anthology-of-the-sota

ADR-061: Theories get a chain of their own: lines of explanation

3 min read

Properties3
StatusActive
Tagsrecord, mechanism
Influenced byADR-011 — Lineage between documents is a field, not a paragraph repeated in every note, ADR-031 — A scheme for explanations, complementing the practice registry

Graph View

Context

ADR-011 turned the lines of work into a rendered chain: extends: and corrects: walked transitively, one page per scheme, so no document has to re-describe the line it sits in. It did that for the papers (lineage) and the practices (practice).

ADR-031 then split explanations out of practices into the THEORY scheme, and gave THEORY the same two relations. extends refines an account that holds. corrects replaces one whose reasoning broke. Fourteen theories now carry one or the other. Nothing walks them. A reader who lands on THEORY-003 learns from its frontmatter that it corrects THEORY-001. They cannot see that the record holds six other corrections of the same shape, or that correction is the normal way an explanation’s line develops here.

Decision

Declare a third chain in luria.yaml:

theory:
  scheme: THEORY
  relation: [extends, corrects]
  invariant: tags
  output: docs/theory-lines.md
  title: Lines of explanation
  facet_by: [status]

It is listed in docs/README.md beside the other two. The lint requires this: it reports a top-level view with no index entry.

Three choices in it are worth stating.

  • No sibling. The other chains render compared_against: as a cross-link. THEORY declares no such relation, since no explanation here states a comparison, so there is nothing to render.
  • invariant: tags, as on the other chains. An account that refines or corrects another is about the same thing, and a shared topic is what says so. On the record as filed, all 11 lines bind. That is the check doing its job at zero rather than a reason to leave it off. A future line that does not bind should get the reading ADR-049 gives: the invariant is missing or the relation is wrong.
  • Faceted by status only. The practice chain adds consensus. THEORY has no consensus field, and its status is the whole verdict. That is what shows THEORY-001 (Rejected) above THEORY-003 (Active), which is most of what the page is for.

Alternatives considered

  • Fold the theories into the lineage chain. One chain walks one scheme, and the question differs anyway. The paper line asks what a paper built on. This one asks which reason replaced which. ADR-031 separated explanations so that a practice could outlive its stated reason. A page that mixed the two would put a rejected explanation in line with the papers that are still good evidence.
  • Leave the relations unrendered. Status quo. They are already declared and linted, and each document names its neighbours. That is the arrangement ADR-011 replaced for papers: a line that exists only as each member’s local view, which nobody can read whole.
  • Add compared_against: to THEORY so the chain has a sibling. The scheme’s own comment says why not: a relation nobody writes is a converse the fixer can never complete. Add it when a theory states a comparison.

Consequences

docs/theory-lines.md is generated on main like every other view. On the record as filed it shows 11 lines. Seven of the fourteen theories on them correct an earlier account; one, THEORY-070, corrects two. Among them are THEORY-001 → THEORY-003 (batch normalization and internal covariate shift), THEORY-017 → THEORY-018 (influence functions), THEORY-068 → THEORY-067 (in-context learning as gradient descent) and THEORY-079 → THEORY-080 (cold diffusion).

The lint’s unbound findings now cover THEORY relations too. Today they report nothing new.

Backlinks