nucleation

record/status-review/README.md

8 min read

Graph View

How a status-review pass is run under ADR-057. The decision says what a pass is for; this file is the working procedure the sixteen passes of 2026-10-27 to 2026-10-30 converged on. It replaces briefs that lived only in a session’s scratchpad. That is how one of them carried a wrong precedent, about a narrowing’s dropped clauses, for seven passes before a reviewer checked it against ADR-047.

1. Build the worklist

On a freshly fetched origin/main, after the last pass’s temporary codes are concretized (main’s latest commit is the docs bot’s “regenerate views”):

python -I record/status-review/worklist.py --top 30
python -I record/status-review/worklist.py --show CLAIM-123

The first six sections must be empty: inversions, stale converses, sound arguments whose conclusion is not held, unwritten concessions, retired premises, and Open questions with an Active answer. Anything listed there is the pass’s first work. Then take work in this order:

  1. Ready for a ruling. These claims have every premise held, no live objection and no standing blocker note. They are grouped by primary line, and a line’s group is a natural cluster.
  2. Sole blockers. These are the unheld premises that alone hold a claim back. Review the premise and the edge together: in passes 13 to 16 the blocker was more often a ground that is really corroboration (ADR-058) than a premise that could be promoted. A sole blocker is counted, not listed, while a recent blocker note holds the premise or every claim it holds back: that note already says what the wait is.
  3. Blocker notes something they name has moved since. An entry the note names was revised, or a work it names got a reading, after the note was written. The wait may be over, so look before the timer runs out. When it is not, leave a new dated note that restates the wait and names the move that was checked, after the old one; replace the old note only where it is now wrong. The new note is the record that the look happened, and it takes the entry off this list.
  4. Proposed readings no source of which is read in full. These need reading, not review. Leave them out of clusters.
  5. Live claims that need a work no NOTE reads. An unread work takes no edge (ADR-058). Either the work is read, or the edge goes and the body keeps the citation; an Active claim on the list is checked first.

The held timer is a floor, not the review. An entry with a blocker note under 14 days old is listed as held. The timer guarantees it is looked at again at least once; it does not stop anyone looking sooner, and section 3 brings it back early when what it waits on moves (ADR-057 v7).

The load-bearing ranking by reach came last on 2026-10-30. By then the most any Proposed entry was relied on was 1, and the ranking had stopped ordering work.

2. Review, read-only

Size a pass at about five clusters, one reviewer each. A reviewer reads every entry in full, with its premises, objections, dependents (grep the code) and the NOTEs on its sources. It proposes exact text and edits nothing.

Statuses.

  • A CLAIM goes Active when ADR-057’s four criteria hold:
    • every premise is Active;
    • each live objection is answered or written in;
    • nothing in the record meets its defeated_if;
    • the ruling is made.
  • A THEORY is a reading. It goes Active when it is read in full and its promote_when holds, or when its title claims no more than one study read in full measured (ADR-047’s evidence rulings).
    • A promote_when asks only for what the title asserts (ADR-065). A title that reports what a source proposes, argues or finds, or what one study measured with its hedges kept, goes Active on its reading; a demand for replication or an independent test moves to a QUESTION, or, where the source asserts the wider thesis, to a Proposed THEORY that extends the entry and keeps the old condition. A demand for the version of record of a copy already read is dropped unless a discrepancy is named; the NOTE records which copy was read. A general title keeps its strict condition.
    • A checked, step-by-step re-derivation meets a request for a proof. Write it out, and put any computation in a script with controls that someone else re-runs before the ruling.
  • A TERM is adopted when its definition is settled and every use says which sense it means.
  • An ARG whose body names its own rejection condition goes Rejected when the condition is met.

Narrowing. Narrow a title only to make it promotable, and route every clause the narrowing drops by ADR-047’s “Where a narrowing’s dropped claims go”:

  1. A generalization the source asserts becomes a Proposed THEORY. It keeps the old wording, extends the narrowed entry, and inherits the old promote_when. Models: THEORY-164 and THEORY-174.
  2. The record’s own extrapolation becomes a Proposed CLAIM with warrant: interpretive, grounded in the narrowed entry.
  3. A mis-statement, or a clause shown false, gets a dated note only.
  4. Only open scope that nobody asserts becomes a QUESTION.

