Chapter 23

How Evidence Attaches to a Relationship

The roles of Chapter 22 make relations as they work. As the graph grows, accurate Nodes stop being enough. That a relation exists between two Nodes has to be trustworthy too.

That two Nodes are joined by a relation does not say why the relation exists. So relations need evidence as well.

1. Relation and evidence are different information

The relation itself is a connection. Evidence is the material that supports it.

A relation joins A to B; its evidence and its confidence hang from it as separate pieces of information A B relation evidence the sentence that supports it confidence the trust placed in it

A relation existing and a relation being verified are not the same state. In this graph that was literally so. The evidence and confidence columns existed, and nothing filled them. Measured, zero of eighty directed relations carried any evidence, and the relation panel always said “no evidence”. A channel wired to a field nobody fills is worse than no channel, because it looks like information.

2. Evidence comes from a sentence

What filled the empty column was the source text of Chapter 12. There are two rules. When a Chunk mentions two Concepts and their titles appear verbatim together in the same sentence, that sentence becomes the relation's evidence. When the titles do not appear verbatim but the key words of both are all in one sentence, that sentence is used too, and trusted less. Either way, appearing in the same document is not evidence. Sentence splitting is a regular expression; there is no language model.

For verbatim titles the confidence falls from 0.9 to 0.5 with the distance between them; evidence found by key words is set at 0.4, so a value below 0.5 marks the looser rule. No value exceeds 0.9, so that the number itself says no person confirmed it. Only a relation a person confirmed can go above. A relation that already has evidence is left alone, so the fill can run again safely.

3. Confidence is not evidence

If evidence explains why a connection was made, confidence says how far to lean on it. Merge the two and support gets tangled with judgment. So a relation keeps them as separate fields.

These values are visible to people as well. In the View every relation line is a hairline, and its ink density carries the confidence. Chapter 25 shows it.

4. A relation has provenance too

When a document mentions a Concept, the MENTIONS of Chapter 12 joins the Chunk to the Concept. When that Concept attaches to a Record, another relation forms, and the Record carries its own source.

Provenance follows the path back from answer material through a Record, a Concept and a Chunk to the sentence in the source answer material Record when, which file Concept Chunk MENTIONS, by which rule source sentence where it began

It has to be possible to walk this path backwards. That is provenance. A Record about code carries the repository and file path, when it was read, and at which commit, so that two claims about the same file can be compared rather than both standing as assertions.

Relations also have a birth time. A relation a person authored carries the moment it was first connected; a relation the system re-derives does not, because the birth of something a rebuild remakes would read “last rebuilt” rather than “first connected”. An old relation's empty birth is left empty. Filling an unknown time with the migration date piles the whole timeline onto one day.

5. Automatically made relations still need verification

The extraction of Chapter 15 and the mention linking of Chapter 12 are convenient, but a relation is not true because it was made automatically. A long title or a vague phrase can be linked to the wrong Concept.

So generation and verification are different stages. What extraction draws from the source is a candidate relation, and it becomes a usable one only after its evidence is checked.

Mention linking writes onto the relation which rule linked it: a title match, a stem match, a provenance match. That is why Chapter 19's brief returns the rule with every excerpt. Different rules do not deserve the same trust.

6. brief carries the evidence as it is

When the material of Chapter 19 carries a relation, it carries more than its name and direction. Whether there is evidence, which sentence it is, the confidence, and the validity period of Chapter 24 come with it. The Agent receives not “these are connected” but “connected because of this sentence, with this much confidence, over this period.”

Validity on the Edge itself is deferred. No incident has yet been measured in which a relation went stale as a relation. Node and Record validity come first; intervals on Edges wait for a measured trigger.

7. Evidence becomes part of the graph

In the end evidence does not live only in documents. Because a relation itself carries evidence, the graph widens from a list of connections into knowledge that carries its own support. Evidence, confidence and a time of birth are added to the Node and the Edge.

This closes Part 2. Memory was found, resolved, made into material, handed through the door, and the roles that use it and the evidence on the relations they make were seen. Part 3 is where that memory lives. Join when a relation held and in which space it is visible, and the graph becomes a structure with time and space. Time comes first.

Continue with Chapter 24, What Time Exists in Memory.