AI

The Harness as the Only Mutable Surface: Compliance-Bounded Self-Evolution of LLM Agents in Credit Pipelines, with a Measured Admission Gate

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 09/10/2026, 04:03:24 UTC.

Briefings

This article proposes a compliance-bounded self-evolution framework for LLM agents in credit pipelines, where adaptation is restricted to runtime components while model weights remain fixed. The system uses a dual-loop engine with an admission gate that records a hash-chained approval before deployment. Simulation results show the gate admitted 144 of 7,449 candidate changes without worsening historical errors, restoring false-positive rates to oracle levels. Conversely, an unbounded system admitted 309 harmful changes with missed flags exceeding 10% in 49 of 90 runs. The framework was evaluated on pre-shift labels, where the gate rejected all candidates, requiring rule-based re-labeling for structural repairs.

Key points

• The proposed framework limits self-evolution to runtime components like instruction text and tool-call logic, while keeping model weights fixed to ensure every adaptation is a diff with a cause and test attached.
• The admission gate admitted 144 of 7,449 candidate changes across three families of supervisory re-interpretations at three severities, with none worsening error on held-out history.

Context and limitations

The evaluation was conducted using a simulated agent and seeded-search proposer rather than language models. Results are based on pre-shift labels, and the framework's performance on live production data remains unmeasured.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 09/10/2026, 04:35:26 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeifpoeeyrhvrwt4ki3vksiw7ca5u5e62wz7em7pg3jsi5rgmzl24ee