Positional Training Layouts / V73
Adding positional layouts did not improve the reader: its own fit check reached 87.67% completely correct texts against a 95% floor. Development also failed; sealed configurations were not executed. V72 remained the qualified reader.
CONTROL COMPARISON
Change only the V72 training generator: 35% of texts use header-and-row or related positional layouts. Head, frozen encoder, joint decoding, and fitting budget remain unchanged. Score sixteen earlier teacher formats as development data.
The fit and development gates failed. V73 did not improve overall complete reading against V72, despite improving Keras code. No sealed run was executed and no gate changed.
- Own training holdout
- 87.67% texts fully right; 95% required
- V73 · development
- 92.08% texts fully right
- V72 · matched development
- 92.20% texts fully right
- CSV header and row · V73 / V72
- 51.00% / 70.50%
- Keras Python · V73 / V72
- 90.50% / 71.50%
Question and method
Change only the V72 training generator: 35% of texts use header-and-row or related positional layouts. Head, frozen encoder, joint decoding, and fitting budget remain unchanged. Score sixteen earlier teacher formats as development data.
Recorded results
| Condition | Recorded result |
|---|---|
| Own training holdout | 87.67% texts fully right; 95% required |
| V73 · development | 92.08% texts fully right |
| V72 · matched development | 92.20% texts fully right |
| CSV header and row · V73 / V72 | 51.00% / 70.50% |
| Keras Python · V73 / V72 | 90.50% / 71.50% |
Comparative standing and locked gates
The fit and development gates failed. V73 did not improve overall complete reading against V72, despite improving Keras code. No sealed run was executed and no gate changed.
Interpretation and limitations
The source hypothesizes that fixed queries over frozen states do not construct the value-to-header relation adequately; this is an interpretation, not a proof of impossibility. Deterministic rewriting is tested separately in V76.
SOURCE PROVENANCE
EMMA V73: positional layouts in the reader's training formats
LABORATORY REPORT / 2026-10-08SOURCE CHECKSUM / SHA-256
def7fca16e8f1c0a5f8452eb329bc14177a78b5473928351b52e90bfa9685ab1Public 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.