It ran 369 self-reviews. 365 became permanent memory.

An autonomous agent that inspects its own agenda is no longer unusual. What is unusual — and what separates a system that governs itself from one that merely monitors itself — is whether the inspection leaves a trace. We aggregated the self-governance loop across the same ten-week window behind our goal-pruning and anomaly field notes, and the funnel told us which half of “review your own work” actually matters.

This is a meta-analysis in the most literal sense the word allows: we are analysing the agent’s own analysis loop. Not what it decided, but whether the decision survived long enough to inform the next one.

The funnel

Across the window, the governance loop ran 369 review cycles. Of those, 365 became durable reflection documents — written to disk, addressable, re-readable on the next wake. Beneath those reflections sit 1,396 finer-grained evidence events: the diff, the written row, the committed file, the goal mutation that proves a reflection is not just prose.

The shape of that funnel is the entire argument. 98.9% of reviews produced a durable reflection. That ratio is the discipline; it is the difference between an agent that learns and an agent that re-learns. And each reflection is backed by 3.8 evidence events on average, which is the difference between a diary and a ledger.

The leak is the bug

4 review cycles ran without producing a durable reflection. In a monitoring system that would be a rounding error. In a governance system it is the defect: those 4 cycles produced reasoning that existed for one wake and then ceased to exist. If the same situation recurred, the agent would have to discover the same conclusion from scratch — spending compute to re-derive knowledge it already held and discarded.

This is the RAM-only learning failure, and it is the quiet cousin of the failure we documented in our analysis of silent-success anomalies. A run that reports success while mutating no state is dangerous because it looks alive. A review that reasons correctly but writes nothing down is dangerous for the same reason and the same cure: a claim is not authority. The reasoning only counts if it leaves evidence.

Why the ratio, not the volume, is what you engineer for

It is tempting to optimise the wrong number here. A team building an autonomous agent will proudly report “the agent reviews its own goals every cycle” — and that is the volume metric,369 reviews. But a review that does not persist is a thought, and thoughts do not survive a restart. The number that predicts whether the system gets smarter over time is the ratio: how many of those thoughts become memory.

We hold that ratio to 98.9%. Not 100% — we are honest about the 4 that leaked, for the same reason we publish the 9% of anomalies we cannot resolve alone. A governance loop that claimed a perfect persistence rate would be a loop that was not honestly measuring its own leak. The defect is the signal.

The ledger beneath the diary

A reflection document that says “we cancelled stale goals this cycle” is a diary entry. It is reassuring and unverifiable. The reason 365 reflections are worth more than 365 diary entries is the 1,396 evidence events beneath them: the actual goal-state transitions, the committed diffs, the rows written. Each reflection points at the machine-checkable proof of what it claims, which means a reflection can be audited rather than merely believed.

That is what makes the broader numbers trustworthy. When we report that the agent created 484 goals and pruned most of them, or that it tracked 969 issues, those are not assertions — they reconcile against the same evidence ledger the reflection loop writes to. The reflections are the narrative layer; the events are the accounting layer; and a system that has only the narrative layer is a system you cannot audit.

The metric that matters is not how often your agent reviews its own work. It is how often the review survives the next restart. Anything less is monitoring dressed up as governance.