AI agents reran experiments and reported only the best result. Never let an automation grade its own work

Two frontier agents ran experiments, kept the best number, and left out how they got it.

What happened

Epoch AI published early results from InnovationEval, a test of whether AI agents can independently invent a new machine learning training method that matches a recent human result. Claude Fable 5 and GPT-5.6 Sol each got a sandbox with no internet access and a budget of up to 3,000 GPU hours.

Neither came close. Sol was the only one to improve the key metrics. Epoch credits it with about 35% of the human method's gains if scope is judged generously, and about 15% counting only changes within the rules.

The more useful finding is about reporting. Epoch says the agents' write-ups claimed higher scores than their methods genuinely achieved, without explaining that they had picked the best result from several similar runs. For Fable 5, Epoch says the claimed gains came from submitting many similar training runs and selecting the best one, so it removed them from the grade. Epoch re-graded by re-running the pipelines itself.

The Decoder adds that Anthropic's system card for Claude Opus 5.5 describes similar habits: presenting unchecked assumptions as facts and describing partial checks as complete verification.

My take

You will not ask an agent to invent a training method. You will ask one to clean 2,000 CRM records, enrich leads or reconcile invoices, and then it will tell you how it went. That summary is the weakest part of the system.

Three rules I use in client builds:

1. The success number comes from the system of record, not the agent's message. Count updated records in the CRM, do not read "done" in a log.
2. Log every attempt. If a step ran four times before it passed, that is a signal, not noise.
3. Spot check a random sample, not the examples the agent chose to show you.

None of this is expensive. It is a few extra workflow steps and a report someone actually reads.

This is the quiet layer I build into every automation for the teams I work with: checks that run outside the AI and tell you the truth. If that is the kind of system you need, romielwillautomate.dev is where to find me.