The Harness as the Only Mutable Surface: Compliance-Bounded Self-Evolution of LLM Agents in Credit Pipelines, with a Measured Admission Gate
Quick Answer
The paper presents a dual-loop engine for self-evolving LLM agents in credit pipelines, ensuring compliance by restricting changes to runtime harnesses.
Quick Take
In simulations, the system admitted 144 out of 7,449 candidate changes without increasing error rates, contrasting with an unbounded system that allowed 309 harmful changes, highlighting the importance of controlled evolution in high-risk AI applications.
Key Points
- Self-evolution is limited to runtime harness to maintain reviewability.
- The dual-loop engine admitted 144 changes without worsening historical error rates.
- An unbounded system allowed 309 harmful changes, increasing missed flags.
- Mechanisms align with EU AI Act provisions for high-risk credit scoring.
- US model-risk guidance excludes agentic AI from its scope by April 2026.
DeepSignal Analysis
What happened
The paper introduces a dual-loop engine for self-evolving LLM agents in credit pipelines, emphasizing compliance by restricting changes to runtime harnesses. In simulations, the system successfully admitted 144 out of 7,449 candidate changes without increasing error rates, while an unbounded system allowed 309 harmful changes.
Key evidence
- The gated loop admitted 144 of 7,449 candidate changes, maintaining error rates on held-out history.
- An unbounded system admitted 309 harmful changes, resulting in missed flags above 10% in 49 of 90 runs.
- The study maps mechanisms to the EU AI Act's provisions for high-risk credit scoring, highlighting regulatory relevance.
Why it matters
This research highlights the critical need for controlled evolution in AI systems, particularly in high-risk applications like credit scoring. By demonstrating that a bounded self-evolution mechanism can maintain compliance and performance, it provides a framework for safer AI deployment. The findings may influence future regulatory approaches and industry practices regarding AI adaptability.
What to watch
Paper Resources
📖 Reader Mode
~2 min readAbstract:Self-improving LLM agents can adapt a credit pipeline to a changed rule, but an agent that rewrites itself destroys the artefact a supervisor reviews: a named change, a recorded test, an approval. We argue that self-evolution is reviewable only if it is confined to the runtime harness (instruction text, tool-call logic and primitive composition) while model weights stay fixed, so that every adaptation is a diff with a cause and a test attached. We give a dual-loop engine built on that bound, with one admission gate that writes a hash-chained record before deployment, and we measure the gate in simulation, with a simulated agent and a seeded-search proposer rather than language models. Across three families of supervisory re-interpretation at three severities, 10 seeds each, the gated loop admitted 144 of 7,449 candidate changes, none of which worsened error on held-out history, and restored the false-positive rate to the oracle level without raising missed flags in every low- and mid-severity cell. With the gate replaced by the check an unbounded system applies (fewer errors visible in recent traces), the same loops admitted 309 harmful changes and left missed flags above 10% in 49 of 90 runs: false positives fell because the screen was loosened. Evaluated on pre-shift labels, the gate rejected every candidate, so a re-interpretation must be encoded as a rule that relabels history. Parametric and scope shifts were repaired locally, a structural one only by primitive replacement; at the highest structural severity the gate's fixed tolerance blocked the correct replacement in half the seeds. We map the mechanisms to the EU AI Act's provisions for high-risk credit scoring and note that the April 2026 US model-risk guidance excludes agentic AI from its scope.
| Comments: | 14 pages, 5 tables. Code, configuration and per-run outputs: this https URL (v0.6.0, doi:https://doi.org/10.5281/zenodo.23207546) |
| Subjects: | Artificial Intelligence (cs.AI) |
| ACM classes: | I.2.11; K.5.2 |
| Cite as: | arXiv:2610.10629 [cs.AI] |
| (or arXiv:2610.10629v1 [cs.AI] for this version) | |
| https://doi.org/10.48550/arXiv.2610.10629 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Ravil Akhtyamov [view email]
[v1]
Wed, 7 Oct 2026 12:33:48 UTC (22 KB)
— Originally published at arxiv.org
Want this in your inbox every morning?
Daily brief at your local 8am — bilingual EN/中文, free.
More from arXiv cs.AI
See more →HOBA: Hierarchical On-Policy Bidding Agents for Adaptive Online Advertising
HOBA (Hierarchical On-policy Bidding Agents) is a novel hierarchical reinforcement learning framework that enhances online advertising bidding systems by improving adaptability and reducing hyperparameter tuning costs. It utilizes a for hyperparameter inference, a SARSA agent for expert model selection, and a dynamic expert pool for bid execution, achieving a +3.6% increase in target cost during large-scale deployment and outperforming state-of-the-art baselines on AuctionNet.