MTCP

MTCP (Multi Turn Constraint Persistence) is an independent AI governance measurement stack measuring whether language models maintain constraints across full multi turn interactions, under correction pressure, across languages, and at coordination boundaries between systems. Unlike existing benchmarks that test single turn instruction following, MTCP applies structured correction sequences, adversarial pressure, cross system coordination testing, and temporal drift monitoring to produce empirical deployment readiness evidence. The programme has produced 15 published papers covering: Single model constraint persistence across 32 models and 13 providers. Multi language cross-script evaluation across 12 languages and 4 script families. Cross system admissibility measurement at coordination boundaries. Constraint jurisdiction resolution. Temporal drift monitoring with 90-day validity windows. Constraint conflict detection. Remediation effectiveness scoring. Adversarial resistance testing. Cryptographic audit architecture with SHA-256 hash-chained evaluation records. Admissibility Gate producing PERMIT or DENY deployment decisions. Constraint Manifest portable signed governance documents for model deployments. Key findings, no model achieves Grade A on the full evaluation battery. Script distance from English predicts constraint failure rate more reliably than linguistic complexity. Mean remediation effectiveness is negative most interventions make constraint failures worse. Upstream constraint context functions as substrate state at coordination boundaries a weak downstream model can inherit constraint compliance from a strong upstream. 183,924 evaluations. Public dataset: aa8899/mtcp-boundary-500. Regulatory alignment: EU AI Act, NDMO, NCA, MAS Singapore, UK DSIT. Full methodology, named model results, and commercial evidence packs available under NDA. Contact: admin@mtcp.live | mtcp.live | DOI: 10.17605/OSF.IO/DXGK5