A Planted-Structure Testbed for Auditing Second-Order Modulation: Positive Controls, a Falsified Pre-registration, and What They Imply for an Archived Audit

An earlier audit of an archived simulation reported that same-target threshold sums, not chain-mediated interactions, produce almost all “strong” synergistic pairs, and that a substituted-intermediate comparison found no intermediate-specific advantage. Those are negative results from a procedure that had never been tested against a known answer. This note builds a small synthetic system in which a relation can multiplicatively modulate the plasticity of another relation, with the modulation graph planted and therefore known, and uses it as a positive control for three audit statistics (an update-level intervention score U, a loss-level pair synergy S, a substituted-modulator statistic C) and a cycle-count proxy for reflexive organisation. A pre-registered version (v2.1, hard-clipped gain) was falsified on its confirmation seeds. A revised design (v2.3, smooth gain, new seeds; revised after seeing the v2.1 confirmation data) recovered planted outside-column edges with U (AUC 1.000 in 11 of 12 cells, 0.997 in the twelfth, no false positives) and separated modulated pairs by |S| (AUC 0.86–0.99), but signed S and C did not separate them (AUC 0.43–0.50) because modulation interactions have mixed sign; cycle detection worked for moderate coupling and degraded at strong coupling, so the pre-registered usability rule is still triggered in 2 of 14 cells. Applied back to the archive: archived chain interactions are one-signed (17,029 positive, 0 negative of 1,285,200 chain pairs), and a planted intermediate-specific amplification of 2x or more would have been detected in 100% of planted chains at a 0.7% false-positive rate. Unpublished draft, not peer reviewed; synthetic systems only; it does not prove or support the conjecture of the research programme. Companion to doi:10.5281/zenodo.23006539 and doi:10.5281/zenodo.23192515. The bundle (five zip parts; extract into one folder) reproduces every quoted number with verify_paper3.py (48 checks). AI use: the simulations were written and executed by an AI coding model under written specifications and audited by a second AI system; the author is responsible for all content.