Chapter 18

How Search Results Get Confirmed as Memory

Chapter 16 ended at candidates. This chapter is the path from candidates to one Node. Search narrows where memory might be; resolution decides which memory it is.

1. Candidates arrive by several paths

The pieces search returns are only one kind of candidate. Node candidates come from three arms.

The title arm. Titles that share words with the question. Any of Concept, Record, Decision or Episode can qualify.

The graph arm. The Concepts the top pieces mention, followed back along the MENTIONS of Chapter 12.

The node semantic arm. Nodes whose title vector lies near the question's vector.

The candidates differ in shape. A piece is part of a document, a Concept is a meaning, a Record is an observation made at one moment. Resolution begins by pooling them under one standard.

2. Candidate and resolution are different states

If two candidates look equally plausible, the answer is not to find a rule that picks one. The absence of grounds for picking is itself a result. One wrong Node drags every later step in the wrong direction, so not picking is a normal response of the system.

Resolution answers with one of three states.

Read resolution as a similarity threshold and this distinction is lost. Search asks what could be relevant; resolution asks which of these is meant.

3. The pool is the ceiling

Resolution can only choose among candidates that entered the pool. However clever the ranking, it reorders the pool; it cannot reach a Node that was never in it.

That is not a hypothesis. Measured in September 2026 on sixty answerable questions, more than half of the gold Nodes never became candidates at all, and digging the pool four times deeper barely moved that. Turning on the vector arm moved node resolution by not one place, because the title arm never sees an embedding. That ceiling rose afterwards as candidacy widened: on 2026-09-29, 48 of the 60 gold Nodes were in the pool.

What moved it was candidacy, and two things showed themselves.

An Episode could not be a candidate. The pool was built from Concepts, Records and Decisions, so a store's Episodes, however many, could never enter it. It was closed the day it was measured, and no other kind lost a single question.

Records won too often. The title arm scores by the count of matched tokens, and in this store a Record title has a median of 22.5 words against a Concept's 5. A sentence-shaped Record title carries four times the surface of a noun-phrase Concept title and takes other questions. Normalizing by title length looked like the fix and measured the other way: a short title then wins on one token and Record collapses from four of ten to none. The cause was right and that remedy was wrong.

4. The bar for confidence

The first rule was confident on a single matched token. Confident wrong answers stood at 36 of 46.

Raising the bar to two tokens and ordering by how many arms agreed brought confident wrong down to 15. Letting the node semantic arm decide by the gap between its first and second candidate rather than by the distance itself brought it to 3. The raw vector distance is unusable on this corpus: an exact match and an unrelated sentence sit in overlapping ranges. The gap separates cleanly. Measured separately, both changes would have been rejected: the arm alone moves nothing, and the bar alone lowers the score. They pay only together.

In the measurement of 2026-09-29, the first candidate in the list was the gold on 37 of the sixty questions. But the system resolved to the right Node on only 12, and on 45 it returned choose with up to five candidates. Confident wrong answers fell not because wrong answers became right ones but because they became candidate lists. In September 2026 this was ruled the place the product stands: rather than resolve to the wrong Node, it hands the choice back to whoever asked.

5. When there is no answer

Twenty-one questions with no answer in the graph were measured alongside. The danger there is resolving with confidence.

The system resolved 0.14 of them. One gate brings that to zero: if the part of the question the resolved title did not consume touches nothing in what search returned, do not be confident. Correct resolutions are almost untouched.

That gate ships off. Not for performance: fifteen of the twenty-one unanswerable questions were written in the same session as the rule that grades them. A gate whose rule and whose exam share an author ships off by default; that is this graph's standing policy, and the measurement reopens when someone else writes the exam.

No candidate, or ambiguous candidates, means no resolution. That is not a failed search but a state of the system. The most dangerous outcome in a memory system is resolving a target without grounds and letting every later step treat it as fact.

6. Each kind names a different thing

A Concept answers what something means, a Record preserves a particular fact, an Episode answers what happened and when. Search landing on a Concept does not make its attached Records the same thing, and a Record being found does not make it the concept. After resolution these different objects are read as connected and distinct.

One thing is set apart for Episodes. An Episode's narrative summary stays outside what search matches against. Used for matching, a summary is wide enough to win questions the graph has no answer to, so it stays in the answer material and out of the search.

7. Language is a candidacy problem

For text, vectors close the distance between languages. A Korean question reaches an English piece that shares no word with it.

Naming a Node is different, and it was measured. The title arm's candidacy follows how many of a title's words the question carries, and chunk vectors reorder pieces without putting a title on the candidate list. So at first a Korean question against an English title failed before ranking, not at it. On that axis the score was zero.

At first it read as a language problem. Counted, nearly half the Concepts in this store had Korean titles, and once title-token reuse was held fixed every language axis performed about the same. What decides is not the language but how much of the title the question reuses. A title in another language that never enters the pool is a candidacy problem, not a ranking one, and the remedy lies in widening candidacy, not in reranking.

What opened that path was the node semantic arm. With the title itself held as a vector, a question that does not reuse a title's words can still reach it. Switching that arm and the question encoder to a 1024-dimension multilingual model (2026-09-13) raised the questions whose first candidate is the gold from 29 to 42 of 60, and almost all of the gain was on questions that cross languages. Correct resolved answers stayed at 13. Candidacy rose; commitment did not.

8. Only what holds now, by default

The time axes of Chapter 24 are used here first. Both the title arm and the node semantic arm put forward, by default, only claims that hold now. A claim with no stated validity reads as current; a retired claim is returned only on request.

That filter taught one thing. When it was first added, one of two tied candidates was a retired claim and dropped out, so the survivor was promoted to resolved. A question with no answer got one. The fix refuses to resolve when a dropped claim tied with the survivor. A read filter must change what is visible, never what the answer is. Anything that removes candidates before the tie-break can cause the same accident.

Once one Node is resolved, material can be drawn from around it. That is the next chapter.

Continue with Chapter 19, How Memory Becomes Material.