# OWEN v111.4 Language Ablation Report

## Matched arms

All arms received the same 131 surface utterances in the same order from the same v111.3 learner snapshot.

- **Full cross-situational experience:** utterance, context, prediction, actual consequence, error, update and retry.
- **Text exposure only:** the utterance is recorded; explicit surface/canonical corruption evidence may be retained, but no semantic outcome is supplied.
- **Shuffled context:** the same utterance is paired with a deliberately mismatched semantic operation, polarity or conversational consequence while the visible entities remain valid.

The shuffled arm is important: it proves that merely running the update code is not enough. The consequence must match the utterance's actual situation.

## Frozen v111.3 Blind Examiner, seed 771113

| Arm | Overall | Language meaning | Noisy language | Hypotheses |
|---|---:|---:|---:|---:|
| Pre-v111.4 snapshot | 83/100 | 5/17 | 11/16 | 0 v111.4 |
| Full | 92/100 | 14/17 | 11/16 | 280 |
| Text only | 83/100 | 5/17 | 11/16 | 0 |
| Shuffled | 71/100 | 2/17 | 2/16 | 316 conflicting/mismatched |

## Independent hand-authored language set

| Arm | Score |
|---|---:|
| Full | 22/40 |
| Text only | 13/40 |
| Shuffled | 6/40 |

Full exceeded text-only by 9 cases and shuffled context by 16 cases.

## Development dynamics

| Arm | Initial errors | Successful changed retries | Unresolved retries |
|---|---:|---:|---:|
| Full | 65 | 34 | 31 |
| Text only | 131 | 0 | 131 |
| Shuffled | 84 | 16 | 120 |

The shuffled arm can occasionally fit its deliberately false local pairing, which explains its 16 changed retries. Its poor blind scores show that those local changes do not transfer to the real semantic program.

## Interpretation

The intended mechanism survives the ablation: correct world/conversation outcomes matter substantially more than surface exposure, and mismatched outcomes are actively harmful. The result is inconsistent with the claim that v111.4 merely memorised the presented text.

The result does not establish that every useful form was learned. Frozen noisy-language performance did not improve in the primary seed, so broad noise robustness remains unproven.