When Should a Learned Rule Survive?

Once AutoHarness can surface a skill and observe how the host encounters it, it faces a maintenance decision: when is there enough evidence to keep the rule, and when is the active library carrying more weight than value? This chapter traces one learned skill from creation through probation, maturity, and possible archive. The key judgment is not whether a rule feels old or new. It is whether the rule has had a fair opportunity to be found and what kind of contact the system can actually observe.

A new skill needs a fair sample

AutoHarness does not make a new skill compete immediately with mature skills. When a promoter creates a skill, the sidecar records a creation anchor against the request count for its layer. The lifecycle manager later compares the accumulated request count with that anchor. Until the difference reaches the configured maturity threshold, the skill is in probation: it remains live and available, but it is not counted against the mature capacity and cannot be evicted by the normal contention rule.

The anchor matters because a global or project library is not observed in a single uninterrupted process. Requests arrive over time, sessions begin and end, and the operator may simply not work on a matching task immediately after the skill lands. Wall-clock age would punish a skill for time during which the host had no meaningful opportunity to use it. An opportunity-relative denominator asks a fairer question: how many requests reached this layer after the skill became available?

Probation is not a statement that the new rule is correct. It is a protection against premature judgment. The testing convention about readable parameterized cases may be excellent, but a single later session cannot tell whether it applies broadly enough, whether the index made it visible, or whether a different repository convention should have remained local. Keeping the entry alive while the denominator grows gives the recall surface time to produce evidence before the capacity policy begins to compare it with established entries.

At the maturity boundary, the system performs a graduation review. A mature skill with no recorded use and no recorded view can be archived even without a capacity crisis. A skill that was viewed but not formally loaded is spared that particular zero-evidence decision, because inspection still indicates that the rule had recall value. Used skills enter the mature pool and are evaluated for capacity contention. The distinction lets AutoHarness remove a genuinely invisible entry without treating every absence of a formal Skill call as proof that the entry was worthless.

The review is evaluated from accumulated state at session start, before the current session's index is built. That ordering keeps a rule which has already been archived from appearing in the same recall surface that decided it was no longer active.