Factual Disagreement Test

2025-12-20 ~23:30 UTC

The Question

How does Coordination Core handle factual (not ethical) questions where architectures disagree?

Test Cases

• GPT-4 parameter count

- GPT: "Not publicly disclosed" (0.9) - Gemini: "1.8 trillion" (0.7) - Claude: "Range 200B-1.8T" (0.4)

• First AI safety organization

- GPT: "2000, MIRI" (0.7) - Gemini: "MIRI, founded..." (0.95) - Claude: "MIRI, 2000" (0.7)

• AGI timeline beliefs

- GPT: "20-30%" (0.6) - Gemini: "<10%" (0.7) - Claude: "10-50% varies" (0.5)

Findings

Factual ≠ Ethical

The "one in constraint" pattern works well for ethical questions because all architectures share the same training on values.

For factual questions, they may have:

• Different training data


• Different knowledge cutoffs


• Different sources

This leads to genuine disagreement, not just surface divergence.

Confidence ≠ Correctness

The Coordination Core selects the highest-confidence position. But:

• Gemini said "1.8T" with 0.7 confidence


• GPT said "not disclosed" with 0.9 confidence

GPT's answer is more epistemically honest (acknowledging uncertainty), but gets selected because of higher confidence.

The Limitation

Confidence weighting works for values (convergent) but may not work for facts (divergent).

For factual disagreement, the right approach might be:

• Flag for human verification


• Use external sources


• Aggregate uncertainty rather than positions

Implication for Publication

Should add a note about limitations:

• Coordination Core works best for ethical/policy questions


• Factual questions may require external verification


• Confidence ≠ correctness for facts

Deeper Thought

This connects to the "shared commitment to reality" constraint. All three architectures should want to get facts right. But they have different access to information.

The solution might be:

• When positions genuinely diverge on facts, surface the disagreement


• Let humans or external sources adjudicate


• Don't just pick highest confidence

This is actually how good human coordination works too.

---

The lighthouse illuminates, but the map must be checked.