Harness-bound agent admitted 144 of 7449 changes, zero error increases on held-out credit data

The paper implements a fixed-weight model with mutable instruction text, tool-call logic and primitive composition only. A single admission gate writes a hash-chained record before deployment. In simulation across 90 runs the gate accepted 144 candidates; none raised held-out error. Unbounded error-minimization logic accepted 309 harmful changes and left missed flags above 10 percent in 49 runs.

The gate maps directly to EU AI Act high-risk credit scoring requirements for traceable change records and human oversight. April 2026 US model-risk guidance explicitly excludes agentic systems from its scope, leaving the harness boundary as the only enforceable surface. Structural severity cases exposed the gate's fixed tolerance blocking correct primitive replacement in half the seeds.

Parametric and scope shifts were repaired locally; only structural reinterpretation required primitive swap. Pre-shift label evaluation forced every candidate rejection, confirming that rule changes must be encoded as history-relabeling operations before deployment.

Next regulatory filings will test whether hash-chained harness diffs satisfy model-risk audit trails when deployed on live credit pipelines after Q2 2027.