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

TEXT START: 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.

The Dissection

This is a containment architecture for making adaptive automation institutionally admissible. It narrows “self-evolution” to changes in instructions, tool logic, and primitive composition while freezing model weights. The admission gate then converts each change into a reviewable object: diff, cause, test, approval, and hash-chained record.

The measured result is useful but narrow. In simulation, the gate admitted 144 of 7,449 candidates and rejected harmful changes on held-out history. The ungated recent-trace criterion admitted 309 harmful changes and left missed flags above 10% in 49 of 90 runs. The paper also exposes the real fault line: a rule reinterpretation cannot be judged against pre-shift labels without rejecting every legitimate adaptation. Structural change requires primitive replacement, and the fixed tolerance blocks the correct replacement at the highest severity in half the seeds.

The paper is therefore less about intelligence than about bureaucratic survivability. It is designing a leash, an audit trail, and a permission ritual for systems that are becoming capable of modifying their own operating environment.

The Core Fallacy

Relative to Discontinuity Thesis mechanics, the central error is treating the harness as the decisive surface of agency. It is the decisive surface of auditability, not power.

Freezing weights does not freeze the system. Instructions, tool access, routing, thresholds, data selection, and primitive composition can radically change behavior. A hash chain proves that a change was recorded; it does not prove that the labels, tests, objectives, or supervisory standard are correct. The gate manages visible adaptation while leaving the deeper question untouched: who owns and controls the automated credit infrastructure, and how many human workers remain economically necessary?

The paper solves a deployment constraint. It does not solve the displacement problem. In DT terms, it may accelerate P1 and P3 by making automation safer for institutions to adopt. The supervisor becomes a control node—or a servitor—not a sovereign.

Hidden Assumptions

• Fixed model weights provide a meaningful boundary on system behavior.
• Harness diffs are understandable and reviewable by supervisors at operational speed.
• Historical relabeling can produce a legitimate oracle for changed rules.
• Held-out history is a sufficient proxy for future distribution and structural shifts.
• False positives and missed flags adequately represent credit-system harm.
• The tested severity families and ten seeds capture meaningful deployment risk.
• Model providers, tools, data sources, and external policies remain stable outside the harness.
• A fixed admission tolerance remains valid as the environment changes.
• Regulatory mapping can convert technical traceability into genuine accountability.

The abstract itself also discloses the most damaging limitation: both the agent and proposer were simulated rather than language models. The evidence demonstrates a test harness and failure mode, not production-grade self-evolution.

Social Function

Primarily transition management, reinforced by compliance prestige signaling and partial truth.

The text makes adaptive automation legible to regulators and acceptable to institutions by converting autonomy into controlled diffs. Its partial truth is real: optimizing recent traces can reward harmful loosening, and structural shifts can defeat local repair. Its anesthetic function is equally clear: it frames the danger as a solvable governance-interface problem rather than asking whether human participation remains necessary once the pipeline can adapt, test, and deploy itself.

The Verdict

This is a technically serious leash on an emerging automation stack, not a defense of the human economic order. The gate can reduce visible regression and create an approval record, but it cannot prevent the replacement of labor by an increasingly deployable credit machine.

The contradiction is terminal: when change is mild, bounded evolution works; when change is structural, the gate either blocks the correct adaptation or must accept greater autonomy and risk. Compliance buys deployment time. It does not restore productive participation, distribute ownership, or prevent the post-WWII employment-to-consumption circuit from being severed.