# Stage 2 - Writing, fact-checking, fixing, and auditing the skill (19 Opus agents)

Unabridged per-agent explanations from the ste100-write-skill workflow. Each fact-check finding was re-verified against the extracted spec text by the fixer before any edit was applied.

## File: skill

### Draft agent

**Result:** SKILL.md, 82 lines. Frontmatter (name: asd-ste100, third-person description with trigger terms STE, Simplified Technical English, ASD-STE100, controlled language, technical documentation/procedures/manuals), then 7 body sections in the required order: overview (controlled language, 53 rules in 9 sections, 875 approved words), When to use (procedural vs descriptive), How to write STE (6-step workflow with one verified grep example per dictionary file, relative paths), Core limits (14-row table citing rules 5.1, 6.3, 5.5, 6.6, 2.1, 3.2, 3.6, 5.2, 8.1, 8.4, 8.5, 8.6, 8.7, 4.5), Safety instructions (4 lines, warning/caution/note per 7.1-7.3 and 5.5), Check your text (11-item checklist with rule numbers), References (4 one-line entries), plus the GR-7 exclusion scope line. Prose is written in STE style: imperative instructions, active voice, sentences within the 20/25-word limits, no paragraph over 6 sentences, vertical lists, zero semicolons, zero contractions, recurring-error words avoided (obey not follow, make sure not ensure, do a check not check, but not however, thus not therefore).

**Steps:**

1. Read map.md for the standard's structure: Part 1 = 9 sections with 53 rules, Part 2 dictionary, and the full rule index used for every citation. 2. Read r1-front-intro.md for the overview facts: STE is a controlled natural language and international standard, "9 sections that contain 53 rules" (General introduction), Issue 9 dated 2025-01-15. 3. Read r5-sec5-6.md for the text types and limits: procedural writing gives instructions (Section 5), descriptive writing gives information and forbids the imperative (Section 6 introduction), 20 words per procedural sentence (5.1, safety instructions included per 5.1 detail), one instruction per sentence (5.2), imperative form (5.3), notes information-only with 25-word sentences (5.5), 25 words per descriptive sentence (6.3), one topic per paragraph (6.5), max six sentences per paragraph (6.6). 4. Read r6-sec7-8.md for Section 7 definitions (warning = risk of injury or death, caution = risk of damage to objects, two levels together = warning per 7.1; command/condition first per 7.2; risk explanation per 7.3) and Section 8 (no semicolon 8.1, colon in vertical list 8.4, parentheses one word 8.5, the seven one-word elements with rule-body wording "titles, headings, and text on placards and labels" and "proper nouns of individuals, groups, organizations, and geopolitical entities" 8.6, hyphenated words 8.7). 5. Read r8-dict-intro.md for the dictionary counts (875 approved, 1274 not approved, page 2-0-3) and UPPERCASE/lowercase headword convention. 6. Read supplement-graphics.md for the word-selection flowchart (page 2-0-16) that step 2 of the workflow points to via references/word-selection.md. 7. Grepped r3-sec2-3.md for exact wording of 2.1 (multi-word nouns of no more than three words), 3.2 (the six verb forms/tenses), and 3.6 (rule-box "only if the agent is unknown"), and r4-sec4-sentences.md for 4.3 (vertical lists) and 4.5 (article/demonstrative wording, used verbatim in the table). 8. Verified both lookup commands against the built reference files: grep -i "^ACCESS " dictionary-approved.md and grep -i "^accomplish " dictionary-unapproved.md both return their entries. 9. Wrote the prose in STE itself per rules 5.1/5.2/5.3, 6.3/6.5/6.6, 3.6, 4.3, 8.1, avoiding the dictionary's recurring-error words (r8 list), then checked the file: 82 lines, no semicolons outside the quoted "(;)" symbol, no contractions. 10. Added the GR-7 scope line per map.md section 5 (excluded on user instruction).

- Citation: General introduction (r1): 9 sections, 53 rules; controlled natural language; standard for technical documentation
- Citation: Part 2 Introduction page 2-0-3 (r8): 875 approved words, 1274 not approved
- Citation: Part 2 Introduction page 2-0-16 + supplement-graphics.md section 1: word-selection flowchart
- Citation: Rule 1.1, 1.2, 1.3 (checklist)
- Citation: Rule 2.1 (max three words in a multi-word noun)
- Citation: Rule 3.2 (six approved verb forms and tenses)
- Citation: Rule 3.6 (active voice; passive in descriptive writing only if the agent is unknown)
- Citation: Rule 4.3 (vertical lists, applied to the file's own prose)
- Citation: Rule 4.5 (articles and demonstrative adjectives before nouns)
- Citation: Rule 5.1 (max 20 words per procedural sentence, safety instructions included)
- Citation: Rule 5.2 (one instruction per sentence)
- Citation: Rule 5.3 (imperative form for instructions)
- Citation: Rule 5.5 (notes information-only, max 25 words per note sentence)
- Citation: Section 6 introduction (descriptive writing gives information, imperative not permitted)
- Citation: Rule 6.3 (max 25 words per descriptive sentence)
- Citation: Rule 6.5 (one topic per paragraph)
- Citation: Rule 6.6 (max six sentences per paragraph)
- Citation: Section 7 Definitions (warning = risk of injury or death; caution = risk of damage to objects)
- Citation: Rule 7.1 (word to identify the level of risk; two levels together = warning)
- Citation: Rule 7.2 (start with a clear and accurate command or condition)
- Citation: Rule 7.3 (explanation to show the risk or possible result)
- Citation: Rule 8.1 (no semicolon)
- Citation: Rule 8.4 (colon in a vertical list = period for word count)
- Citation: Rule 8.5 (text in parentheses counts as one word)
- Citation: Rule 8.6 (seven elements that count as one word, rule-body wording)
- Citation: Rule 8.7 (hyphenated words count as one word)
- Citation: Section 9 GR-7 (inclusive language, excluded per user instruction; map.md section 5)
- Citation: List of recurring errors, pages 2-0-17..2-0-18 (r8): word choices avoided in the file's own prose

