Temporary Learning and Decay / V69
A temporary residual unit learned within a new synthetic regime but failed calibration across three seeds. Primary validation accuracy was 79.33% versus 72% for the frozen base; calibration error was 0.098 against a 0.08 ceiling. Units decayed and the sealed set remained unopened.
CONTROL COMPARISON
Primary seed development validation: 300 new-regime runs. These are not sealed results.
All three seeds failed calibration. Units decayed and the sealed set stayed unopened; V69.1 is a separate follow-up.
- Temporary unit
- 79.33% · calibration error 0.098
- Frozen composition
- 72.00%
- Locked calibration ceiling
- 0.08 · failed
Model and protocol
A 305-parameter residual MicroModel learns from 1,200 verified training-run outcomes in twelve blocks, predicting before updating. The frozen V67 composition remains separately owned. Promotion requires improvement, calibration, preservation of older behavior, and successful controls.
Results
The primary unit reached 79.33% on 300 validation runs versus 72% for the frozen base. Calibration error was 0.098, above the locked 0.08 ceiling. The other two seeds also failed at 0.086 and 0.092. All units decayed; the 800-run sealed set was not executed. All 800 older-regime answers remained bit-identical. Direct fine-tuning reached 81.67% in the new regime but reduced old-regime accuracy from 85.50% to 80.37%, changing 137 older answers. A permuted-outcome unit also failed and decayed.
Interpretation and limitations
This demonstrates bounded learning and enforced isolation, but not an accepted capability. Improvement alone did not satisfy promotion. V69.1 subsequently tests an explicit calibration revision with fresh calibration and validation data; it does not retroactively change this failed gate.
SOURCE PROVENANCE
EMMA V69: a temporary unit that learns during operation, then is promoted or decays
LABORATORY REPORT / 2026-10-08SOURCE CHECKSUM / SHA-256
ee2deb3a4c5306e16f6fc9301a5bf25837168bc932258cf444a37a565a56caa9Public 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.