conflict-adjudication

Conflict Adjudication

Two memories say incompatible things. This skill decides what to recommend about them: retire the older one, let both stand, or put it in front of a human.

A memory here is one stored entity — a decision, a lesson, a working agreement — read through its summary and traced back to the episodes it came from. The relationships around it are evidence about it, not the thing being judged.

It runs on one pair you already hold — typically because a capture path searched before writing and turned up something that contradicts what was about to be stored, or because a reader noticed two stored memories disagree. So the pair arrives in one of two shapes: two stored memories, or one stored memory and one candidate — content not yet written, whether a claim being made live or a queued draft carrying its own provenance. A candidate is adjudicated like any other side; only where its evidence comes from differs.

The evidence is the organization itself: who or what asserted each memory, what standing that carries, and whether the two apply to the same thing. It is extracted from whatever happened to be ingested, so it is uneven — most of this skill is about knowing when it is good enough to act on.

Hard rules (non-negotiable — read first)

1. A pair comes in — never go hunting for one. Two specific memories, one recommendation. This skill does not sweep the graph looking for contradictions; "find all the conflicts" is a different job and not this one. With no pair in hand, ask which two memories are in question and stop — and when the question is really whether a single memory is still current, with no second memory in play, hand it to graph-traversal.

2. Recommend, never act. Name what should happen and why — the caller decides and writes. Never call a delete tool, never write a correcting memory, and never describe a memory as "superseded" in the past tense. It isn't, until someone acts on the recommendation.