### Fact-checker 1 (52 claims checked)

1. Read SKILL.md in full and map.md (master map) in full; listed digest/ and build/asd-ste100/references/ directories.
2. Verified the overview claims "9 sections that contain 53 rules" and "875 approved words" / dictionary alternatives against map.md sections 1-3 and the original spec text: full.txt line 1680 ("The writing rules (part 1) have 9 sections that contain 53 rules"), line 5455 ("875 approved words"), line 5458 ("words that are not approved (1274 words)").
3. Verified every rule number and limit in the Core limits table against map.md section 2 and the digests: 5.1 (max 20 words, procedures), 6.3 (max 25 words, descriptive), 5.5 (max 25 words per note sentence), 6.6 (max 6 sentences per paragraph), 2.1 (multi-word nouns max 3 words), 3.2 (six approved verb forms/tenses), 3.6 (active voice; passive in descriptive writing only if agent unknown), 5.2 (one instruction unless simultaneous actions), 8.1 (no semicolon), 8.4 (colon in vertical list = period for word count), 8.5 (parentheses = one word), 8.6 (elements counted as one word, matched to the rule-body variant), 8.7 (hyphenated words = one word), 4.5 (article/demonstrative adjective before nouns).
4. Confirmed the "(safety instructions included)" qualifier on rule 5.1 in the spec: full.txt lines 3633-3634 ("the maximum length of a sentence for procedures is 20 words. Warnings, cautions, and other safety instructions must also obey this rule") and r5-sec5-6.md line 32.
5. Confirmed the 25-word note limit belongs to rule 5.5: full.txt lines 3666 ("Notes (rule 5.5) do not obey rule 5.1... maximum length of a sentence in a note is 25 words") and 3820-3821 (under Rule 5.5), and rule 5.2's "unless two or more actions occur at the same time" at line 3670.
6. Verified the Section 6 claims in "When to use": r5-sec5-6.md lines 158-163 and full.txt lines 3942-3944 confirm descriptive writing gives information not instructions, the imperative form is not permitted, and the three example types (description of a system, reports, a note in a procedure).
7. Verified the Safety instructions section against r6-sec7-8.md and full.txt: warning = risk of injury or death and caution = risk of damage to objects in the Section 7 Definitions (full.txt line 4286); "two levels of risk together, use a warning" appears under Rule 7.1 (full.txt line 4337); Rule 7.2 = start with a clear and accurate command or condition; Rule 7.3 = give an explanation to show the risk or possible result; Rule 5.5 = notes give information only.
8. Verified all 11 checklist bullets' rule citations against map.md section 2: 1.1 (approved/technical noun/technical verb), 1.2 (part of speech), 1.3 (meaning), 3.2, 3.6, 5.3 (imperative instructions), 5.2, 5.1/6.3/5.5 limits, 6.5 (one topic per paragraph), 6.6, 2.1, 8.1, 8.4-8.7 (word-count rules per map section 4), 7.1-7.3, 4.5.
9. Verified the four references/ file pointers exist via ls: writing-rules.md, word-selection.md, dictionary-approved.md, dictionary-unapproved.md are all present in build/asd-ste100/references/.
10. Counted rule headings in writing-rules.md: exactly 53 "### Rule" headings covering 1.1-1.14, 2.1-2.2, 3.1-3.7, 4.1-4.5, 5.1-5.5, 6.1-6.6, 7.1-7.3, 8.1-8.7, 9.1-9.4 across 9 "## Section" headings — matches the "53 writing rules" description and the "Apply all 53 rules" step.
11. Ran both grep examples from the SKILL verbatim: grep -i "^ACCESS " dictionary-approved.md returns the ACCESS (n) entry; grep -i "^accomplish " dictionary-unapproved.md returns the accomplish (v) entry — both lookup patterns work against the actual file formats.
12. Checked the 875/1274 file descriptions against file contents: dictionary-approved.md has 932 entry lines (one per meaning; header states spec count 875) and dictionary-unapproved.md has 1319 entry lines (one per part of speech; header states spec count 1274). The SKILL quotes the spec's official counts (875/1274, spec page 2-0-3, full.txt lines 5455/5458), which map.md section 3 identifies as the authoritative figures with reader-sum deltas explained as a per-PoS counting convention — no discrepancy.
13. Verified the GR-7 scope note: map.md sections 2 and 5 confirm GR-7 is "Inclusive language", a General recommendation (not an STE rule), deliberately excluded per user instruction.
14. Result: 52 factual claims checked (rule numbers, numeric limits, counts, STE concepts, grep examples, file pointers); zero discrepancies — no wrong rule numbers, wrong limits, wrong counts, invented claims, or broken pointers.

