Current case: from direction repair to mechanism ambiguity

This file records the concrete work that opened the foundations question. Counts are dated baselines from successive reviews, not timeless corpus facts.

What changed

The work began by correcting the idea that links are simply forward-authored and asymmetric, with backlinks as the only reverse surface. The current rule is more precise: the source collection governs outbound authorization; each direction is independently authored when it serves a reader need; semantic symmetry does not require mirrored Markdown; and a useful reciprocal edge is allowed rather than discouraged. Backlinks remain a derived inbound view, not a prohibition on reciprocal authoring.

The collection contract can authorize local collection destinations, the reserved external destination, or literal any, which includes external. This made destination permission independent from label semantics.

Directional grammar

ADR 058 adopted the invariant:

Every directional identifier completes source <label> target.

The source is the artifact containing the Markdown edge. A reader should not need a collection table to discover that the identifier silently describes the target's role instead.

Evidence migration

The old evidence identifier was used for both assertion→observation and observation→assertion. The review found that both reader journeys were real rather than accidental reciprocity. ADR 058 therefore adopted:

  • source evidenced-by target for assertion→observation;
  • source is-evidence-for target for observation→assertion.

The migration covered 207 active old-label edges, including 26 inverse-journey cases, and updated contracts, current guidance, and the mutable corpus together (b3f7e126). It established the reusable rule that semantic classification precedes mechanical replacement.

Rationale and off-pattern grounds

The old rationale identifier also described the target: the target was the rationale on which the source depended. Corpus review found 134 active edges in three semantic groups rather than one renameable relation. ADR 060 adopted source rests-on target for design, description, procedure, rule, and system-definition dependence.

The migration (dac39977) assigned the 134 old rationale rows as:

successor rows
rests-on 114
evidenced-by 15
is-evidence-for 2
implements 2
compares-with 1

It also reclassified 38 off-pattern grounds rows: 8 to rests-on, 20 to evidenced-by, and 10 to is-evidence-for. The remaining 276 note→note grounds edges were deliberately preserved for a separate premise review. The evidence and rationale runs supplied the repeated mechanical core now recorded in Migrate a directional link label.

Grounds evaluation and adjudication

The grounds direction review rebaselined 283 active rows: the preserved 276 note→note cohort plus seven newly authored sources→notes evidence mappings. It recommended premised-on for 160 clear premise dependencies while finding 49 mechanism/explanation rows, 23 extensions or specializations, 39 mixed definition/lineage/comparison/navigation rows, and 12 evidence rows.

The maintainer adjudication accepted premised-on and assigned exact successors to the evidence, extension, and mixed cohorts. Eight rows from the coarse X/D buckets proved to be premise dependencies on closer inspection:

disposition rows at adjudication baseline
premised-on 168
extends 23
defined-in 13
exemplifies 12
is-evidence-for 9
evidenced-by 8
see-also 1
deferred mechanism evaluation 49

This reconciles the 283-row review baseline. A later mechanism scan found 285 active grounds rows—278 notes→notes and seven sources→notes—showing again that review counts drift while work is in flight. The two newer note→note rows still require disposition before implementation.

The mechanism problem

At the next baseline, the active surface contained 79 edges literally labeled mechanism—78 notes→notes and one reference→notes—plus the 49 grounds rows deferred as mechanism-like. Nine source files author both old labels. Their local use does not establish a stable distinction: mechanism sometimes names a prerequisite or explanatory theory, while grounds sometimes names an operational path.

The unadjudicated mechanism direction review reconciles all 128 candidate tuples and recommends a split:

semantic class rows proposed treatment
explanatory relation 41 candidate explained-by
operational mechanism 72 candidate operates-through
enabling condition or prerequisite 10 defer to enables / precondition review
extension or exemplification 3 exact existing relation still selected per row
premise 1 premised-on
evidence 1 is-evidence-for

The report was requested for Luna, but its own provenance says Luna execution was not verified; the attempted local --model luna dispatch failed because this account does not support that model. Treat its semantic split as evidence awaiting maintainer review, not an adopted decision. Its 128-row ledger reconciles structurally, but structural validation does not establish that the artifact/claim/phenomenon distinction was applied correctly.

Why this is a foundational problem

The candidate split exposes at least three different assertions:

  1. a source claim is explained by a target claim;
  2. a source phenomenon or capability operates through a target process or component;
  3. a source assertion is premised on a target claim about how something works.

These are not merely alternative names. explained-by is an epistemic or discourse relation between claims; operates-through is a stronger subject-matter claim about productive organization; premised-on records an inferential dependency. Yet Markdown endpoints are artifacts, whose claim titles sometimes stand in for propositions and sometimes describe systems or phenomena. A grammatically correct identifier can therefore still cross levels invisibly.

The distinction also changes maintenance:

  • rejecting an explanation calls for revisiting the source's explanatory account but need not falsify the phenomenon;
  • changing a process or component calls for operational or interface review;
  • rejecting a premise reopens the source assertion's truth or applicability;
  • losing evidence changes its support record;
  • removing an associative neighbor need not invalidate anything.

The current reader-need test catches part of this—“verify,” “understand how,” “inspect the process”—but does not by itself say what the source asserts or what revision should propagate.

What the foundations workshop must settle

  • Whether formal identifiers primarily encode a subject-matter assertion, an inferential role, a discourse function, or a revision consequence.
  • Whether explained-by and operates-through remain reliably distinguishable when endpoints are notes rather than literal phenomena and mechanisms.
  • Whether a relation signature must state source role and target role, or whether collection plus claim title supplies that context cheaply enough.
  • Whether the operational-mechanism category from the 128-edge review survives independent adjudication using the stricter philosophy-of-mechanism test.
  • Whether prerequisite relations should be handled as a separate family before any mechanism migration.
  • Which distinctions improve agent follow/skip, reasoning, or maintenance behavior enough to justify their authoring cost.

Until these questions are answered, the consistency workshop can preserve inventories and exact non-mechanism dispositions, but it should not adopt or migrate explained-by or operates-through.