The second axis: derivation confidence × domain relevance, with the derivation that makes a rejected group a query

[Hafeok]

Implements the top-ranked finding of the G0 session report (§7.1), with §7.4 arriving at the same
change from the review side. corpus-backend read read-only at 3b0a56b3, full solution, on a
machine with private-feed access.

The two-axis model

§5.2's assurance column measured one property while being read as two, and G0 proved them
independent: 31 assertions at the top grade, every mechanical check passing, all 31 wrong to
ratify. It splits:

• Derivation confidence — high · mid · low · not-derivable. The value set is unchanged,
  deliberately: the old column was already measuring derivation and nothing else, so this renames it
  rather than re-marking it, and no ratified assertion changes.
• Domain relevance — high · low · unknown, with a basis (computed · defaulted · not-computed · not-applicable). Not ordinal: unknown is the absence of a verdict, not a value
  between the other two.

Every row is unknown / not-computed at G0, and says so. The column is declared-empty, not
undeclared — a different and better state than the one it replaces, where relevance was undeclared
and silently read as high.

The derivation record, and where it lives

Two marks say what a class of assertion is, not how a reviewer finds it again. Each assertion
gains reg:derivedByRule, reg:standsOn, reg:memberKind, reg:containerKind, reg:sourcePath —
recorded facts, with the classification left to the reviewer's query. They live in a per-run
evidence graph under runs/, keyed by the assertion's content-addressed id, leaving graphs/ pure
ratified assertion.

Template 0.3.0 — backward compatible, verified

Shapes target the class reg:GradedAssertion, stamped by the writer and entailed by nothing. A
shape keyed on predicate presence would pass in the authority repo and fail at the projection, where
owl:equivalentProperty manufactures reg:derivationConfidence on the older cohort; a fixture
validates the projection, not just the repo.

product registry check over an instance holding all 561 pre-0.3.0 assertions plus the new run
evidence: conformant, 9 667 data files, 48 constraints, file rule clean, no file changed.

Full-solution re-run

16/16 projects restored, 951 source files, 967 declarations, 9 662 assertions read, 18 review
batches, 0 orphaned evidence entries.

Every row's behaviour confirmed; no row behaved differently at full scale. §5.2's provisional
pending a full-solution run qualifier is discharged. subclass-transitive still fires zero over
172 edges; supertypes-outside-subset drops to 0. Three findings the subset could not produce:

1. A probe verdict is itself a Reading. workspace/symbol did not answer on one run and
   answered on two others, same corpus, same ref. Not "over-advertises at scale" — intermittently
   unavailable under load. Verdicts are now recorded in the run evidence with their as-of, typed
   reg:ProbeVerdict , reg:Reading so the existing Reading shape demands the four-tuple. Two runs
   that disagree differ visibly, and a reviewer can tell "this row found nothing" from "this row was
   gated off that day". Retry-then-record is proposed, not built (PRD §12 item 17).
2. The modules row does not stand on workspaceSymbol. On the run where the operation was
   unavailable the row still fired 1 514. Table corrected to documentSymbol — a correction that
   could only come from a failure.
3. §7.5's collision prediction held. 21 cross-module simple-name collisions where the subset had
   0, so range resolution is now unsafe on this corpus. definition is wired, answers, and is
   consumed by nothing. Filed as PRD §12 item 16, scoped to the definition route.

The fixpoint ratio inverted (0.84 → 1.69) at 509 ms: g-dec-02's condition not met, so the
decision does not flip; direction now measured.

The two groups, recovered from the graph

	Query	Full solution	In the 561 ratified	G0 reported
Group B	reg:memberKind = reg:field	195	31	31
Group A	any reg:sourcePath under a generated-migration folder	18	18	18

Both reproduce G0's counts exactly, by query where G0 needed a hand count that was wrong twice.
All 561 ratified assertions now have derivation; 0 without — the backfill ruling confirmed on
the real cohort, with no ratified byte changed. Both readings of the path predicate are recorded in
the PRD (every recovers 14 of 18, any recovers all 18) so the next reviewer inherits the choice.

Idempotence

Second full run at the same ref: 0 proposed, nothing written, 0 tracked files changed. The new
fields stay outside the assertion hash at 9 662 assertions, not only in a fixture — g-dec-04's
identity is intact.

The instance defect, and the clause behind it

G0 ruled 512 accepted / 49 rejected; the instance holds all 561. The ruling was prose, the act
was a merge button, and nothing checked that the second carried the first — presumed discharge at
the ratifier's seat. Disposition ruled: retraction, not deletion, and not this session — a
retraction decision naming the 49 by id, basis the Gate 4 ruling, files removed in the same act.
Emil drafts it.

PRD §5.5.2 files the general clause: a ratification act must be checkable against the ruling it
discharges — a stated accept/reject set by id before the merge, the merged tree diffed against it.
Per-triple review that ends in one undifferentiated merge is per-triple review in name only.

The 49 were unfindable until this change and are now two queries: the recoverability argument, made
in the abstract at G0, paid out on a real defect in the ratified set.

What this does not do

It does not filter Group A. It does not filter Group B. It does not widen the declaration subset.
It makes both groups visible and batchable.

Still open

The subset-widening decision (§12.15) · the simple-name collision defect (§12.16) · retry-then-record
for probe verdicts (§12.17) · the retraction decision Emil drafts · Kotlin's workspace import · G0.5
and the comparison instrument.

Gates

cargo t 1 495 / 0 (baseline 1 454) · clippy 0 · cargo xtask check 0 errors · ddd validate
conformant (277) · contract surface 88 events, 0 undischarged · product registry check
conformant · file/function length and single-responsibility clean.

🤖 Generated with Claude Code

https://claude.ai/code/session_01AdDM2VGG3qyDAN5vGrphS2