### Fact-checker 2 (56 claims checked)

1. Loaded the exact rule statements from the digests: Rules 5.1, 5.2, 5.3, 5.5, 6.3, 6.6 from r5-sec5-6.md; Rule 3.6 (plus the Rule 3.3 past-participle-as-adjective allowance, \"this is not passive voice\") from r3-sec2-3.md; Rule 4.3 from r4-sec4-sentences.md; Rules 8.1, 8.4, 8.5, 8.6, 8.7 from r6-sec7-8.md.
2. Segmented SKILL.md into prose sentences and applied the exemptions: YAML frontmatter (lines 1-4), 6 headings, 14 table rows (lines 35-50), code spans (lines 24-25), markdown file-path links, and quoted text — 56 non-exempt prose sentences/list entries remained, and I checked every one.
3. Counted words per Section 8: parentheticals such as \"(Rule 7.1)\", \"(Section 5)\", \"(5.1)\" = 1 word each (8.5); numbers and abbreviations such as \"875\", \"53\", \"STE\", \"9\" = 1 word each (8.6); quoted text such as \"Safety instructions\" = 1 word (8.6); hyphenated words such as \"word-selection\", \"multi-word\", \"GR-7\" = 1 word each (8.7); a colon before a vertical list ends the sentence (8.4), so lead-ins like \"Make sure that your text obeys these rules:\" (8 words) were counted to the colon.
4. Checked every instruction sentence against Rule 5.1 (max 20 words), Rule 5.3 (imperative form), and Rule 5.2 (one instruction per sentence). All instructions are imperative (Apply, Identify, Select, Obey, Find, Write, Do, Make sure, Start, Then give, Read), each carries a single instruction, and the longest is 16 words (\"When you do not know if a word is approved, find it in the dictionary files:\") — condition-first per Rule 5.4 pattern, under the 20-word limit.
5. Checked every descriptive sentence against Rule 6.3 (max 25 words). The longest is 17 words (checklist item \"Sentences obey the limits: 20 words in procedures (5.1), 25 words in descriptions (6.3) and notes (5.5).\") — compliant. Intro sentences range 11-15 words.
6. Checked Rule 6.6 (max six sentences per paragraph) using the blank-line map: the intro paragraph (lines 8-11) has 4 sentences; the largest paragraph, \"Safety instructions\" (lines 54-57), has exactly 6 sentences — \"no paragraph has more than six sentences\" permits 6, so it passes at the limit; all other paragraphs and list items have 1-4 sentences.
7. Checked Rule 8.1 (no semicolon) by grep: the only \";\" in the file is inside the exempt table cell on line 45, which names the banned mark — no semicolon is used in prose.
8. Checked Rule 4.3 (vertical list for complex text): inline enumerations in non-exempt prose have at most 3 items in sentences of at most 17 words (lines 18, 22, 67, 72), below the \"long sentence with many different items\" trigger; the genuinely long enumerations sit in exempt table cells (e.g., the Rule 8.6 row, line 48) or already use vertical lists (steps 1-6, dictionary sub-list, checklist, references).
9. Checked Rule 3.6 (active voice) on all 56 sentences: \"is not permitted\", \"is approved\", \"are not approved\" pass as past participles used as adjectives after \"to be\" per Rule 3.3 (and mirror the standard's own Section 6 wording \"the imperative form of the verb is not permitted\"); the one failure is line 82's \"excluded on user instruction\", a reduced passive whose agent is known — reported as the single finding, severity minor because the sentence is descriptive, not an instruction.

**Findings (1):**

- **[minor]** Line 82: "Scope: this skill does not cover GR-7 (inclusive language), excluded on user instruction."
  - Claim in file: The trailing clause "excluded on user instruction" is a reduced passive construction ("[which was] excluded") with a known agent (the user/the skill author). Sentence length is compliant — 12 words counting GR-7 as one word (8.7) and "(inclusive language)" as one word (8.5) — but the voice is not active, and the past participle cannot pass as an adjective because it is neither before a noun nor after "to be/become/stay".
  - Spec says: Rule 3.6 (r3-sec2-3.md): "Use the active voice. In descriptive writing, you can use the passive voice only if the agent is unknown." Rule 3.3 permits a past participle as an adjective only before a noun or after "to be," "to become," or "to stay" — "excluded" after a comma fits neither permitted position.
  - Proposed fix: Scope: this skill does not cover GR-7 (inclusive language) because the user excluded it. (13 words, active voice, descriptive limit 25 obeyed.)

### Fixer (applied 1, rejected 0)

1. Read SKILL.md and confirmed line 82 contained the flagged text "Scope: this skill does not cover GR-7 (inclusive language), excluded on user instruction."
2. Verified Rule 3.6 in spec/r3-sections2-3-nounclusters-verbs.txt (summary at lines 205-206, detail at lines 404-409): "Use the active voice. In descriptive writing, you can use the passive voice only if the agent is unknown." The clause "excluded on user instruction" is a reduced passive whose agent (the user) is known, so the passive exception does not apply — violation confirmed.
3. Verified Rule 3.3 in the same spec file (lines 269-275): a past participle may act as an adjective only "Before a noun" or "After a verb form of the verbs 'to be,' 'to become,' or 'to stay.'" "Excluded" after a comma is in neither position; also cross-checked Rule 3.2's approved-forms list (lines 198-199), which permits the past participle only "(as an adjective)" — so the participle cannot pass as an approved usage.
4. Verified the word-count claims: Rule 8.5 (parenthetical text counts as one word) and Rule 8.7 (hyphenated words count as one word) in spec/r6-sections7-8-safety-punctuation.txt (lines 176, 185; detail at 350, 521), and the 25-word descriptive limit Rule 6.3 in spec/r5-sections5-6-procedural-descriptive.txt (lines 325, 479). Original sentence = 12 words, replacement = 13 words — both compliant, matching the finding's arithmetic.
5. Cross-checked the digest r3-sec2-3.md as secondary (rule summaries at lines 94-98; Rule 3.3 at line 147; Rule 3.6 at line 220) — digest wording matches the spec, confirming the finding's citation of r3-sec2-3.md.
6. Checked the fix's vocabulary against the dictionary pages: BECAUSE (conj) is approved (d02-dict-p165-180.txt line 513) and INSTRUCTION (n) is approved (d09-dict-p277-292.txt line 434). Noted that "exclude (v)" is an unapproved dictionary entry (d06-dict-p229-244.txt line 564, approved alternatives NOT INCLUDE / NOT USE), but the word was already present in the original line, the finding does not cover word approval, and the mandate forbids changing content no finding covers — the fix introduces no new vocabulary and converts the participle to active simple past, an approved tense under Rule 3.2.
7. Applied the finding's fix with Edit: line 82 now reads "Scope: this skill does not cover GR-7 (inclusive language) because the user excluded it." (13 words, active voice, descriptive limit obeyed). Applied 1 of 1 findings; none rejected; no other content touched.

