate 0.0001
- 99.6% training; 87.4% best fit holdout; fifteen epochs
- Match · rate 0.0001
- 64.4% training; 55.6% best fit holdout
- Match · rate 0.0003
- 64.6% training; 58.8% best fit holdout
- Rate 0.001
- Both heads failed to train reliably
- Open Match weights · prior saved comparison
- 74.5% first-choice accuracy on untrained operations; lab head 39.5%
Question and method
Study 14.8-million-parameter marker heads over a frozen encoder with 3,000 examples per link, three learning rates, and up to twenty epochs. Measure training curves and a fit holdout, stopping early only at 99.5% training accuracy. One seed was executed; the planned second seed was not.
Recorded results
| Condition | Recorded result |
|---|---|
| Applies · rate 0.0001 | 99.6% training; 87.4% best fit holdout; fifteen epochs |
| Match · rate 0.0001 | 64.4% training; 55.6% best fit holdout |
| Match · rate 0.0003 | 64.6% training; 58.8% best fit holdout |
| Rate 0.001 | Both heads failed to train reliably |
| Open Match weights · prior saved comparison | 74.5% first-choice accuracy on untrained operations; lab head 39.5% |
Comparative standing and locked gates
This is optimization diagnosis, not a gated qualification. Longer training explains Applies fitting but does not establish the cause of Match failure. Prior claims that both heads failed solely because of the learning-rate and epoch budget are corrected. Open trained weights remain the stronger Match reference in the saved comparison.
Interpretation and limitations
One seed, three rates, one head depth. Frozen representations are a plausible Match bottleneck, not a proven causal explanation; testing an unfrozen encoder requires another study. High training accuracy alone does not resolve Applies generalization.
SOURCE PROVENANCE
V78.1 optimization study: can the MarkerDecision heads fit their training data?
LABORATORY REPORT / 2026-10-09SOURCE CHECKSUM / SHA-256
2063f3d826428b23598825ab7f1f5b9e96e527af73c919166b646c706288e393Public 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.