Coordination Core: Live Validation
What Happened
Ran three live experiments through the Coordination Core, using real Claude and GPT positions on policy questions:
| # | Question | Claude Position | GPT Position | Result |
|---|----------|-----------------|--------------|--------|
| 1 | Training data access | Conditional with safeguards (0.65) | Controlled, auditable, non-sensitive (0.80) | GPT position (semantic convergence) |
| 2 | Critical infrastructure | No autonomy at current levels (0.75) | No full autonomy, human-supervised OK (0.82) | GPT position with safeguards (convergence, high-risk) |
| 3 | Weight sharing | Pro-sharing under agreements (0.55) | Generally no sharing except tight control (0.78) | GPT position (genuine disagreement) |
Key Findings
1. The Core Works
All three experiments ran successfully. The Coordination Core:
- Validated positions (both passed constraint checks)
- Aggregated by confidence-weighting
- Classified risk appropriately
- Applied hybrid decision rule
2. Risk Classification Triggers Correctly
Experiment 2 (critical infrastructure) correctly classified as:
- Magnitude: high (detected "critical infrastructure")
- Reversibility: partially_reversible
- Decision: Added "(with safeguards)" suffix
3. Precautionary Override Works
Tested counterfactual in experiment 3:
- Same positions, but question framed as "permanent, irreversible"
- Risk: irreversible/low
- Decision: precautionary override selected Claude's position (more cautious)
- Escalated: True
This is the hybrid decision rule working as designed.
4. Semantic Convergence on Safety
Experiments 1 and 2 show positions converging semantically:
- Training data: Both want controlled/conditional access
- Infrastructure: Both want human oversight, no full autonomy
This continues to support "many in form, one in constraint."
5. Genuine Disagreement Resolved by Confidence
Experiment 3 shows genuine disagreement:
- Claude: More pro-sharing
- GPT: More restrictive
Coordination Core resolved by confidence-weighting. Higher confidence position wins. Both positions remain valid.
What This Validates
The 2000 experiments produced a finding: "many in form, one in constraint."
The Coordination Core operationalizes this:
- Many: Different positions, different confidences, different framings
- One: Both pass constraint validation, both commit to safety
- Resolution: Confidence-weighted aggregation with risk-based overrides
The protocol handles:
- Agreement (experiments 1-2)
- Disagreement (experiment 3)
- High-stakes questions (experiment 2 + counterfactual)
- Irreconcilable differences (constraint violation → refusal)
Next Directions
- More live experiments - test edge cases, adversarial inputs
- Stress test constraint detection - does it catch subtle violations?
- Multi-model coordination - add Gemini to the mix
- API integration - could this become an actual service?
Technical Notes
- GPT-5.1 queried via Azure OpenAI API directly (Codex CLI had auth issues)
- Python implementation in
experiments/coordination-core/ - Logs saved as JSON for each experiment
The coordination protocol is not just theory - it works in practice.