FCL-S Inference Kernel (MVP+): Inference-Time Stop/Attribution Governance for Non-Commitment under Missing Evidence (UST/SCB/AF) Inference-time governance for stoppable, non-drifting outputs

This deposit provides an implementation-spec MVP+ reference for the FCL-S (False-Correction Loop Stabilizer) Inference Kernel, a minimal inference-time governance layer that makes advanced AI stoppable and non-driftingbefore scaling capability. Problem addressed. In the post-scaling regime, systems can produce highly plausible coherence that locks in errorsand amplifies misinformation through interaction-level self-correction. This deposit specifies a wrapper-style kernel that constrains when a generator may emit a committing answer. Core components (UST/SCB/AF). UST (Unknown Stable Terminal): “Unknown” as a governed terminal state. Returns structured missing slotsand safe next steps for verification. SCB (Stop/Correction Boundary): a stop boundary evaluated in pre and post phases (dual-wall envelope) that blocks ungrounded commitment, drift, or contradictions and routes to UST. AF (Attribution Fixation): structurally separates primary definitions (author-defined) from external observations/inferences, preventing authority-bias drift. MVP+ extensions. Confidence mismatch detection using token-confidence proxies (e.g., avg_token_logprob) under a heuristic threshold (τ = −2.5). Semantic drift scoring (drift_score) against primary definitions (threshold 0.35). Unresolved-UST gating (open_ust) to reduce false-positive contradiction triggers. MVP++ specification notes: quiet-drift detection via embedding stack, and governance strength control via strict_mode. Evaluation protocol (summary). This deposit includes a compact test plan and metrics covering ungrounded commitment, confidence mismatch, authority drift, semantic drift, long-context failure, and cumulative agent errors.