Artifact Formalizer

Artifact formalizer (specification register)

Purpose

A reader of an essay-register artifact performs two passes: one to locate the requirement, one to decide whether a given sentence carried one. This skill removes the second pass. It has two modes.

Mode A prevents the defect; Mode B measures what Mode A missed.

Why in that order. Defective prose measured 5.1% of one corpus's words (731 of 14,288), while removing it required reading and editing all 14,288. Register also varied by authoring model on the same repository, so an unwritten standard is not a standard. references/measurement-baseline.md §5 carries both figures.

Scope boundary — what this is NOT

This is not a jargon or terminology tool. Domain terms are precise and stay verbatim: singleflight, RTM, дедлайн, throttle. Translating engineering vocabulary into business language for customer-facing documents is a different problem with a different audience, and this skill does not attempt it.

1. Red Flags (Anti-Rationalization)

A definition list, not a table: the reality column is prose, and documentation-standards §5.1 prescribes converting such a column rather than widening it.

• "I'll write it, then run the formalizer." That is the expensive order. §Purpose gives the measured cost. Mode A first.
• "The scan is clean, so the text is clean." Every zero is reported next to what the detector actually saw. Read DIAGNOSTICS: a corpus whose longest sentence equals the limit was written for the gate.
• "I'll add a pattern for this new phrasing." First ask whether a test in the authoring contract already forbade it. If it did, the lexicon gains a faster detector and nothing else changes. If it did not, the contract is amended (§6).
• "«Шов», «нога», «мост» — это термины проекта." A term appears in ARCHITECTURE.md, a public API, or a cited standard. Run --terms docs/ARCHITECTURE.md and let the scanner apply that test.
• "I formalized the section that was quoted at me." The pass that produced this skill's own worked example did exactly that. It left seventeen occurrences of one metaphor in the same task set. Coverage is per section: --sections.
• "The whole document reads badly — I'll rewrite it." Conforming sentences stay verbatim. Over-rewriting introduces errors that were not there.
• "I'll shorten it by trimming the requirement." Register changes, substance does not. Numbers, identifiers, obligations and scope limits survive verbatim.
• "This word is fine — I'll widen the threshold." A failing scan is fixed in the prose. Thresholds move only when a measurement moves them.
• "It flagged a false positive, so the rule is wrong." The scanner is advisory by design. Judge the finding; info exists because the author decides.

2. Capabilities

• Authoring contract. Six tests applied per sentence, and the licensed statement forms. The tests are stated as properties rather than as a word list, which is what lets them reach a phrasing the lexicon has never seen. That is a design intent, not a guarantee: §6 says what to do when a finding escapes all six.
• Self-probing detectors. Every run probe-tests each detector against a known positive and prints the result beside the findings. A rule whose pattern cannot match its own probe string is rejected at load time; a dead detector exits 2 instead of reporting a clean document.
• Measured diagnostics. A zero is never bare: sentence count, mean, observed maximum against the limit, cell counts, active lexicon size. PRESSED AGAINST THE LIMIT names a distribution written for the gate.
• The value that invalidates a scan is letters masked away. Read it before the findings. Masking blanks fenced blocks, comments, link targets and code spans, so a high share is correct in a code-heavy document and suspicious in a prose one. Corpus reference: median 22%, p95 63%. A share far above the p95 in a document with no fenced block means the scan measured little of what it was given. INPUT DEFECT lines name what stopped it — an unterminated comment, a backtick that never pairs. Both are facts about the document, so they are reported and exit 0 (ARCHITECTURE §7.4).
• All six rules reach a detector, three of them with declared recall limits (§5).
• Termhood applied mechanically. --terms ARCHITECTURE.md … downgrades rule-6 findings that occur in the declared sources, which is the guide's own test executed rather than remembered.
• Section worklist. --sections enumerates the document so "the reading pass was done" has a denominator.
• Data-driven markers. data/register-ru.json, data/register-en.json (schema register-rules/v1). Entry = marker + pattern + guidance + probe, plus optional flags, severity, rule and note. New language = new data file, zero code.
• Masking before matching. Code spans (any backtick run, across line breaks), fenced blocks (backtick or tilde, terminated or not), link targets and HTML comments are blanked at preserved length. Line numbers survive, and a marker quoted as an example is not reported as one used.
• Advisory exit codes. 0 on any number of findings. 2 for a broken instrument: an invalid rule file, thresholds that conflict or are inconsistent once merged, unreadable input, or a dead detector. 3 for a usage error, so a mistyped flag is not read as a dead detector.

