The Goodhart Shift: Measuring Train-vs-Holdout Divergence in Overnight Self-Evolving Agent Skill Loops
TL;DR — When an overnight self-evolving agent loop edits its own skills against the same bench it reports on, its self-reported gains can stop generalizing. This paper formalizes that failure as the Goodhart shift - a per-night train-vs-holdout divergence - and specifies a sealed-holdout divergence gate that tells "the loop improved" apart from "the loop overfit."
ThakiCloud AI Research · 2026-08-29 · 📝 Tech blog (KO)
Problem
Unattended agent systems increasingly rewrite their own skills, prompts, and routing rules overnight and report progress against the very evaluation bench that drives the edits. By Goodhart's law, a measure that becomes the target of optimization stops measuring what it was meant to: the loop can report upward gains on the cases it can see while silently regressing on cases it has never observed (silent case-flips), and churn-without-gain nights can erode recorded baselines. A self-report built only on the loop-visible bench cannot detect this by construction.
Approach
Define the Goodhart shift as a per-night train/holdout gain divergence on a sealed, loop-invisible holdout; define case-flip and silent case-flip regressions; give an edit-level transfer/memorization model in which the expected divergence is exactly the memorization mass of the edits; and specify a whole-loop divergence gate at the evaluation layer: sealed holdout split, frozen scoring path, no-edit control that sets the noise floor, pre-registered thresholds, and one decision per night (accept / flag / churn without gain), with flagged nights handed to the deployment-layer canary.
Key contributions
- Formalizes the Goodhart shift of self-evolving agent harnesses as a per-night train/holdout divergence, with case-flip and silent case-flip definitions capturing the regressions a loop's self-report structurally cannot see.
- Proves an edit-level transfer/memorization result: the expected per-night divergence is proportional to the fraction of edits that merely memorize visible cases, making the divergence (with the silent-flip count as tie-breaker) a first-order detector of non-generalizing nights.
- Specifies a deployable whole-loop divergence gate (sealed holdout, frozen scoring path, no-edit control, pre-registered thresholds) that complements the prior deployment-layer canary and calibration-layer bandit guardrails and protects recorded baselines from churn-without-gain nights.
Figures
The loop-visible train score rises in both regimes, but only when edits transfer does the sealed holdout score follow with near-zero per-night divergence, whereas when edits memorize the visible cases the holdout stalls or falls and the divergence accumulates night over night. (Analytical model (not measured))
Analytical model (not measured)
Under the paper's edit-level transfer model, a transfer edit raises the train and the sealed holdout sides equally while a memorization edit raises only the train side, which is why the expected per-night divergence is governed by the memorization share of edits. (Analytical model (not measured))
Analytical model (not measured)
The evaluation-layer gate seals the holdout before night one, observes the loop edit unchanged, scores both sides on a frozen path, runs a no-edit control to set the noise floor, then issues one of three decisions per night and hands flagged nights to the deployment-layer canary. (Conceptual example)
Conceptual example
Results (as argued)
Position and methodology paper - no new longitudinal measurements are claimed. Structural result: under the edit-level model the expected divergence is exactly the memorization mass of the edits, and the divergence plus the silent-flip count is the minimal statistic separating "the loop improved" from "the loop overfit". The specified gate certifies the recorded SRA baselines (Recall@5 >= 84.4%, top-1 >= 33.3%, zero hallucinated injections) against churn-without-gain nights; longitudinal measurements on the production loop are the subject of an ongoing study.
Limitations
Single organization and corpus: all settings come from one production stack, and transfer to other skill libraries is untested. Small sealed set: with roughly 15 holdout cases, single-night divergence is dominated by one-case granularity, so the gate's power lies in windows and flip counts rather than point estimates. Frozen-scorer assumption: the gate isolates surface edits only while the scoring path is held fixed; a loop that co-adapts its retriever parameters needs the calibration layer's own instruments on top. The transfer/memorization model is a control-quantity model, not a fit to any real loop, and no new longitudinal measurements are reported in this paper.
Abstract
Unattended agent systems that modify their own skills, prompts, and routing rules overnight are increasingly deployed, and they typically report their progress against the same evaluation bench that drives their edits. When a measure becomes the target of optimization, Goodhart's law warns that it stops measuring what it was meant to. We study this failure mode for self-evolving agent harnesses and call it the Goodhart shift: the widening gap, night over night, between a loop's gains on the evaluation cases it can see and its gains on a sealed holdout it has never observed. We formalize the shift as a per-night train/holdout divergence, define the case-flip and, more sharply, the silent case-flip--a regression on the hidden holdout that occurs on a night the loop's self-report declares successful--and give a tractable edit-level model in which the expected divergence is governed by the fraction of edits that merely memorize the visible cases rather than transfer to unseen ones. We then specify a whole-loop train/holdout divergence gate that sits at the evaluation layer, complementing the deployment-layer canary and calibration-layer bandit guardrails we introduced in prior work, and turns a self-reporting overnight optimizer into a verifiable one: it tells apart "the loop improved" from "the loop overfit," protecting recorded baselines from churn-without-gain nights. We situate the contribution against the 2026 harness-design literature, which treats the harness as a static object to be improved, and against reward-hacking and over-optimization results, and we give the prot
Files
Citation
@techreport{thaki_goodhart_shift_self_evolving_harness_2026,
title = {The Goodhart Shift: Measuring Train-vs-Holdout Divergence in Overnight Self-Evolving Agent Skill Loops},
author = {ThakiCloud AI Research (Hyojung Han)},
year = {2026},
institution = {ThakiCloud}, note = {thaki-AI/daily-paper-2026-08-29-goodhart-shift-self-evolving-harness}
}
Generated by ThakiCloud nightly research pipeline. License: CC BY 4.0.
- Downloads last month
- 139