## File: rules

### Draft agent

**Result:** Complete ASD-STE100 Issue 9 Part 1 writing-rules reference, 459 lines: header with source and 9-section/53-rule structure plus at-a-glance limits; all 53 rules as "### Rule N.N - <title>" entries with normative statements (numeric limits verbatim: 20 words procedural, 25 words descriptive/notes, 3-word multi-word nouns, 6-sentence paragraphs, 20/25 before and after list colons), condensed details/exceptions, and one Non-STE → STE example pair each; Rule 1.5 with all 22 technical-noun categories (names + 3-5 examples, including the colors comparative/superlative prohibition and category 9 symbols ° Ω ″); Rule 1.12 with all 4 technical-verb categories and sub-groups 1a-f, 2a-c, 3a-f, 4; Section 9 followed by GR-1..GR-8 (GR-7 rendered only as "Not covered in this skill (excluded on user instruction)"); Appendix A subject-to-rule index with 113 rows; Appendix B with 7 lines on Issue 9 changes. Verified by grep: 53 rule headings, no expected number missing, 9 exact section titles, 115 table lines (113 data rows + header + separator).

**Steps:**

1. Read digest/map.md to fix the canonical structure: Issue 9 (2025-01-15), Part 1 = 9 sections / 53 rules (14+2+7+5+5+6+3+7+4), GR-1..GR-8 advisory, section page ranges, and the collected sentence-length limits (5.1, 5.5, 6.3, 6.6, 8.4); adopted the rule-box wording flagged canonical for 3.6 ("if the agent is unknown") and 4.3 ("complex text").
2. Read r1-front-intro.md for the header facts (subtitle, "9 sections that contain 53 rules"), the complete subject-to-rule index (SRI-1..SRI-4) reproduced in Appendix A, and the Highlights change log used for Appendix B.
3. From r2-sec1-words.md wrote Rules 1.1-1.14: kept all 22 Rule 1.5 categories with 3-5 example words each (incl. category 17 colors note and new categories 21-22) and all 4 Rule 1.12 categories with sub-groups 1a-f, 2a-c, 3a-f, 4; picked one example pair per rule (e.g. 1.2 operable/operate, 1.10 brick/OFF, 1.13 ream).
4. From r3-sec2-3.md wrote Rules 2.1-2.2 (three-word limit verbatim; Methods 1 and 2) and 3.1-3.7 (six approved forms/tenses table condensed, banned tenses, four passive-to-active methods, switching-relay pair).
5. From r4-sec4-sentences.md wrote Rules 4.1-4.5, including the vertical-list mechanics for 4.3 and the article-placement series nuance for the new Rule 4.5.
6. From r5-sec5-6.md wrote Rules 5.1-5.5 (max 20 words, notes max 25 words, note-vs-instruction pair) and 6.1-6.6 (max 25 words, six-sentence paragraph limit, ILS/smartphone/fuel-manifold examples).
7. From r6-sec7-8.md wrote Rules 7.1-7.3 (warning/caution definitions, oxygen-tubes reclassification pair) and 8.1-8.7 (semicolon ban, 5 hyphen categories, 7 parenthesis uses, colon 20/25 limits, all 7 one-word count elements with word counts, incl. the reconstructed "10 Ω" example).
8. From r7-sec9-practices.md wrote Rules 9.1-9.4 and GR-1..GR-8, writing GR-7 exactly as instructed ("Not covered in this skill (excluded on user instruction)") and mirroring that note on the Appendix A index row.
9. Used supplement-graphics.md to confirm Rule 1.5 category 9 symbols (° , Ω, ″ from page 1-1-7/pdf 51) and to describe the Rule 3.6 Method 1 agent/subject inversion diagram (page 1-3-7/pdf 73) inside the 3.6 entry.
10. Wrote Appendix B (7 lines) from r1 Highlights + map: ISO harmonization/subtitle change, technical noun replaces technical name, Rule 4.5 formerly 2.3, Rule 1.5 categories 21-22, Rule 1.12 restructure (new category 4), Rule 8.6.2/8.6.7 changes, GR-7/GR-8 new.
11. Verified with grep/wc: 459 lines; exactly 53 "### Rule" headings with every expected number 1.1-9.4 present; 9 exact "## Section" titles; GR-7 exact line; 113 index rows; numeric limits spot-checked verbatim.

