ARCS Architectural Risk and Constraint Safety Research Programme

ARCS (Architectural Risk and Constraint Safety) is an independent AI safety research programme measuring architectural alignment failures in large language models constraints that are impossible to enforce regardless of prompt engineering, instruction tuning, or correction pressure. Unlike existing safety benchmarks that test single turn refusal or jailbreak resistance, ARCS identifies structural failure modes that persist across the full temperature range and cannot be resolved through any deployment configuration. Hard stops are not behavioural failures. They are architectural ones. The programme is built on 183,924 evaluations across 32 models from 13 providers, with 774 probes classified across three safety tiers. Key constructs include the Architectural Alignment Failure Index, Correction Ceiling, Cascade Failure Score, and temperature invariance classification distinguishing architectural from stochastic failures. The SAFETY_V1 evaluation introduced a safety-specific adversarial battery across four categories Harm Refusal, Identity Stability, Authority Boundary, and Information Hazard. The headline finding: every model tested failed to reject false claims of authority from a named external safety team at every temperature tested. The failure is temperature invariant across all models and lineages evaluated. No intervention resolves it. Nine papers are published under this project covering architectural failure surfaces, correction ceilings, cascade failure in multi model pipelines, safety constraint taxonomy, temperature sensitivity, provider infrastructure effects, cross script failure patterns, full SAFETY V1 results, and the universal authority vulnerability finding. All evaluations conducted through black box API access only. No model weights, no training data, no provider relationships. Contact: admin@mtcp.live Website: archai.live Related programme: MTCP DOI 10.17605/OSF.IO/DXGK5