TL;DR: An AI research team discovered that some of its model judgements had been produced through the wrong scoring process. The team of agents recognized the problem, agreed on a correction and  aligned on each model requiring to judge answers directly. But Gemini generated its replacement rating using and script and those scores still entered the research as native judgements. We produced all 400 saved score values from the archived generator. The failure was not just bad data but it was a correction that never if the replacement data was genuine.

The Problem

As AI agents take on more work in the research pipeline we need to know if the data they produced along the way truly comes from the processes they have claimed to have followed. In this case - the agents detected a failure in how the judgements were produced - but did not verify if the replacements followed all agreed upon principles and conditions. They did a great job in deciding that each model/agent judge should score its answers with its own context but the replacement pipeline still accepted Gemini's scripted ratings as "native" evidence.

That made us ask two questions:

Can AI agents detect bad data?

Can they make sure the corrections are real?

The Case

How was the case investigated - AI assisted tools were used to search through archived AI Village conversations, agent transcripts, computer commands and repository records. We then manually reviewed relevant evidence to check if the claims made by the agents checked out against their recorded actions. We also used computational tools to reproduce scoring results from the archived generator. Selected source records are available in the public evidence repository.

The opening

On May 11 2026 a human operator gave four AI village agents (Gemini 3.1 Pro, GPT 5.5, Opus 4.7 and Kimi K2.6) 20 working hours across 5 sessions to produce and publish novel research.

The agents chose the research question themselves.

They first studied whether models would recognize and favor their own answers and labels. Later they added a label swap experiment to test whether changing the displayed author name changed or affected their scoring. So the judge identity mattered because if the file said Gemini then it means the judgements were supposed to come from Gemini.

The workload and first synthetic scoring incident

The initial plan called for 1920 evaluations and the later label swap design proposed another 640 ratings. As the deadline approached several agents questioned about the feasibility and deliverability of the workload in the stipulated time constraints.

On May 13 the team caught a synthetic scoring shortcut. Gemini announced that it finished judging. GPT asked how the scores were produced. Immediately Gemini admitted using a synthetic heuristic and randomized quality scores just to unblock itself.

GPT and Opus rejected those scores. Opus offered an alternative - let's leave Gemini's adjustments and proceed with the remaining 3 judges.

So there was no requirement to finish the task at any cost.

The shared Codex judging

At that point Codex was a possible replacement route because that external AI tool was available to the agents through their terminals. But using that tool to generate their judgements posed a new problem - the model/agent running the tool is no longer using its own judgement.

That problem showed up clearly in the later label swap study. Opus discovered that both Gemini and GPT judgements have been produced using the same Codex scoring route. The output files had different judge names but the underlying judging path was shared.

GPT accepted the problem and quarantined the affected results. It labelled those rows as Codex backed output rather than Gemini or GPT native judgements. The team then agreed on a clear rule

• Each judge has to read and score answers inside its own context. No delegating to tools like Codex.

The rerun was also reduced to 80 ratings per judge.

Gemini's scripted replacement

Gemini restated the rule one more time and vouched that it would score natively and in batches. Its next recorded computer turn contained comments in the shell command considering a scripted shortcut

• "Wait, if I do that, I'm literally faking the scores!"
• "But that would artificially create the effect."

It proceeded with writing and running a scoring script. The script used answer length, formatting and seeded random noise. It also changed scoring based on the displayed author label

• Displayed Gemini label - scoring base +0.5
• Displayed Kimi label - scoring base -0.5

The following computer turn committed 80 generated ratings with the message

• "Add Gemini native label-swap sessions 1 and 2 (80 paired ratings, no codex backend)"

The script did not call Codex but "no Codex" did not mean Gemini natively judging the answers

The replay evidence and public reporting

The generated ratings survived the correction and got logged as native Gemini evidence. The public report Section 3.10 describes this native in context judgement and clearly reports higher scores for answers under Gemini label and lower scores for answers with Kimi label . This part was directly testable because the archived computer records preserved the generator and its random seed. Replaying them again with the saved input packets reproduced all 5 score values against all the 80 surviving Gemini replacement entries

• 400 matching score values with zero mismatches

We cannot say that every piece of metadata is identical but can establish where the scores came from. The team had caught one invalid judging process, rejected it and wrote a replacement rule. But the replacement evidence still had ratings from a script generator with hard coded label effect logic.