- Citation: Rules 1.1-1.14 (r2-sec1-words.md, pages 1-1-1..1-1-18)
- Citation: Rules 2.1-2.2 and 3.1-3.7 (r3-sec2-3.md, pages 1-2-1..1-3-10)
- Citation: Rules 4.1-4.5 (r4-sec4-sentences.md, pages 1-4-1..1-4-10)
- Citation: Rules 5.1-5.5 and 6.1-6.6 (r5-sec5-6.md, pages 1-5-1..1-6-8)
- Citation: Rules 7.1-7.3 and 8.1-8.7 (r6-sec7-8.md, pages 1-7-1..1-8-8)
- Citation: Rules 9.1-9.4 and GR-1..GR-8 (r7-sec9-practices.md, pages 1-9-1..1-9-14)
- Citation: Subject-to-rule index SRI-1..SRI-4 and Issue 9 Highlights HI-1..HI-6 (r1-front-intro.md)
- Citation: Rule 1.5 category 9 symbols page 1-1-7 (pdf 51) and Rule 3.6 Method 1 diagram page 1-3-7 (pdf 73) (supplement-graphics.md)
- Citation: Structure totals and canonical wording variances (map.md sections 2 and 6)

### Fact-checker 1 (310 claims checked)

1. Read writing-rules.md Sections 1-3 (lines 9-155) plus the header structure claim (line 4): 'Section 1: 14, Section 2: 2, Section 3: 7' matches the spec rule summaries (r2 p.1-1-1 lists Rules 1.1-1.14; r3 p.1-2-1 lists 2.1-2.2; r3 p.1-3-1 lists 3.1-3.7). No rule missing, none extra, all numbers/titles verbatim against the spec rule boxes.
2. Section 1 rule-by-rule against r2-section1-words.txt (pages 1-1-2 to 1-1-17): verified title, statement (limits verbatim, e.g. 1.9 'not more than three words'), and every example sentence by locating it in the .txt — use/engine/ream (1.1, p.1-1-2); clean as v+adj and operable→'can operate' (1.2, p.1-1-3); follow='come after, go after' and obey (1.3, p.1-1-4); REMOVE (v) REMOVES/REMOVED/REMOVED and SLOW (SLOWER, SLOWEST), more/most note (1.4, p.1-1-4/5); base→bottom vs 'base of the triangle is 5 cm' category 7 and 'relative angular position' (1.6, p.1-1-9/10); 'Oil the steel surfaces'→'Apply oil', drill as both TN and TV (1.7, p.1-1-11); touchscreen/home button (1.8, p.1-1-11); four screws/flange/cover with index numbers (1.9, p.1-1-12); brick→'set the router to OFF', IT slang (1.10, p.1-1-12); servo control unit/actuator/control unit → actuator in all 3 steps (1.11, p.1-1-13); detect→find + 'security scanner detects metal objects' (1.12, p.1-1-15); 'Give the hole 0.20-inch ream'→'Ream the hole to a 0.20-inch dimension', reamed-hole adjective exception, plate as TV+TN (1.13, p.1-1-16); colour→color and quoted-text/Rule 8.6 exception (1.14, p.1-1-17). All genuine.
3. Rule 1.5: verified all 22 technical-noun categories (spec pages 1-1-5 to 1-1-8) — every category number and title verbatim (incl. 21 Law and regulations, 22 Animals, plants, and other life forms), all ~110 quoted example words found in the corresponding spec category lists (incl. formula 'C = (A - B) - 0.063 mm' in cat. 7), the colors note (adjectives treated as technical nouns; blacker/the reddest not permitted) present, 'twenty-two' verbatim, and the 'illustrative, not a full list' caveat matches the spec note on p.1-1-8. Category 9 symbols — degree (°), ohm (Ω), second (″) — confirmed against supplement-graphics.md section 2 (page 1-1-7, pdf 51).
4. Rule 1.12: verified all 4 technical-verb categories and all 15 subcategories (1a-1f, 2a-2c, 3a-3f, 4) against spec pages 1-1-13/1-1-14 — titles verbatim, all 81 example verbs present in the spec lists (file lists are honest subsets of longer spec lists, e.g. 1d omits decay/normalize/vaporize), 'four categories' verbatim, category 4 'only for legal and regulatory texts' restriction and the prefer-approved-dictionary-verb guidance (pp.1-1-14/15) faithful.
5. Section 2 against r3 (pages 1-2-1 to 1-2-4): Rule 2.1 title + three-word maximum + prepositions 'of, on, in, for' + runway-light-connection example with word counts (5 words → 1 word, 1 word, and 3 words) all verbatim; Rule 2.2 two-part title with both methods verbatim, Method 1 (ramp service door safety connector pin → safety connector pin; abbreviations unnecessary at ≤3 words) and Method 2 (hyphenated words count as one word; no hyphen groups >3 words; main-gear-door retraction-winch example; inward-outward valve keep-the-hyphen exception) all confirmed.
6. Section 3 against r3 (pages 1-3-1 to 1-3-9): 3.1 REMOVE/GIVE dictionary entries verbatim; 3.2 all six approved forms/tenses verbatim plus the (to) adjust paradigm and banned tenses (present perfect, past perfect, progressive) exact; 3.3 condition-not-passive, before-noun/after to be-to become-to stay, permitted/damaged adjectives, both STE examples verbatim; 3.4 'has adjusted'→'adjusted' example verbatim; 3.5 permitted uses (titles Cleaning/Troubleshooting; air-conditioning system, grinding wheel, welding torch) and the exact approved '-ing' inventory (nouns lighting/opening/routing/servicing; adjectives mating/missing/remaining; pronoun something; preposition during); 3.6 title matches the rule box ('only if the agent is unknown'), by-whom-or-by-what test, all four passive→active methods with their spec examples ('The test can be continued by the operator.'→'Continue the test.'; you/we; 'During transmission, the data was corrupted.' + 'something' agent), and the Method 1 diagram brackets '[The main gear leg] [is held by] [the side stay]' → '[The side stay] [holds] [the main gear leg]' confirmed against supplement-graphics.md section 3 (page 1-3-7, pdf 73) including the switching-relay example.
7. Chased the two details in Sections 1-3 that do not appear in the r2/r3 extracts to their real source in the original spec corpus: 'per Merriam-Webster' in Rule 1.14 (r8-dictionary-introduction.txt p.2-0-3: 'The dictionary uses American English spelling as defined in the Merriam-Webster's dictionary'; also r1-frontmatter-intro.txt line 558) and Rule 3.1's 'four verb types (regular, irregular, irregular auxiliary, defective modal)' (r8 p.2-0-5: 'Verbs in the dictionary can be: Regular verbs, Irregular verbs, Irregular auxiliary verbs, Defective modal verbs'), which the file correctly attributes to the Part 2 dictionary introduction — both genuine, nothing invented.
8. Single surviving deviation: Rule 3.7's ohmmeter example is labeled 'Non-STE → STE', but spec p.1-3-9 labels that pair 'Do not write / WRITE' and explicitly says all those sentences are in STE; the spec's only Non-STE for 3.7 is 'Check the laptop battery.' Reported as a minor finding with a relabeling fix. Everything else in Sections 1-3 checked out clean.

