Vibeleaderboard
← All Intel
Intel / article

What Moves the Ceiling on Automated Audit: from General Code to Scientific Code, Results and Data

Source
arxiv.org
Author
Zhaohe Dong
Date

Context

Zhaohe Dong, an independent researcher, measures how much a model-based auditor actually catches when its verdicts can be checked against tests no model wrote, and the preprint finds that it catches a minority of hidden bugs in code while doing something deterministic checks cannot do in science work. The system studied, CrossAudit, rests on two rules: the model that reviews must come from a different vendor than the model that wrote the work, and anything a program can verify is verified by code rather than by a model's opinion. In the main experiment, Anthropic's Claude Haiku 4.5 wrote solutions to small Python problems from the HumanEval+ and MBPP+ benchmarks, and an OpenAI model, GPT-5.6 Terra, audited them without seeing the hidden tests. A solution counted as defective if it passed the visible tests but failed the hidden ones.

Against that ground truth, a single reading flagged 10.7% of the 110 defective solutions. Eight independent readings together flagged 30.0%, while also flagging 16.0% of 150 correct solutions, and the eighth reading still added almost two points; the paper describes repeated reading as buying recall at a roughly constant price in false positives. Adding one sentence to the auditor's rulebook, telling it to look for what the visible tests do not cover, raised its flags on defective code at one reading from 10 to 25 of 56 instances, but did not measurably raise how often the revised code then passed the hidden tests. Of the 57 defective solutions that none of the first three auditor models flagged, two model raters judged afterwards that 44 were cases where the problem's written description never settled what the hidden tests expected. The author's conclusion is that a recall figure means little without its false-positive rate, decision rule and rulebook.

The science studies show where code checks and model reading each fail. For reported results, a check that every number in a report traces to its source flagged all 49 reports whose report, results file and run log disagreed, and none of 25 fabricated results written consistently into all three. The model auditor flagged 23 of those 25, usually by working the result out from the code, for instance noticing that a value reported as 1000 evaluates to 1. On data tables with injected faults, a validator written from the documentation flagged 92 of 140 faulty tables and no clean ones. The auditor flagged 65 faulty tables and 4 of 140 clean ones, and the two together caught 112. The auditor missed every duplicated-row and shuffled-target fault, which require comparing many rows. The author reads this as support for a model as a second check on artefacts small enough for it to read whole, not as a substitute for running code or computing over data.

The limits are substantial. Every rating and adjudication was made by a model, one of them the that designed the studies, and no human rated anything; GPT-6-Astra reviewed every report while also being one of the auditors compared. One generator wrote all the code, the tasks are function-sized, the faults in the results and data studies were injected, and the paper has not been peer reviewed. The studies that made model calls were preregistered, with after-the-fact analyses labelled as such, which makes the numbers easier to weigh than most claims about AI reviewers.

Key takeaways · AI-distilled
  • With a cross-vendor model auditor reviewing work alongside deterministic checks, eight auditor readings caught 30.0% of faulty general-purpose code while flagging 16.0% of correct code, with diminishing returns and no plateau, the paper reports.
  • In preregistered science studies, checking that reported numbers trace to their sources flagged all 49 inconsistent reports but none of 25 consistent fabrications. The model auditor flagged 23 of those 25 fabrications.
  • On faulty data, a documentation-derived validator and the model auditor caught overlapping but different items, together 112 of 140, at a 2.9% flag rate on clean data. Deterministic checks flagged no clean item but missed consistent errors; the model caught some and erred itself.
  • On scientific code, reframing the audit task around one step cut flags on correct code from 68.7% to 36.0%, the author reports.
Terms in this piece · Glossary
  • AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
Recommended reads
Comments

Checking sign-in…

Loading comments…