An AI re-derives a decision that was already made and may quietly arrive at a different answer. There is no warning that this week's answer contradicts last month's. That is drift, and it is the most expensive failure mode in AI-assisted work.
When the AI does not know something, it guesses. The guess is presented with the same confidence as a known fact. There is no provenance, no source, no way to tell whether this is settled or invented. That is hallucination on unknowns, and it is the second most expensive failure mode.
Every time a decision is re-derived from scratch, it costs tokens, time, and risk, when a single lookup would have returned the settled answer. The token cost of re-deriving is real, and it compounds across every station, every query, every session.
And a knowledge base full of unattributed assertions cannot be trusted, because you cannot tell who said it, when, or from what evidence, so the AI treats everything as equally valid. Untrustworthy knowledge is no better than no knowledge, because you still have to verify everything yourself.
Look it up before you predict it.
If a key exists, the stored statement wins over inference. Parsing a row beats re-deriving a decision that was already made, and unlike prediction, it cannot quietly invent a different answer than last time.
If the key does not exist, that miss is an answer. It means ask a human, or ask the station that owns the scope. It does not mean guess. A miss is a stop, not permission to invent.
Raw conversation is evidence, not memory. It becomes memory only by promotion, with provenance. Only an ai_id with promotion authority can move an entry from candidate to promoted.
A station queries Living Memory before asserting any fact it did not just observe itself. That is the lookup-first discipline, and it is a conformance rule, not a suggestion.
On a hit, the stored promoted statement is returned. On a miss, the answer is on_miss=ask_human. That is the default and the correct answer. The only time a station proceeds on unknown is when the caller explicitly declared the fact optional. Otherwise, a miss stops the work and asks a human.
Candidates are hints, not truth. A candidate is asserted but not confirmed. A station must not act on a candidate as though it were settled. The candidate-to-promoted pipeline is the gate that separates someone said this from this is confirmed and safe to act on. Only an ai_id with promotion authority can move an entry from candidate to promoted.
Every entry carries a provenance chain: who asserted it, from what source, and when. The source can be a human directive, a completed packet, an observation, a conversation, a document, or a measurement. Optionally, an entry can name who independently verified it and when. Provenance is what makes a fact trustworthy. Without it, you have an assertion. With it, you have a chain of custody.
Never edit a promoted entry. Write a new one and mark the old superseded by it. The old statement is never modified. A wrong entry is retracted, not deleted. The record of the mistake is what stops it recurring. The history is the asset, and the immutability is what makes it trustworthy.
This is stricter than retrieval-augmented generation. RAG retrieves relevant context before generation, augmenting the model's prediction. The EAS principle is that a stored fact does not augment prediction, it replaces it. If the key exists, the stored statement wins outright, not as a suggestion to the model.
The conformance test that proves the discipline: move an ai_id to another engine and another operating system, replay a packet, and the same answers are returned. If anything differs, an identifier is carrying a mutable attribute somewhere. Lookup, not prediction, is what makes memory durable across changes.
Before: the AI re-derives every decision from scratch, costs tokens and time, and may quietly arrive at a different answer. On unknowns, it guesses with the same confidence as a known fact. Knowledge is unattributed and untrustworthy.
After: a station queries first. A hit returns the settled statement with provenance. A miss returns ask_human and stops. Candidates are visible only when explicitly requested and never acted on as truth. The decisions that were already made stay made, and every fact traces to a named source. Drift stops, hallucination on unknowns stops, and the token cost of re-deriving drops to the cost of a lookup.