Passes 8 to 14 sent every dropped clause to a QUESTION, and the ADR-047 kind-1 sweep of 2026-10-30 repaired it.

Narrow or split only when it frees something. Narrow or split only when doing so frees a dependent, or when the reading is promotable on its own merits. A claim that needs only what an author proposes can stand on an Active “X proposes that …” THEORY split off on THEORY-1110’s pattern; the original reading stays Proposed.

Relations (ADR-058, ADR-060, ADR-061).

  • Corroboration, not grounds. A Proposed reading the claim cites as an instance, a parallel, prior art or “weak support” is corroborated_by. This freed most of the claims promoted in passes 13 to 16.
  • No edge. Each of these takes no edge, with the body citation as the link:
    • a rival account the claim’s test must exclude;
    • an attribution;
    • an unread or Unavailable work.
  • Dropped. A premise or ground the body never uses is dropped.
  • answers against bears_on. answers is for an offered answer to the question as posed. A part-answer gets an ADR-049 split: an Answered child, and a narrowed parent that stays Open. Anything else is bears_on.
  • Field, not prose. A body citation doing an objection’s work goes in objects_to.

Concessions (ADR-059). A conceded objection is recorded by revising its target. The target’s history note must name the objection, or the target carries a note headed ## Note of YYYY-MM-DD: filed with its concessions that names it. Promoting an objection makes its concession due: re-run the worklist after each cluster lands.

Kept Proposed. An entry kept Proposed gets one to three sentences under ## Note of YYYY-MM-DD: what holds it back, stated as a missing result or an unanswered objection. The heading is exact, because the worklist reads it.

The traps.

  • Inverted ground. Promoting a claim that a Proposed THEORY or LIT still lists under supports fails the lint as inverted_ground. List every such reader in the report, and move or remove the edge.
  • Answered but open. Promoting a claim that answers an Open QUESTION fails the lint unless the question changes status in the same change.

3. Rule, then apply

Rulings. The coordinator rules on each report. The owner delegated these rulings to the status review on 2026-10-29, so the pull request lists the interpretive ones for the owner to see, and merging it is the assent.

Applying. Each reviewer applies its own cluster. Clusters touch disjoint files, and each reviewer is told the others’ file lists. Reviewers do not run git or any luria command except luria new. The conventions:

  • Codes. Write codes bare. Never hand-write a link target.
  • Version and history. Any content change bumps version and appends a history entry. The entry is dated the day of the change, says what changed and why, cites the decision, and quotes any old title or condition verbatim. A ruling is worded “at the status review’s ruling (the owner delegated these rulings)”.
    • A blocker note alone takes no bump.
    • Removing or changing a converse edge does take one.
  • Converse edges. When an edge moves, remove its stale converse on the other entry by hand. The repair bot adds missing converses and never removes stale ones.
  • Promotion. Promoting a THEORY deletes its promote_when.
  • Titles. The # CODE: title heading matches the frontmatter title. A single-quoted YAML title doubles its apostrophes.
  • New entries. New entries come from luria new under a temporary code. Overwrite the scaffold with clean frontmatter, using QUESTION-303 and THEORY-1117 as models.
  • Checks. Check that every frontmatter parses.

4. Land it

Commits. Commit each cluster as its reviewer finishes, and push. Open the pull request with the first cluster, and pull before each push: the docs-check bot commits its repair onto the branch. Do not run luria lint, repair or index locally (ADR-054).

Before the last push.

  • Re-run the worklist on the branch. Its sections that must be empty include stale converses: an Active claim still listed under a Proposed reading’s supports.
  • File one workbench entry for the pass and one changelog fragment.

The pull request lists each cluster, the checks that were re-run, and the calls the owner may want to see. It is ready when docs-check is green on the last human commit; the bot’s repair commit sits on top of it.

After merging, wait for main’s docs-generate run to concretize the temporary codes before cutting the next branch.