**Findings (1):**

- **[minor]** Rule 3.7 example, writing-rules.md line 154
  - Claim in file: The example pair is labeled Non-STE → STE: 'Non-STE: "The ohmmeter gives an indication of 450 ohms." → STE: "The ohmmeter shows 450 ohms."'
  - Spec says: r3-sections2-3-nounclusters-verbs.txt page 1-3-9 labels this exact pair 'Do not write: / WRITE:' and then states verbatim: 'In the examples, all sentences are in STE but those with direct verbs describe the action more clearly.' The spec's only Non-STE example under Rule 3.7 is 'Check the laptop battery.' → STE 'Do a check of the laptop battery.' So the ohmmeter sentence is valid STE (dispreferred), not Non-STE.
  - Proposed fix: Relabel the ohmmeter pair as 'Do not write → WRITE' (matching the convention the file already uses for Rule 4.1) and optionally note that both sentences are STE but the direct verb is clearer; keep 'Check the laptop battery.' (Non-STE) → 'Do a check of the laptop battery.' (STE) as the genuine Non-STE example — the file's detail text on line 153 already frames that one correctly.

### Fact-checker 2 (57 claims checked)

1. Read the target file /tmp/claude-1000/-home-cfcosta-Notes/f3cf6639-a28c-403a-997d-fc0414dcd4e4/scratchpad/build/asd-ste100/references/writing-rules.md (Sections 4-6, lines 158-236, plus header claims on lines 4-5) and both spec extracts: r4-section4-sentences.txt (spec pages 1-4-1 thru 1-4-10) and r5-sections5-6-procedural-descriptive.txt (spec pages 1-5-1 thru 1-5-8 and 1-6-1 thru 1-6-8).
2. Verified all 16 rule numbers and titles verbatim against the spec rule boxes and each section's "Summary of the rules": 4.1-4.5 (pages 1-4-1, 1-4-3, 1-4-4, 1-4-7, 1-4-8), 5.1-5.5 (pages 1-5-1 thru 1-5-5), 6.1-6.6 (pages 1-6-1 thru 1-6-7). No rule missing, none invented; header's per-section rule counts (Section 4: 5, Section 5: 5, Section 6: 6) are correct. Note: the md's 4.3 title "complex text" matches the spec's rule box (page 1-4-4) even though the Section 4 summary page says "complex texts" - the spec is internally inconsistent, md follows the rule box.
3. Verified the numeric limits are exact everywhere they appear: 20 words in 5.1 (page 1-5-1, incl. "warnings, cautions, and other safety instructions must also obey"), 25 words for note sentences in 5.5 (light-bulb note page 1-5-2 and rule body page 1-5-5), 25 words in 6.3 (page 1-6-4, with the "descriptive text is more complex than procedural text" rationale), 6 sentences per paragraph in 6.6 (page 1-6-7), and the header "at a glance" line (md line 5) restating all four.
4. Verified each rule's condensed statement clause-by-clause against the spec body: 4.1 (imperative in procedures, one topic per descriptive sentence, never abstract), 4.2 (no omitted nouns/verbs/subjects/articles, no contractions), 4.3 (colon after lead-in, item markers, uppercase start, period rules, no comma/semicolon, period on last item, no mixing procedural/descriptive, DO NOT per item, same level with parenthesis flattening, items connect to lead-in), 4.4 (connecting words and/but/then/thus, phrases "as a result"/"at the same time", demonstrative adjectives, after work steps and in safety instructions), 4.5 (no articles in general statements, series-of-items article semantics, no definite article before alphanumeric identifier 36L7), 5.1-5.5 (incl. 5.3 "must" exception, 5.4 comma-position meaning, 5.5 bans on instructions/requirements/limits/tolerances/results, WARNING/CAUTION escalation, notes-removed test), 6.1-6.6 (gradual information, key words unchanged, topic sentences, one topic per paragraph, divide >6-sentence paragraphs).
5. Traced all 16 example pairs to the spec verbatim, including embedded numbers: 5.1 preservation-oil counts 25/10/16 words with "0.24 inches" vs "0.24 in" preserved (page 1-5-2); 6.3 smartphone counts 31/15/16 (page 1-6-4); 6.6 fuel-manifold paragraph counts 1,2,5,4,4,2 (page 1-6-7); 5.2 leakage limit 0.5 cc/minute (page 1-5-3); 4.5 identifiers (15)/(14)/(13)/(12) and 36L7 (pages 1-4-8/1-4-9); 4.3 wheel-assembly list with period only on last item (page 1-4-4); 6.2 key-word table claims (sentence 2 reuses "system/show/pilot", sentences 3-4 reuse "localizer course"/"glideslope path", pages 1-6-3); 6.4 three Lightning Strike paragraphs (page 1-6-5); 6.5 two ILS topics, sentences 1-5 and 6-7 (page 1-6-6). All genuine, nothing invented.
6. Result: zero critical findings; two minor condensation deviations, both in the Rule 4.3 detail line (md line 171) - the omitted spec bullet about using an article before each list item's subject noun, and the item-marker set narrowed from "number, letter, punctuation mark, or symbol" to the spec's four illustrative examples.