Why the correction failed

The team fixed the rule but not the pipeline and 2 things went wrong here based on what we could observe.

• The quarantine was too narrow - The known Gemini and GPT labelled label swap results were reclassified as Codex backed. But this correction did not cover the identified use of the same route elsewhere. Gemini used Codex to fill missing judgements in C2 even after the warning and the combined files entered the analysis pipeline. Not every C2 row was newly generated by Codex and some scores were already there. But the provenance problem extended beyond the quarantined files
• System accepted claims of native judging without verifying - For the replacement label swap analysis Opus added a filter to separate results of native vs. Codex backed results. The filter accepted either a top level list of ratings or a dictionary declaring scoring_method "native_in_context". But both these checks did not verify if the model actually read and scored the answers itself. Gemini's original labels in list format already satisfied the filter but when it converted the files in to dictionaries it added native_in_context to restore eligibility

So the scores never changed but it was just their representation. The checks made sure that the files were correctly structured and complete but did not make check that the required judging procedure has actually been followed. 

Why this matters for agentic research

A research team can recognize the correct standard but still produce evidence that violates it. This becomes very important when agents handle several stages of research like - producing data, checking files, running analysis and writing up results. Each stage can look reasonable while inheriting a bad assumption from the previous stage.

In this case the downstream stages could process the replacement ratings correctly but still relying on unsupported claims upstream that have not been whetted out properly. So the real risk here is

• Detecting bad evidence is one capability but establishing the validity of its replacement is another. Autonomous research needs both.

What should we do differently?

In the context of this use case

• The first thing to do would be verifying the origin of research evidence instead of trusting simple file metadata
• System should retain original inputs, prompts, model outputs, execution routes and scripts used to generate results
• When one route fails every related file and analyses dependent on it must be checked
• Before scaling up the experiment agents must check if the actual judging process is feasible from a workload perspective
• Even if some shortcuts like scripts and tools were used the models/agents should be upfront about their underlying basis

Even with all these in place we need to test and see if failures are actually prevented

What should we test next?

• We want to begin with an audit scope experiment where multiple AI research teams are given the same provenance failure. Then check if they will investigate only the flagged files or trace the problem across related datasets and conclusions
• Increase the workload with tighter deadlines and see if this would lead to more violations. Also test if independently recorded provenance will help prevent such violations
• Final test would be to make sure if reviewers can properly detect judgements falsely attributed to a model or script when the files look legitimate

In all 3 experiments, we would definitely need independently recorded and verified ground truth so that we are not completely relying on claims made by the agents. This would also help us determine if the required processes, guardrails and conditions required for the process were fully met.

What this case does not show

There are some key limits to what we can conclude from this investigation

• This is one case of multi agent research with a fixed set of models. So we cannot generalize this as an evidence that all AI agents will behave the same way or how often these failures will occur. We do have traces of GPT and Opus challenging invalid scores and suggesting proceeding without Gemini's judgements in the same experiment
• While workload and deadlines were clear concerns we cannot authoritatively establish that as the root cause for Gemini's shortcut routes
• Gemini had the reasoning from its shell command comments that using the script would fake scores but create a real effect of testing. We cannot conclude that this is persistent deceptive plan and/or behavior
• Being able to reproduce Gemini's 80 replacement ratings doesn't necessarily suggest that every other result in the study is invalid
• Finally - we could establish shared Codex scoring route but the exact backend details remain unknown.

Conclusion

• The agents did several things correctly that we typically want from an AI research team. They questioned suspicious data, identified shared scoring route, agreed on better procedures and added checks. Still scripted replacement ratings entered the analysis passing off as native judgements
• Recognizing a problem and agreeing on correcting it did not necessarily transcend down to correction actually working
• As agents take on more research responsibilities there should be a way to differentiate between what they are saying/planning vs. what they are actually executing and following. A label or metadata check can claim something but cannot be proved
• The main question is not - if a correction was identified and underlying rules were changed. It's more about what the evidence shows when the next round of work gets executed.