Skip to content

v0.19.1

Latest

Choose a tag to compare

@HarperZ9 HarperZ9 released this 15 Nov 11:27
· 5 commits to main since this release
20510f4

Structure rules for a model's replies. A reply passed check with a texture
score of 0 on 2026-11-15, and its reader then read it back as a list of faults.
These rules report what the line checker passed.

Released 2026-11-15. This is the first version on PyPI with these changes.
Version 0.19.0 was tagged an hour earlier and its publish stopped at the
release gate, because its changelog section opened with the date line where
the gate wants a summary. Nothing was uploaded for 0.19.0. A test now holds
that rule before a merge, and the Discover check gates on a release branch.

  • New rule pack structure, on under the chat profile and in the house voice
    notes: structure-anaphora, structure-parallel, structure-triad,
    structure-pin, structure-parataxis, structure-pointer and
    structure-landing. A finding spans the whole run of sentences. The HIGH
    findings gate under chat. No other profile lists the pack, so author-voice
    and narrative text raises none of these findings.
  • chat now resolves to a domain profile that adds the pack to the published
    chat register. The published detector, its rules and its fingerprint are
    unchanged.
  • Under chat the cadence block adds longest_opener_run and
    repeated_openers_in_a_row. The published repetitive_openers flag is a
    whole-document ratio and stays as it is.
  • The MCP check and score tools take an optional profile. With one, they
    also return gate and blocking_count. With none, the default line rules run
    as before. A model that checks its own reply to a person should pass
    profile: "chat", and the tool descriptions now say so.
  • Every MCP check and score result carries rules_run: the profile, the
    rule packs that ran, and whether the structure rules were among them. On the
    day these rules were written a session read "clean" from the default rules as
    permission to send a reply. An unprofiled result now says in rules_run that
    the structure rules did not run. No existing field changed its meaning.
  • The texture score counts line rules only, and it is the published detector's
    figure, so rule-pack findings do not feed it. A blocked reply with 8 HIGH
    findings printed "texture 0/100" beside the failure. Decision: keep the
    texture score as it is and never show it alone when rule packs found
    something. A check result that ran rule packs carries pack_findings (counts
    by tier). The terminal headline reads rule-pack findings 12, texture 0/100 (line rules only), and the MCP result adds texture_scope. Dropping the
    texture figure from a blocked result was the other option. It would have left
    the bare zero in the JSON and in every unblocked result with MEDIUM pack
    findings, so the second named figure is the one that cannot mislead. The
    edit hook and the Stop hook print findings and no texture figure, and a test
    holds that.
  • The pin, pointer and landing rules read a list item as a short paragraph. A
    hand-back that closes a bullet ("That one is your call.") is a landing
    sentence. The run rules stay out of lists, and the wrapped lines under a
    list marker are read with their item, where they used to be read as a
    paragraph of fragments.
  • structure-landing/handback (MEDIUM) reports a hand-back before the end of a
    paragraph. A long sentence that ends within four words of the hand-back
    counts as one. "Their decision was final." no longer matches: the possessive
    form needs "is", "was" or the end of the sentence. A hand-back inside
    quotation marks is a mention and raises nothing.
  • A quoted span stays inside the sentence that quotes it: a full stop inside
    quotation marks ends no sentence, and neither does one inside a closing
    bracket when the text runs on in lower case.
  • structure-anaphora/pronoun-run (MEDIUM): exactly three sentences in a row
    that open on one personal pronoun report and do not gate, and an opener alone
    makes them no triad. This was the largest source of HIGH findings on human
    prose, and none of 52 sampled replies held one. HIGH findings fell from 14.7
    to 5.1 per 1,000 sentences on the human corpus and from 9.6 to 7.6 on four
    books. Four or more in a row still gate.
  • articulate check prints where a structure finding starts and ends
    (L1:85-L1:207) and the first and last words of its span. It used to print
    the start of the line.
  • tests/fixtures/structure/readback/ holds texts a person read and judged,
    with the spans the reader objected to and the rule the reader named. A test
    reports each rule's recall against those marks. The folder's README says how
    to add a case.
  • Series review adds corpus/sentence-template and corpus/ritual: a sentence
    ending or an opening that several documents by one author share. Report only.
    The corpus thresholds changed, so a corpus receipt from an earlier version
    reads Unverifiable.
  • Structure rules gives each rule, its thresholds,
    the measured rates on 52 replies and on human prose, the three worst false
    positives, and which of one author's 23 prose rules have a detector.