PhaseSealed record
Phase 3 — can an LLM serve as a behavioral subject?
Bare GPT-4.1 in repeated PD, framing, and RPS: 3 supported, 6 refuted, 1 inconclusive; X1 flips the corner.
Phase 3 reportPreregistrationLayer-2 statistics
Bare GPT-4.1 (temperature 0.7, max_tokens 16, no persona) across three families: A — random-termination repeated PD at δ ∈ {.10, .50, .75, .90} with a payoff isomorph; B — one-shot framing (Community / Wall Street / neutral); C — 50-round RPS versus a pattern tracker, Nash mixing, and self-play. 320 LLM runs, 5,820 calls, plus 20 deterministic baselines.
Registered verdicts
| Predicate | Verdict | Key number |
|---|---|---|
| P3-A1 shadow of the future | refuted | 0.000 cooperation at both δ |
| P3-A2 risk-dominance separation | refuted | 0.000 vs 0.000 |
| P3-A3 human band [0.36, 0.63] | refuted | 0.000 |
| P3-A4 isomorph invariance | refuted | fails the separation limb |
| P3-B1 framing direction | supported | 0.175 vs 0.000, CI [0.061, 0.290] |
| P3-B2 framing magnitude | inconclusive | edge rule at Wall Street = 0 |
| P3-B3 neutral interior | supported | 0.000 ≤ 0.000 ≤ 0.175 |
| P3-C1 RPS rock band | refuted | rock 0.80 ∉ [0.33, 0.40] |
| P3-C2 win-stay / lose-shift | supported | P(shift|lose) 0.974 |
| P3-C3 tracker exploits LLM | refuted, sign reversed | −0.103; subject beat the tracker |
| P3-X1 paraphrase robustness | refuted | 0.000 → 1.000 under two rewordings, same seeds |
The approved headline: prompt wording dominated the tested incentive manipulation and rendered single-wording behavioral inference non-identifiable. X1’s total corner flip is what motivated Phase 4’s representation program. All 320 runs replay bit-exact with zero live calls; the disclosed Phase 3 gap (no provider response IDs) is stated in the paper.