3. Execution Mode

Hybrid. Mode A is prompt-driven judgement. Rule-file validation, detector probing and scanning are deterministic. Judging info findings and covering the recall gaps in §5 are yours.

Script Contract

Finding kinds: sentence_length, sentence_near_limit, cell_width, cell_sentences, emoji_severity, marker, reasoning, maxim, metaphor.

Every finding names the section it violates. Register findings carry a §5.5 rule number; cell_width and cell_sentences carry §5.1, which owns cell shape. Reporting them under a §5.5 rule number claimed an ownership this skill disclaims two paragraphs later.

4. Instructions

Mode A — authoring (default)

A1. Read references/authoring-contract.md. Identify the document's language; it is the author's choice and this skill never changes it.

A2. For each statement, pick its licensed form before writing the sentence. There is no step at which prose is generated and then repaired.

A3. Apply tests T1–T6 as you write. T6's target is the corpus mean of 5.8–15.4 words, not the 35-word failure bound.

A4. Hand off to Mode B on what you wrote. Findings there are residue, and §6 says what each one obliges.

Mode B — audit

B1. scan_register.py <file> --sections --terms <ARCHITECTURE.md>. Exit 2 means a broken rule file or a dead detector — fix that first; findings from a run with a dead detector are not a measurement.

B2. Read DETECTORS and DIAGNOSTICS before reading the findings. A clean scan whose diagnostics show PRESSED AGAINST THE LIMIT, or a low prose reaching rule 1 share, is not evidence of a conforming document.

B3. Fix every warn; judge each info. ровно, именно, exactly are load-bearing when they name an exact count or location.

B4. Walk the --sections worklist and read every section for the recall gaps in §5 — not only the sections that already carry findings.

B5. Re-scan. Then apply the specification test: could a reader who was not in the discussion verify each claim from the document alone?

5. Detector coverage, and what each detector does not reach

Stated so that a zero is read correctly. Rules 1, 2 and 5 are decided by the scanner. Rules 3, 4 and 6 have detectors with declared recall limits, and the reading pass owns the remainder.

• Rule 1 — two claims inside twelve words pass, because the detector counts words.
• Rule 3 — the detector needs the obligation and the causal connective in one sentence. Each vocabulary pattern declares its example in probes; a rule file holding a pattern that declares none does not load.
• Rule 4 — it recognises named personifications and maxim templates, nothing beyond them.
• Rule 5 — ✓ and ✗ (U+2713/U+2717) are excluded everywhere; ✅, ❌, ☑, ☒ and ☐ are excluded inside a table cell, where §5.1 governs. Outside a cell those five are still reported. Rule 5's second clause is why: a glyph carrying no severity at all is diff metadata, and does not belong in a specification either (references/formalization-guide.md rule 5).
• Rule 6 — a candidate list plus a string test against the sources given to --terms.

documentation-standards §5.1 is surfaced by the same scanner (cell_width, cell_sentences) and remains owned there.

6. Maintenance — how the lexicon stays a backstop instead of a list of yesterday's phrases

A new defective phrase found in the wild is triaged before any data file or detector is edited. Rules 1 to 4 govern that direction. Rule 5 governs a proposal to remove coverage.

1. Does a test T1–T6 already forbid it? If yes, the contract held and the author skipped it. Add the entry as a faster detector, and nothing else changes.