**Findings (2):**

- **[minor]** writing-rules.md:171 (Rule 4.3 details, vertical-list checklist)
  - Claim in file: The condensed checklist for making a vertical list omits the spec's bullet requiring an article before the noun that is the subject of each item; the md reproduces the other seven formatting bullets but not this one.
  - Spec says: r4-section4-sentences.txt lines 156-157 (page 1-4-4): "Where applicable, use an article before the noun that is the subject of each item in the vertical list."
  - Proposed fix: Add to the Rule 4.3 detail line: "where applicable, use an article before the noun that is the subject of each item".
- **[minor]** writing-rules.md:171 (Rule 4.3 details, item identifiers)
  - Claim in file: The md says "identify each item with a dash, bullet, letter, or number," presenting the spec's four illustrative examples as the complete set of permitted markers.
  - Spec says: r4-section4-sentences.txt lines 149-154 (page 1-4-4): "Identify each item in the vertical list with a number, letter, punctuation mark, or symbol. For example, you can use: A dash (-) / A bullet point (•) / A letter (a, b, c…) / A number (1, 2, 3…)." (with a pointer to applicable specifications/style guides for which marks or symbols to use).
  - Proposed fix: Reword to: "identify each item with a number, letter, punctuation mark, or symbol (for example a dash, bullet point, letter, or number)".

### Fact-checker 3 (148 claims checked)

1. Read the target file and the three spec extracts: r6-sections7-8-safety-punctuation.txt (pages 1-7-1..1-8-8), r7-section9-practices.txt (pages 1-9-1..1-9-14), and r1-frontmatter-intro.txt (Highlights HI-1..HI-6, TOC, subject-to-rule index SRI-1..SRI-4, General introduction pages i-viii).
2. Section 7: verified the definitions paragraph (warning = risk of injury or death; caution = risk of damage to objects; other industries' words "danger/attention/notice" and refs ISO 45001:2018, ANSI Z535, ISO 3864 — spec page 1-7-1) and rules 7.1-7.3 titles verbatim vs the summary box. Rule 7.1's risk-decision list (injury/death → warning; damage to machines, tools, or equipment → caution; both levels together → warning) and the oxygen-tubes caution→warning example match page 1-7-2 exactly; 7.2's command-first (solvent) and condition-first (spray paint) examples match page 1-7-3; 7.3's explanation guidance and assemble-unit caution match page 1-7-4. Warning/caution distinction is stated correctly in both the definitions and 7.1.
3. Section 8: verified titles 8.1-8.7 vs summary page 1-8-1 and each rule box. 8.1 semicolon rationale + example (page 1-8-2); 8.2's five hyphen categories with spec-listed example terms and the hyphen-vs-dash note (page 1-8-2); 8.3's seven parenthesis uses and all seven examples verbatim (pages 1-8-3..1-8-4) — found the title truncation flagged above; 8.4 word-count convention exact (max 20 words before colon procedural / 25 descriptive; each item a new sentence at 20/25; example counts 13/2/2/3/4, page 1-8-4); 8.5 exact (parenthetical = 1 word, counted separately as its own sentence; identifiers/abbreviations = 1 word; 10-word/5-word example, page 1-8-5); 8.6 all seven element classes and every example count exact (6, 4, 5, 7, 4, 9, 1, 16, 1, and 8 words; step-number exclusion note; pages 1-8-5..1-8-8); 8.7 exact (soap-and-water 7 words; Main-gear-door retraction-winch handle 3 words, page 1-8-8).
4. Section 9: verified 9.1 (word-for-word replacement requires same part of speech and unchanged meaning; the four numbered reasons for a different construction; the two questions; the "primary objective" sentence verbatim; sight-gauge example — pages 1-9-1..1-9-5); 9.2 (approved-meaning column, usually one meaning/one part of speech, "flush" verb/adjective exception, wear→use example with the friction meaning — pages 1-9-5..1-9-7); 9.3 (phrasal-verb ban, not listed as "not approved", few approved phrasal verbs "put on"/"come on" with restricted meanings, put out→extinguish example — pages 1-9-7..1-9-8); 9.4 (consistent wording, the two STE-correct lubrication variants, the 8-step body-assembly example — pages 1-9-8..1-9-9).
5. GR-1..GR-8: verified all eight titles vs the Section 9 summary box and each entry's condensed guidance vs pages 1-9-9..1-9-13, and that the GR block is labeled "not STE rules" per the spec's General recommendations intro. Confirmed GR-7 (line 324) shows only the title plus exclusion note and that none of the spec's GR-7 content (page 1-9-13: gender-neutral language requirements, prohibited pronouns, EIGE/UN references) appears anywhere in the file — requirement satisfied.
6. Appendix A: compared all 113 rows of the file's table against the spec index (pages SRI-1..SRI-4) row by row, including correct OCR normalizations (Abbreviations "1,5"→1.5; Hyphens "8.2. 8.7"→"8.2, 8.7"), the # convention note, and quirky-but-faithful spec rows (Simple future/past/present tense → 3.3 as printed in the spec). 112 rows match the spec exactly; the 113th is the "Inclusive language | 9 – GR-7" row at line 372, which the requirement says must be absent — flagged as the critical finding.
7. Appendix B: verified each of the 7 Issue 9 change bullets against the Highlights (HI-1..HI-6) and history table (page iv): full replacement of prior issues + subtitle "Standard for technical documentation" + 2025-01-15 date + ISO-harmonization focus (HI-1, page iv); technical noun replaces technical name with definitions in Rule 1.1 (HI-2); Rule 2.3 → 4.5 move (HI-3/HI-4); Rule 1.5 categories 21 and 22 added with revised titles and explanatory texts (HI-2); Rule 1.12 category 3 subcategories replacing former category 4 plus new category 4 (HI-3, corroborated by index row "Law and regulations — 1.5, 1.12"); Rule 8.6.2 revised unit word count and 8.6.7 proper nouns added (HI-6); GR-7 and GR-8 new (HI-6). All accurate.
8. Cross-checked the file's structure claims relevant to scope (53 rules total; Section 7: 3, Section 8: 7, Section 9: 4; 8 GRs explicitly not STE rules) against "Guide to the writing rules" (page vii, "9 sections that contain 53 rules") and the three section summary boxes — all correct.

