Observation-Equivalence Branching Audit / V75.5
A measurement-only audit branched 300 configurations into seven continuations from identical states. Majority agreement estimated 98.7% at step forty and 99.2% at step 120. These are finite-sample uncertainty estimates, not achieved prediction scores or qualification gates.
CONTROL COMPARISON
Freeze every bit of state at a chosen step and redraw future mini-batch order seven times. Mean agreement with the majority estimates how much outcome uncertainty remains. The executor reproduced the original outcomes on all 300 unbranched runs. An earlier nearest-neighbor audit was uninformative and is not used.
No training or gate applies. Measured agreement suggests room above the opening predictors, particularly at shorter openings. Audit configurations differ from the sealed predictor set, so the side-by-side scores are descriptive rather than matched improvements or a proof of a universal ceiling.
- Initial state · step zero
- 97.6% majority agreement
- Step forty
- 98.7% agreement; V75.3 predictor 92.5% on a separate sealed set
- Step 120
- 99.2% agreement; V75.3 predictor 97.6% on a separate sealed set
- Unbranched executor check
- 300/300 reproduced outcomes
Question and method
Freeze every bit of state at a chosen step and redraw future mini-batch order seven times. Mean agreement with the majority estimates how much outcome uncertainty remains. The executor reproduced the original outcomes on all 300 unbranched runs. An earlier nearest-neighbor audit was uninformative and is not used.
Recorded results
| Condition | Recorded result |
|---|---|
| Initial state · step zero | 97.6% majority agreement |
| Step forty | 98.7% agreement; V75.3 predictor 92.5% on a separate sealed set |
| Step 120 | 99.2% agreement; V75.3 predictor 97.6% on a separate sealed set |
| Unbranched executor check | 300/300 reproduced outcomes |
Comparative standing and locked gates
No training or gate applies. Measured agreement suggests room above the opening predictors, particularly at shorter openings. Audit configurations differ from the sealed predictor set, so the side-by-side scores are descriptive rather than matched improvements or a proof of a universal ceiling.
Interpretation and limitations
Only mini-batch order is varied. Seven branches bias majority agreement upward; 300 configurations do not establish a formal population bound. Initial loss and gradient observations are proposed next inputs, not verified gains.
SOURCE PROVENANCE
V75.5 branching audit: how much of a run's outcome is still undecided at a given step
LABORATORY REPORT / 2026-10-08SOURCE CHECKSUM / SHA-256
f01f3ec8fa71f150ec12d4c7f4c4af64aa9d78c684165d212160fd1b98d3c88dPublic journal edition reviewed 2026-10-09. 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.