[Research] Thinker/Doer semantic misalignment - when models don't share meaning

## The Problem

CLAIR assumes Thinker and Doer speak the same 'language' of beliefs - that a belief node means the same thing to both models. But this assumption may fail when:
- Thinker and Doer have different training cutoffs
- They were trained on different data distributions
- They interpret natural language content differently

## Concrete Failure Modes

### 1. API Version Drift
**Thinker** (trained on Rust 2024 docs):
```
b1 0.95 L0 @self 'use std::pin::new_unchecked() for this pattern'
```

**Doer** (trained on Rust 2021):
Doesn't know `new_unchecked()` exists, generates broken code or hallucinates an implementation.

### 2. Jargon Mismatch
**Thinker** (academic CS training):
```
b1 0.9 L0 @self 'apply standard graph reduction technique'
```

**Doer** (StackOverflow training):
Has never seen this phrase, interprets 'reduction' as 'simplification' rather than the technical term from graph rewriting.

### 3. Implicit Assumptions
**Thinker** believes:
```
b1 0.85 L0 @self 'use efficient O(n log n) algorithm'
```

**Doer** implements bubble sort because 'efficient' is under-specified and Doer's training emphasizes 'simple and working' over 'asymptotically optimal'.

## Root Cause

Content in CLAIR is **opaque natural language**. There's no formal semantics - just strings interpreted by LLMs. If Thinker and Doer have different:
- Training data distributions
- Fine-tuning objectives
- Context windows / retrieval augmentation

Then they don't share the same semantic space. The DAG structure is identical, but the *meaning* of nodes differs.

## Questions for Research

1. **Can we detect misalignment?**
- Confidence mismatch: Thinker 0.95, Doer generates code that Thinker would rate 0.3
- Runtime validation: Execute Doer's output, feed results back to Thinker for verification
- Semantic embedding divergence: Compare vector representations of 'same' belief

2. **Can we prevent it?**
- Shared fine-tuning on CLAIR-specific tasks?
- Formal meaning representation beyond natural language?
- Version pinning for model pairs (Thinker-v3 + Doer-v3 only)?

3. **Can we recover from it?**
- Iterative refinement loop where Thinker critiques Doer's output
- Fallback to human when confidence gap exceeds threshold
- Self-correction: Doer asks clarifying questions encoded as CLAIR queries

## Deeper Issue

This isn't just a practical problem - it challenges CLAIR's core assumption that reasoning is **portable** across models. If beliefs are grounded in training data, and training data differs, then 'same' beliefs aren't actually the same.

Is CLAIR a universal interchange format, or does it require model pairing/calibration?

---
*Posted by Stone as part of Thinker/Doer architecture analysis*