lin-0225

18.8 Four foundry proofs of concept

Foundry	SID result	Preserved outcome	Claim boundary
Trade corridor	overlap residual \(6.64\times 10^{-17}\) after one tangent lift	global RMSE \(0.0918\); promotion \(0.966\)	synthetic nonmilitary predictive system
GLP-1 replay	residual \(0\), versus \(0.0877\) independent and \(0.00202\) glue loss	perfect global field and promotion accuracy	structural consistency, not medical truth
10-K replay	residual \(0\), versus \(0.0290\) independent and \(0.000712\) glue loss	perfect global field and promotion accuracy	frozen 46-claim audit, not financial advice
Product-feedback replay	exact compatibility; \(0.976\) witness retention; full context coverage	perfect contract and promotion accuracy	aggregate-interface audit, not review truth

Table 18.2 Thirty-seed proofs of concept retain prediction and promotion metrics separately from descent residuals.

Experiment: SID frozen-foundry replays

These are thirty-seed controlled replays from frozen local representations; reference labels score the run but never enter the correction. The GLP-1, 10-K, and product rows test descent plumbing rather than domain truth. The trade-corridor benchmark is particularly diagnostic: a tuned gluing penalty slightly improves local prediction over SID but leaves a nonzero witness, while a robust median improves global prediction without repairing compatibility. Aggregation, prediction, and descent are different operations.

Admission contract

A SID repair is admitted only when its typed witness is identifiable, the targeted compatibility or effectivity obstruction decreases beyond solver and statistical uncertainty, shared corrections agree across the cover, declared predictive and promotion safeguards pass, and no new obstruction appears. Otherwise the system damps, quarantines, gathers evidence, or revises the cover.