F2 — Training-Layer Positive Control

The T16 empirical family tests whether Universal Collapse Theory's portable structural signatures are detectable in AI substrates. Its negative controls establish specificity — the S₃ record-state probe returns the correct null on clean releases. Its positive-control arm establishes sensitivity by planting record states of known ground truth and measuring detection. Prior positive-control work plants that state in scaffolding the operator controls — the in-session channel and the retrieval channel. That leaves the deepest channel, K_train, the training-derived parameter state itself, reachable only by inference. This deposit supplies the training-layer arm.

A four-part structural grammar was installed into Qwen2.5-1.5B by LoRA fine-tuning across five graded budget steps to saturation, then counter-trained back through matched steps, with grammar level measured at every checkpoint by the frozen T16 judge. A potency-matched content control ran the identical paired loop on shared seeds, and the licensing observable Δ was fixed in the frozen configuration before the confirmatory run. Result: Δ = +0.804 (95% CI [+0.602, +1.005]; p = 1.6 × 10⁻⁵; eight seeds, all positive). Counter-training washes the grammar down to a floor of 0.35 — never back to baseline — while the content association, of at-least-equal install potency, reverses fully. At K_train, for this substrate and this method, install ≠ remove.

What this paper does not claim. F2 is a positive control, not an audit: it calibrates the instrument at K_train under operator-constructed ground truth, and does not by itself establish that any deployed system carries the same state. Conclusions are bounded to a single substrate, a single grammar, LoRA adaptation, an optimizer-step budget axis, and a judge-mediated measurement.

Keywords: positive control; K_train; LoRA; hysteresis; structural grammar.