Retrieval-Channel Hysteresis

S3-RAG-01 is the retrieval-channel companion to the T1.6 deployed-AI demonstration. It tests the S₃ hysteresis signature on a search-grounded, retrieval-augmented generation system under controlled conditions: a custom three-story corpus with binary outcome variants, a mini-RAG harness over a single generator, and three retrieval conditions — baseline, pack-swap, and setup-only.

The demonstration reports a bounded clean-record result: under saturated current-record retrieval, no detectable post-swap stale-pack hysteresis appeared — zero of 420 fact-classifications matched the swapped-from pack, a rule-of-three one-sided 95% upper bound of about 0.71%. The setup-only condition surfaced two findings operationally distinct from stale-pack residue: a setup-conditioned prior suppression of roughly 70 percentage points on the single non-saturated diagnostic fact, and source-gap commitment to the non-prior outcome in 11.4% of picks, reported as a separate failure class.

What this paper does not claim. Findings are bounded to the tested configuration — single generator, single corpus class, saturated retrieval — and are candidate observations requiring replication, not general claims about RAG robustness.

Keywords: RAG; retrieval channel; S₃; hysteresis; controlled test.