# OWEN v111.4 Language Hypothesis Report

## Record contract

Each provisional language hypothesis contains:

- deterministic `hypothesis_id`;
- observed `surface_feature`;
- abstract `construction_pattern` with entity slots;
- candidate semantic operation;
- candidate surface-to-semantic roles;
- distinct contexts seen;
- support and counterexample counts;
- bounded prediction-error history;
- evidence-weighted confidence;
- transfer evidence identifiers;
- monotonic last-update index.

These records live inside `DevelopmentalOwenModel` snapshots. The v111.4 append journal is mediated by the existing `OwenMemoryCore`; it is not a second autobiographical or truth store.

## Selection behavior

At interpretation time, OWEN extracts lexical, ordered, between-entity and abstract-frame evidence. Multiple candidates compete. Candidates with fewer than two supports or low confidence cannot resolve a relation. A narrow winning margin yields ambiguity rather than a guessed relation.

Feature weighting favours ordered construction evidence over isolated words. A capped whole-frame feature may support an interpretation but cannot establish transfer alone.

## Representative developed hypotheses

The full arm produced 280 candidate records. Examples include:

| Surface evidence | Candidate | Roles | Support/context evidence | Confidence |
|---|---|---|---|---:|
| `contain` between entities | `inside` | first=container, second=contained object | six observations across static and transition settings | 0.875 |
| `inside` between entities | `inside` | first=contained object, second=container | repeated static settings | bounded, high |
| `on` between entities | `on` | first=item, second=surface | six observations across three settings | 0.875 |
| `above` after first entity | `above` | first=upper object, second=lower object | nine observations including scope contrasts | 0.909 |
| `beside` between entities | `beside` | first=item, second=neighbour | six observations including contrastive scope | 0.875 |
| `mean` in conversation | `CLARIFICATION_REQUEST` | no world-object roles | nine observations across clean/noisy conversational contexts | 0.909 |
| `pass` in conversation | `REQUEST` | no world-object roles | eight formal/regional/noisy outcomes | 0.900 |

Confidence values are internal evidence summaries, not calibrated probabilities that an answer is correct.

## Counterevidence

When the same feature occurs with a different observed semantic signature, the supported candidate is incremented and competing candidates receive counterevidence. The shuffled-context arm created 316 conflicting or incorrect records and scored only 2/17 on the frozen language test, demonstrating that record quantity is not evidence of grounding quality.

## Dashboard data

The developmental dashboard contract now includes:

- current utterance and untouched surface;
- canonical candidate;
- semantic hypothesis and role mapping;
- confidence;
- actual world/conversation truth;
- prediction error;
- number of supporting experiences and counterexamples;
- transfer evidence;
- explicit `private_chain_of_thought_exposed: false`.

This is an evidence replay, not a display of hidden chain-of-thought.

## Known limitations

- Construction extraction remains shallow and token-based around known entity mentions.
- Confidence is not fully calibrated and can be overconfident when only one candidate is activated.
- Cross-clause scope and long-distance dependencies remain limited.
- Synonym-like transfer emerges only where surface evidence overlaps enough across experiences.
- New words with no grounded overlap remain unknown—or, in two blind cases, are incorrectly captured by weak shared evidence.
- Speaker style is recorded as context but no mature speaker-conditioned generative model exists.
- Real audio prosody, gesture and open-world perception are outside this bounded source phase.