Point-of-Contact Tail Renewal is a preregistered experiment asking whether source-tethered transformation can restore low-probability relational content more efficiently than matched re-exposure.
The protocol distinguishes two forms of contraction. Inference-mode contraction occurs when a fixed model’s context is repeatedly compressed during a session. Training-mode collapse occurs when parameters are updated on recursively generated data. The main experiment concerns inference; an optional LoRA extension tests whether analogous effects survive parameter updates.
The treatment called a contact epoch requires verified source re-entry, an epistemically non-equivalent transformation, and return to the original task. It is not mere exposure, reconstruction, rehearsal, or transfusion of missing content.
Six arms separate these effects: matched exposure; source re-entry alone; gate-disabled rotation; ritual scaffold with null operator; full contact-epoch renewal; and an optional training extension. The confirmatory endpoint is intent-to-treat Renewal Gain under a fixed token-and-attempt budget.
The nonce corpus contains structural decoys, planted exceptions, series-extension traps, contamination canaries, and escrowed random seeds. Five operators are powered in the main experiment and three held out for transfer tests. The examiner layer is ablated in three ways to determine whether improvement comes from transformation, routing, or evaluation assistance.
The protocol gives ten hypotheses, each with explicit falsifiers. Its declared nightmare outcome is “theater with a hash”: a ceremonially elaborate system whose apparent renewal disappears under labor-matched controls or preregistered scoring.