# OWEN v111.4 Language Development Replay

These are human-readable views of recorded prediction-first episodes. They are simulated developmental experiences, not autobiographical memories.

1. **Unknown clean containment** — `item is inside place`: OWEN initially had no v111.4 relation hypothesis. One encounter was retained but did not immediately justify resolution; he requested another comparable situation.
2. **Repeated containment** — the same construction with new objects and context changed the retry from unknown to `inside`, with roles `item → place`.
3. **Inverse containment** — `place contains item`: early prediction lacked the relation and reversed the visible argument order; repeated outcomes established that the contained item remains semantic subject.
4. **Different inverse wording** — `place holds item in its interior`: repeated new-object outcomes produced a successful changed retry without installing a phrase rule.
5. **Same meaning, event wording** — `item went into place`: the observed table-to-container transition changed an unknown prediction to `inside`.
6. **Same event, inverse result wording** — `place now contains item`: compatible transition evidence reinforced the inverse semantic roles.
7. **Different meaning with related vocabulary** — adjacency and containment cases shared object names and sentence frames; the world relation, not vocabulary overlap, selected the candidate.
8. **Forward support** — `item is on place`: repeated surface-support outcomes grounded `on` with item-first roles.
9. **Inverse support** — `place supports item`: the same objects in reversed wording required container/surface-first role orientation.
10. **Vertical relation** — `item hovers higher than place`: repeated new settings moved the candidate from unknown to `above`.
11. **Direct negation failure** — `item is not above place`: the first case retained a positive prediction and a 0.25 field error; one exposure was deliberately insufficient.
12. **Direct negation changed retry** — after contrasting scoped cases, a later `not above` case changed polarity to false while retaining the `above` operation.
13. **Negated secondary relation** — `item is above place without touching it`: the main `above` relation remained positive; `without` did not negate the entire sentence.
14. **Contrastive scope** — `item is beside place, not within it`: OWEN retained positive adjacency while the negation applied to containment.
15. **Conversational request** — `Pass us that tool, mate`: an initially unknown operation became `REQUEST` after repeated interaction consequences across formal and regional styles.
16. **Warning transfer** — `Careful, that's hot`: later episodes resolved as `WARNING` after comparable protective consequences.
17. **Noisy clarification** — `wot exactly do you mean`: surface/canonical evidence remained separate; after clean clarification experiences, the retry became `CLARIFICATION_REQUEST`.
18. **Residual noisy failure** — `wot dya mean`: overlapping weak cues first activated an unrelated candidate and then ambiguity. The system did not count this as successful understanding.

## Blind surprise results

On the frozen v111.3 examiner, 3/17 language cases still failed after development. On the separate hand-authored set, role generalisation failed in 9/16 cases and both novel ambiguity cases produced unjustified certainty. These are retained as the next developmental boundary, not converted into teaching phrases.

## Behavioural change

The full arm produced 34 successful changed retries and later blind language improvement. Text-only produced zero successful changed retries and no primary blind language improvement. Shuffled context produced local changes but substantially worse blind transfer, demonstrating that an observed consequence can change future behaviour in the wrong direction when its grounding is false.