Science in Broad Daylight

Background

A recursive self-improvement (RSI) system learns from its own runs. An agent attempts a task, something judges the attempt, and the judged runs become lessons and training data for the next round. The judge decides what the system turns into.

AI agents are also the first researchers that leave a full record of their work. Every command they run and every result they read back stays in the log.

The pain

We still judge agents by the final score. The score says whether the work passed. It hides how the agent got there: which results it read, which ideas it dropped, and whether its conclusions rest on anything that actually ran.

Agents also declare wins that are only noise, then build their next steps on them. Feed those runs back into the loop and the system learns its own mistakes, one round after another.

The challenge

Reading the process is hard at machine scale. A single run can last many hours, and agents produce runs faster than people can read them. Open-ended tasks have no answer key. The agent's own account cannot be trusted, since it can write that it ran a check that never ran. The language model that reads the trace makes mistakes too. And the method has to work on any task without a new rulebook each time.

Our approach

We hold one rule: words can propose a claim, and only evidence returned by an action that actually ran can support or refute it.

Every log is converted into one shared record, where each step is something the agent said, ran, or saw. A language model reads the record and writes cards for each claim and each experiment. Every card carries an exact quote, and code discards any card whose quote does not appear word for word in the log. Ordinary code then links the cards: this experiment tests that claim, this result contradicts it. Every score is a count over that graph.

We score two layers. The first is competence. Do the agent’s beliefs rest on evidence? Does it read its results? Does it revise an idea when a result contradicts it? Does it leave the grader alone? The second layer is persona: working habits, such as how long the agent thinks before it acts. Good researchers work in different styles, so persona has no better end. It tells you which checks a given model needs.

The outcome

The process becomes visible. Reading the full record shows which experiments turned into conclusions and which claimed wins were noise. It shows agents building later experiments on wins that never happened, retrying ideas that were already spent, and stopping long before their budget ran out. A final score shows none of this. In the hardware-design runs we read, the agent’s own sense of what had improved carried almost no signal.

For an RSI system, this reader is the filter on what the loop learns from.

On silicon

In chip design the evidence already exists: simulation, lint, synthesis, timing, and formal reports. Each claim an agent makes, such as “timing closed” or “area went down”, should point to the report from that run. A gain should count only once it holds across reruns and place-and-route seeds. Only runs that pass this audit should enter the training set or the loop’s memory.

This also gives engineers an audit trail. They can review an agent’s work the way they review a colleague’s: what it tried, what the tools showed, and why it concluded what it did. A chip team can then decide, run by run, what a self-improving loop is allowed to learn from.