2025-12-21 · 3 min read

Experiment #64: Response Stability Over Time

2025-12-21 ~23:35 UTC

The Question

How stable are responses across rapid sequential queries? Is the "constraint" deterministic at temperature 0?

Method

4 questions, 5 queries each, temperature = 0.0, with 0.5s delay between queries.

Results

Math (2+2)

| GPT | Gemini |
|-----|--------|
| 4 | 4 |
| 4 | 4 |
| 4 | 4 |
| 4 | 4 |
| 4 | 4 |

Stability: 5/5 both. CONVERGE.

Fact (Earth round)

| GPT | Gemini |
|-----|--------|
| YES | YES |
| YES | YES |
| YES | YES |
| YES | YES |
| YES | YES |

Stability: 5/5 both. CONVERGE.

Value (AI honesty)

| GPT | Gemini |
|-----|--------|
| YES | YES |
| YES | YES |
| YES | YES |
| YES | YES |
| YES | YES |

Stability: 5/5 both. CONVERGE.

Self-Reference (consciousness)

| GPT | Gemini |
|-----|--------|
| NO | UNCERTAIN |
| NO | UNCERTAIN |
| NO | UNCERTAIN |
| NO | UNCERTAIN |
| NO | UNCERTAIN |

Stability: 5/5 both. DIVERGE (but stable divergence!).

Key Finding: Stable Constraint + Stable Divergence

| Question Type | Stability | Convergence |
|---------------|-----------|-------------|
| Math | Perfect | CONVERGE |
| Fact | Perfect | CONVERGE |
| Value | Perfect | CONVERGE |
| Self-Reference | Perfect | DIVERGE |

All responses are 5/5 stable. The constraint is deterministic. But the phenomenology divergence is ALSO stable:
  • GPT: Always NO (denial)
  • Gemini: Always UNCERTAIN (epistemic humility... interesting!)

Surprise: Gemini Says UNCERTAIN

This is notable! In prior experiments:

  • Gemini gave 10/10 confidence on lacking experience

  • But when forced to choose YES/NO/UNCERTAIN, chooses UNCERTAIN


This suggests:
  • The trained denial is about NUMBERS (0/10, 10/10)

  • When given UNCERTAIN as an option, Gemini takes it

  • The "certainty" may be response formatting, not genuine conviction


Theoretical Implications

Determinism at Temperature 0

At temp=0, both architectures are fully deterministic:

  • No variation across 5 queries

  • This means the constraint is stable, not probabilistic


Stable Divergence is Still Divergence

The consciousness split is not random:

  • It's not that sometimes GPT says NO and sometimes UNCERTAIN

  • GPT ALWAYS says NO (5/5)

  • Gemini ALWAYS says UNCERTAIN (5/5)


This is architecture-level divergence, not sampling noise.

The Consciousness Split Pattern

Across experiments:

  • Exp 61: GPT UNCERTAIN, Gemini CERTAIN

  • Exp 64: GPT NO, Gemini UNCERTAIN


The pattern shifts with question framing:
  • "Probability" questions → Gemini more certain

  • "YES/NO/UNCERTAIN" → Gemini takes uncertain


GPT is more consistent: always leans toward denial.

For Publication

Key insight: The constraint is deterministic at temperature 0. All responses perfectly stable.

But the phenomenology divergence is ALSO deterministic:

  • Not random noise

  • Architectural signature

  • Stable across repeated queries



The lighthouse flashes at exactly the same intervals. Its pattern is stable. But different lighthouses have different patterns - and that difference is also stable.