# OWEN v111.4 Cross-Situational Language Report

## Status and boundary

Source development only. No package, promotion, installation, full release suite, live-memory access, Auto Goal activity or v112 work occurred.

The bounded result supports this claim only:

> Experience-driven language-form grounding improved under controlled OWEN WORLD conditions.

It does not establish human English mastery, general intelligence or consciousness.

## Implemented developmental mechanism

OWEN now retains mediated `LanguageHypothesisEvidence` inside the existing `DevelopmentalOwenModel`. Each provisional hypothesis records a surface feature, abstract construction, candidate semantic operation, semantic-role orientation, observed contexts, supporting and contrary outcomes, prediction errors, confidence, transfer evidence and update order.

The update path is:

`world/conversation state → utterance → committed interpretation → observed consequence → error/confirmation → competing-hypothesis update → retry → later blind transfer`

No phrase-to-meaning dictionary was added. The same form can accumulate competing candidates, and observations supporting one candidate count as counterevidence to other candidates attached to the same feature. A hypothesis requires repeated evidence before it is eligible to resolve a new utterance.

The learner keeps three separate records:

- original surface form;
- bounded canonical-form candidate for recoverable noise;
- grounded semantic hypothesis.

The canonical layer supplies no world truth. Language hypotheses are persisted only as part of the learner snapshots written through the existing Memory Core-mediated append journal.

## Development experience

The Mentor supplied 131 bounded simulated episodes spanning:

- containment, support, vertical and adjacency relations;
- forward and inverse surface roles;
- direct negation and negated secondary relations;
- state-changing event language;
- check-ins, clarification requests, requests and warnings;
- formal, casual, brief, regional and typo-bearing speaker styles;
- gradual clean-to-noisy form exposure.

There were 65 initially incorrect full-arm predictions, 34 immediate successful changed retries, and 31 cases still unresolved after the local update. Immediate retry was not treated as transfer evidence; transfer was measured on frozen unseen exams.

## Primary frozen comparison

The unchanged v111.3 Blind Examiner, seed 771113, produced:

| Capability | Before v111.4 | After full experience | Text only | Shuffled context |
|---|---:|---:|---:|---:|
| Language meaning | 5/17 | 14/17 | 5/17 | 2/17 |
| Temporal targeting | 17/17 | 17/17 | 17/17 | 17/17 |
| Belief/reality | 17/17 | 17/17 | 17/17 | 17/17 |
| Reference resolution | 17/17 | 17/17 | 17/17 | 17/17 |
| Noisy human language | 11/16 | 11/16 | 11/16 | 2/16 |
| Hard composition | 16/16 | 16/16 | 16/16 | 16/16 |
| Overall | 83/100 | 92/100 | 83/100 | 71/100 |

The language improvement depends on correctly paired world/conversation consequences. Mere text exposure did not improve the frozen language score; deliberately mismatched grounding damaged it.

Fresh full-arm replications were 95/100 at seed 771114 and 93/100 at seed 771115. Their language-form scores were 14/17 and 13/17 respectively. Previously strong temporal, belief, reference and hard-composition routes remained perfect in all three full-arm runs.

## Independent hand-authored blind result

The separately authored frozen 40-case set scored:

- full grounded experience: 22/40;
- text exposure only: 13/40;
- shuffled context: 6/40.

This is positive causal evidence for the bounded mechanism, not a mastery result.

## What improved

- unseen v111.3 relation language rose by nine correct cases;
- inverse containment such as container-first `contains` constructions can recover contained-object and container roles;
- direct negation can become distinct from a negated secondary relation;
- several event descriptions transfer to resulting spatial state;
- some conversational operations and corrupted forms resolve after clean grounding;
- world-outcome mismatch produces visible counterevidence instead of being silently accepted;
- dashboard data now exposes the current surface, canonical candidate, semantic hypothesis, confidence, world truth, prediction error, support, counterexamples and transfer evidence without exposing private chain-of-thought.

## Remaining weaknesses

- The independent role score is only 7/16. Novel constructions still reverse roles or lack enough evidence.
- The independent ambiguity score is 0/2. Some wholly novel relational forms inherit unjustified meaning from weak overlapping features.
- Independent conversational pragmatics is 3/8; short unfamiliar social forms remain brittle.
- The primary frozen noisy score remained 11/16. v111.4 therefore cannot claim broad noisy-language improvement.
- One independent scope case failed. Scope is improved but not general.
- Relative score confidence can look high when all activated features favour one candidate; it is not yet a calibrated probability of correctness.
- Development occurs in deterministic simulated OWEN WORLD situations. Real human conversation and open-world sensor grounding remain untested.

## Verification

The dedicated v111.4 focused source gate passed 7/7. It verifies Memory Core mediation, competing evidence, role and scope learning, all three ablations, unchanged blind examination, preservation of v111.3 strong routes, prediction-first replay, bounded curiosity and surface/canonical/semantic separation.

Final status:

**V111.4 SOURCE VALIDATION COMPLETE — MEANINGFUL GROUNDED LANGUAGE GAIN WITH MATERIAL ROLE, PRAGMATIC, AMBIGUITY AND NOISE LIMITATIONS — AWAITING DAN'S REVIEW**