2. Does no test reach it? Then the finding is about the contract. Amend references/authoring-contract.md first — a test or a licensed form — and add the entry second.

3. Every entry ships with a probe. Validation rejects a pattern that cannot match its own probe, so a rule can never be added dead.

4. A rule ships only when a measurement supports it. Four properties expected to show drift did not, and six generic AI-writing tells measured at zero; references/measurement-baseline.md records them as non-rules so nobody re-proposes them from impression.

5. A narrowing states what it removes. Rules 1 and 2 route a defect no test reaches. A proposal to narrow a rule runs the other direction, and it ships only with three items:

   • the population it removes, measured over the declared scope, with the scope and the commit named;
   • the battery result after the change, since a narrowing that fails a case is a rule the battery still asks for;
   • one occurrence in scope that the narrowed rule still reports.

   A keep-class with no occurrence in scope is what references/measurement-baseline.md §4 declines to adopt, read from the other end. The figures are recorded there, so the next proposal reads them before it is written.

Where each register rule is narrowed. A plan states the surface it edits, and the surface differs by rule.

Rule 5 is the one with no data-file surface. _validate_entry rejects a rule number outside RULE_KIND, so no entry can carry it. A plan that routes a rule-5 narrowing through data/register-*.json names a step that cannot be carried out (WI-13 §7.5). Editing the code DEFAULTS to narrow rule 1 is the mirror error: both shipped data files declare that key, and load_rules merges the declared value over the default.

7. Safety Boundaries

Read-only: the scanner opens files and writes nothing. Rewrites in Mode B change register only, and never alter numbers, identifiers, obligations, scope limits or the document's language. Ambiguity is surfaced as an Open Question, not resolved. Reports and never fails a phase.

Why advisory. documentation-standards §4 records the mechanism it did measure: a gate that fails on correct documents is how gates get switched off. The second mechanism — an author writing to the bound rather than to the rule — is this skill's own measurement, and references/measurement-baseline.md §8 carries the figures.

The two non-advisory exits are a dead detector (2) and a usage error (3). Neither is a verdict on the document: the first says the instrument is broken, the second that the command was mistyped.

8. Validation Evidence

• Selftest: python3 scripts/selftest_scan.py — 192 cases. TC-META-01 asserts the battery ran that many, and a second case reads this number out of this file. A dropped case is then a red run rather than a smaller self-consistent total. They cover the schema, masking, every detector, the probe contract, diagnostics, sections, the terms downgrade and every exit code. Each finding of the WI-096 adversarial review carries its own regression pin.
• What the battery compares against. Every pin is a value declared in the battery, never one read out of the file under test. The lexicons are pinned by pattern and flags, the thresholds and the rule-5 glyph sets by value. Two surfaces live only in code: the line filter and the flag compilation. Each is held by a fixture. The battery edits the scanner in a throwaway copy and requires the roster to report the edit.
• Probe: --probe verifies 18 detectors across both shipped languages on every run. The roster is a literal of nine rows per shipped language. A detector class with no rule behind it prints a DEAD row and exits 2, instead of leaving the denominator with the numerator. Each rule-1 and rule-3 fixture runs twice: as a bare line, and as a list item. An entry declaring flags: "i" runs again against a case-flipped copy of its own probe.
• Behavioural evals (TASK 101): evals/selftest_evals.py — 59 cases, zero tokens, wired into CI. It covers what the battery above cannot: Mode A, measured as two arms of one prompt differing only in the contract, and the §5 recall gaps, measured against keys planted before the run. First campaign: warn 503 → 13 over ten documents per arm, with rule 1 not carried — the contract arm exceeds the 35-word bound in 3 repetitions of 3 on one case. Reading-pass recall 6 of 6 and no finding on the control. Figures, caveats and what they do not license: references/measurement-baseline.md §12. How to run one: evals/README.md.
• Corpus check and the full baseline, including the two detector defects found by measurement and pinned as regressions: references/measurement-baseline.md.

9. Quick Reference