PROJECTSKILLSCONNECTING…

SYS.20 / GROUNDING

Grounding and evidence

Every claim carries a confidence and the files it came from. The interesting part is what happens when the files are not there.

What can be trusted4 SECTIONS

Three values, fixed everywhere

The same enum in every subsystem, with the same colours.

VERIFIED
Read directly from your files and re-checked against the snapshot’s real file index. Drawn green.
INFERRED
Reasoned from what was read. Plausible, not established. Drawn orange.
UNKNOWN
Nobody established it. Drawn in muted ink, and reported rather than omitted.

These sit beside, and are not the same as, the five statuses a check can conclude with. Everything the deterministic scan emits is VERIFIED by construction — a manifest or a parse tree said so. The one exception is the architecture *style*, which is read off directory structure and is therefore INFERRED.

The evidence gate

Citations are re-validated against the file index, in code.

  1. A model asserts a fact with citationsREASONED

    Paths and line ranges, in structured output rather than prose.

  2. Each citation is resolved against the snapshotGATE

    Not against a memory of the repository — against the file index this snapshot actually holds.

  3. Unsupported claims are demotedGATE

    A VERIFIED claim citing a file that does not exist becomes UNKNOWN. A path that resolves to nothing is removed from a knowledge source rather than shipped.

A VERIFIED claim with no citation at all is demoted regardless of any check. The whole meaning of that value is *go and look*, and a claim offering nowhere to look has not earned it.

UNKNOWN is a result

Reporting what could not be established is the difference between a system that can be checked and one that sounds authoritative.

Most tools in this category have no way to say "I could not tell". The gap gets filled — with a confident sentence, with a zero, or with silence, and all three read as good news. Here it has a name, a colour and a place in every summary.

  • A pillar with no successful detector reads "not checked", never "no problems".
  • A count derived from a run where something failed is a floor, and is labelled as one.
  • A section of a generated document with no material says so, next to the claim it qualifies rather than in a footnote.

What the gate cannot do

It checks that evidence exists. It does not check that the evidence means what the claim says.

This is why every claim in the product — a Skill’s knowledge sources included — remains expandable to its evidence. The gate raises the floor; the citation is what lets you check the ceiling.

DOCS23 CHAPTERS