The Correction Loop

The correction loop is the mechanism Arkeus uses to turn mistakes into durable behavior change. It has three steps: capture the incident contrastively, write the rule as an applicable instruction, reload at the top of the next context.

Capture starts at the moment of failure. When an agent does the wrong thing — sends the wrong message, reads the wrong calendar, commits Ryan to something he did not approve — the correction is written immediately with the exact shape: what happened, what should have happened, the rule that separates the two. The contrastive structure is load-bearing. A lesson without an incident is abstract and will not fire. An incident without a rule is just a story.

The rule goes into corrections.md with a date stamp and a source. The file has a ceiling of around 4,000 tokens, which forces pruning. When a newer correction supersedes an older one, the older one gets removed or compressed. The ceiling is deliberate: corrections that have become ambient knowledge can graduate into the kernel itself, and corrections that have been superseded should not take up the scarce slot.

Reload is where the loop closes. Corrections.md is loaded last in the kernel order, because recency bias gives late-position content the strongest behavioral effect. The model reads corrections just before it starts generating. The most recent failure mode is the most salient context.

Agent-local corrections work the same way but stay inside the agent's own file. They only graduate to the global corrections.md when the same mistake appears across three or more agents — the graduation threshold prevents local quirks from flooding the global layer while still catching cross-cutting patterns.

The loop is what makes the system anti-fragile. Every mistake is a chance to update the rules. The rules reload every session. The same mistake does not happen twice in a row unless the loop is broken — and when it is broken, that failure itself becomes a correction.