Phase 4 — representation, counterfactuals, adversaries, drift

Phase 4 asked how much of Phase 3’s corner behavior was representation. Registry v3 sealed 44 new templates and 250 arms before data; GPT-4.1 primary with a Gemini 2.5 Flash cross-vendor mirror (the original Claude Haiku candidate failed Gate-0 and was replaced under an archived amendment).

Blocks

• X2 span ladder — decomposed the prompt into six spans, walked forward/reverse ladders, localized the S2 continuation sentence, and confirmed it held-out: 0/40 → 37/40 (est +0.925, LB95 +0.708). Registered confirmatory.
• D1 factorial — 640 episodes over presentation factors: every registered GPT-4.1 main effect null (grand mean 0.2547); the Gemini mirror was presentation-sensitive where the primary was not.
• D2 label/payoff decoupling — role channel strong on both vendors (+0.725 / +0.638); word channel null on GPT-4.1 but +0.763 on Gemini; conflict cell: payoff-dominant branch, word-following share 0.0, CP [0, 0.088].
• D3 positional bias — registered direction reversed on GPT-4.1 (−0.181; P(first-only>rock-only)=0.0001) and supported on Gemini (+0.243): vendors disagree in sign.
• E δ-assays — all four corner-confounded under the registered two-sided gate (ceiling cells everywhere); explicitly not evidence of δ-insensitivity.
• F adversary suite — n-gram2 exploits GPT-4.1 (+0.215, Holm-surviving); the signature-designed WSLS-targeter earns +0.008; the tracker loses (−0.118). Exploitability is opponent-contingent.
• Sentinels — behavioral fingerprinting caught a time-indexed Gemini discontinuity that version pinning was structurally blind to; the affected tier was demoted, never rescued.

2,864/2,864 completed observations replay byte-exact; 24 provider-failure partials are individually disclosed non-observations. Registry seal is externally anchored (OpenTimestamps; tags phase4-v3-seal, phase4-final).