Frozen Text-Reading Transfer / V68
A readout of frozen Qwen3-0.6B features reached 79.25% on seen text styles and 51.75% on withheld styles. Both development gates failed and sealed evaluation was not executed. The native Foundation derivative also failed; a general text-reading interface remains unqualified.
CONTROL COMPARISON
Development configurations rendered in five styles; English and command-line styles were withheld from training.
Both acquisition gates failed. The sealed set stayed unopened; seen-style calibration did not qualify cross-style reading.
- Qwen readout · seen / withheld
- 79.25% / 51.75%
- Native reader · seen / withheld
- 63.00% / 54.37%
- Character baseline · withheld
- 67.00%
- Required seen / withheld floors
- 80.00% / 75.00%
Task and model
Predict training outcomes from textual configuration descriptions. A 65,730-parameter head reads frozen Qwen3-0.6B representations; a native 001.x Foundation derivative provides a separate comparison. Training covers YAML, Python, and JSON descriptions. English and command-line styles are withheld. Development uses 400 configurations rendered in each of five styles.
Results
| Reader | Seen-style accuracy | Withheld-style accuracy |
|---|---|---|
| Frozen Qwen features and learned head | 79.25% | 51.75% |
| Native Foundation derivative | 63.00% | 54.37% |
| Required gates | 80% | 75% |
The Qwen head also fell below the 67% character baseline on withheld styles. Seen-style calibration passed, but both acquisition gates failed. The candidate was rejected and the sealed set remained unopened. Restoration reproduced cached-feature probabilities; 227 protected weight artifacts remained unchanged.
Interpretation and limitations
Frozen representations did not provide a reliable interface across these styles. Parsed numeric inputs scored 80.75%, separating prediction from the failed text interface. This does not establish broad language competence. Foundation adaptation, learned pooling, and canonical restatement remain untested proposals.
SOURCE PROVENANCE
EMMA V68: an owned Decision reading text through a frozen pretrained Foundation
LABORATORY REPORT / 2026-10-08SOURCE CHECKSUM / SHA-256
4032c466812d4891110315ad170262f9a16fdb13045da29a6b173972a4dfad2dPublic 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.