Evidence direction review

Status: completed pre-migration corpus review. ADR 058 adopted the reviewed journeys as evidenced-by and is-evidence-for; old-label counts below remain historical evidence.

Question

Do the authorized source/review → note uses of evidence represent accidental mislabelling that should be removed, or a recurring inverse reader journey that has earned its own directional label?

The shared catalogue currently defines evidence as asymmetric, with the authored journey running from a theoretical claim to a descriptive observation. The live source and agent-memory contracts also authorize evidence in the reverse direction, from an evidence-bearing analysis or review to a theory note.

Method

The review extracted Markdown footer links whose immediate label is evidence, accepting both the current em dash form and older double-hyphen form. It resolved each relative target and selected:

  • kb/sources/ → kb/notes/;
  • kb/agent-memory-systems/ → kb/notes/.

For each selected edge, it read the containing artifact's connection/relevance section, recommended action where present, and target-side links back to the source artifact. Generated reports and workshop copies were not treated as authorizations or durable library evidence.

Corpus result

The review found 26 edges across 7 source artifacts and 20 target notes:

source artifact shape source→note evidence edges target links back
Language Models Don't Always Say What They Think ingest ingest report 4 1
Gödel Machines ingest ingest report 5 1
Ashby ultrastability source review 4 3
In Search of Lost Domain Generalization ingest ingest report 3 1
Towards Faithfully Interpretable NLP Systems ingest ingest report 3 1
FALSIFYBENCH ingest ingest report 5 0
Fintool review agent-memory review 2 0

Of the 26 edges:

  • 19 have no target→source return link;
  • 6 have a reciprocal target→source evidence footer, sometimes alongside an inline citation;
  • 1 has a target→source abstracted-from footer plus an inline citation.

The concentration is informative. This is not a long-established corpus-wide convention, but neither is it one accidental edge: it recurs across three artifact shapes, with ingest reports contributing 20 edges, a source review 4, and an agent-memory review 2.

They map a source into the claim landscape

The dominant reader journey is:

I am reading an analysis of this source or system; show me the KB claims that this observation supports, qualifies, bounds, or instantiates.

That journey is native to the artifact contracts. An ingest report exists to record how a source fits the KB; its Connections Found section keeps settled, durable connection judgments. A source review explains relevance to the KB. An agent-memory review closes with relevant theory and borrowable implications.

Examples span several evidential roles:

The final case is important: a source-side edge can assert that evidence bears on a claim without asserting that the target note has already incorporated, cited, or accepted it.

They are not lineage edges

Most target notes pre-exist the source analysis, and their substantive claims are not reconstructible from these source artifacts. Relabelling the edges derived-from or abstracted-from would falsely assert an origin or maintenance regime. The source is corroboration, counterweight, boundary evidence, or an instance—not the generator of the target.

Only 7 of 26 target notes link back at all. Where both directions exist, each serves a distinct reader:

  • note → source: inspect the observation that warrants or qualifies the claim;
  • source analysis → note: see where this observation lands in the KB's theory.

The 19 one-sided cases show that source-side authors did not merely mirror a target citation. Conversely, the reciprocal cases are useful two-way navigation when both reader needs are present.

Finding

The inverse journey is real and recurring, so removing source/review → note authorization would discard part of what source analyses are for. But retaining evidence in both directions makes an explicitly asymmetric label ambiguous.

The existing label also collapses two different statements:

  1. This claim points to evidence it uses or wants the reader to inspect.
  2. This source bears on that claim, whether or not the claim has incorporated it yet.

The distinction matters operationally. The second statement can be authored during ingestion and become a future update signal; it is not proof of target-side uptake.

Recommendation

Retain both reader journeys, but defer naming them until the vocabulary adopts a consistent directional grammar. The proposed evidence-for identifier is withdrawn: without a global grammatical rule it does not say reliably whether the source is evidence for the target or the target supplies evidence for the source.

The workshop's candidate invariant is now documented in directional label grammar: every directional identifier must complete source <label> target. Under that rule, the current evidence identifier also needs review because its declared meaning is “target is evidence for source.” The leading replacement pair is evidenced-by / is-evidence-for: both make the source endpoint their grammatical subject. It remains provisional until the whole asymmetric-label audit tests its boundary against neighboring relations.

Whatever names are selected, the semantic boundary established here remains:

journey reader need assertion boundary
claim or descriptive assertion → corroborating observation/source “show me the observation bearing on this assertion” The target artifact is worth inspecting as warrant, qualification, or boundary evidence.
evidence-bearing source/review → claim “show me the claim this observation bears on” The source bears materially on the target claim; it does not assert that the target already cites, incorporates, or accepts it.

Both relations are asymmetric. Authoring either remains an independent reader-aid decision, and a reciprocal pair is allowed only when both journeys help.

Migration outcome

  1. ADR 058 adopted source <label> target as the naming invariant and the evidenced-by / is-evidence-for pair.
  2. The pair explicitly includes corroboration, qualification, and boundary evidence while leaving grounds, supports, and contradiction cleanup to their own reviews.
  3. Contracts authorize assertion→source use beyond the theoretical profile where live corpus evidence required it.
  4. The shared catalogue and affected collection contracts changed together.
  5. All 26 reviewed inverse edges migrated to is-evidence-for; reciprocal edges changed only through their independent claim→observation migration to evidenced-by.

Decision status

The two reader journeys are established and the pair is adopted. Other identifiers that fail the source-as-subject invariant remain explicit migration debt for later scoped runs.