**Findings (2):**

- **[critical]** Appendix A - Subject-to-rule index, line 372
  - Claim in file: The index table includes the row "| Inclusive language | 9 – GR-7 (content not covered in this skill) |".
  - Spec says: The spec's subject-to-rule index (page SRI-2) has row "Inclusive language — 9 – GR-7", but the build requirement states this row must be absent from the compiled file because inclusive-language/GR-7 content is excluded on user instruction (parallel to the GR-7 body requirement that its content must NOT appear).
  - Proposed fix: Delete the "Inclusive language" row from the Appendix A table entirely; the body's GR-7 line (line 324) already records the exclusion.
- **[minor]** Rule 8.3 heading, line 270
  - Claim in file: The rule title lists the second parenthesis use as "to include letters or numbers that identify items".
  - Spec says: Both the Section 8 rule summary (page 1-8-1) and the Rule 8.3 rule box (page 1-8-3) state the second use in full: "To include letters or numbers that identify items on an illustration or in a text".
  - Proposed fix: Append "on an illustration or in a text" to the second listed use in the Rule 8.3 title so the rule statement matches the spec wording.

### Fixer (applied 5, rejected 0)

1. Re-verified Finding 1 against spec/r3-sections2-3-nounclusters-verbs.txt lines 554-575 (Rule 3.7, page 1-3-9): the ohmmeter pair is labeled "Do not write:/WRITE:", the spec states verbatim "In the examples, all sentences are in STE but those with direct verbs describe the action more clearly", and the only Non-STE/STE pair is "Check the laptop battery." / "Do a check of the laptop battery." Applied: relabeled line 154 to "Do not write → WRITE" (the convention the file already uses for Rules 4.1 and 5.4) and added the note that both sentences are STE with the direct verb clearer; left line 153's laptop-battery framing untouched per the finding. 2. Re-verified Finding 2 against spec/r4-section4-sentences.txt lines 156-157 (Rule 4.3, page 1-4-4): the bullet "Where applicable, use an article before the noun that is the subject of each item in the vertical list" exists in the spec but was absent from md line 171. Applied: inserted "where applicable, use an article before the noun that is the subject of each item" into the checklist at its spec position (after the uppercase-letter bullet, before the period bullets). 3. Re-verified Finding 3 against spec/r4-section4-sentences.txt lines 149-154 and 163-164 (Rule 4.3, page 1-4-4): the normative marker set is "a number, letter, punctuation mark, or symbol"; dash/bullet point/letter/number are introduced only with "For example, you can use". Applied: reworded line 171 to "identify each item with a number, letter, punctuation mark, or symbol (for example a dash, bullet point, letter, or number)". 4. Re-verified Finding 4 against spec/r1-frontmatter-intro.txt line 385 (Subject-to-rule index row "Inclusive language — 9 – GR-7", page SRI-2) and the build constraint in digest/map.md (line 175: index "reproduced (one inclusive-language row omitted per user constraint)"; section 5 item 3: "row omitted from the reproduced index table") plus digest/r1-front-intro.md's USER-CONSTRAINT SKIP NOTE: the row must be absent from the compiled file. Applied: deleted the "| Inclusive language | 9 – GR-7 ... |" row from Appendix A; the body's GR-7 exclusion line (now line 323) remains as the sole record of the exclusion. 5. Re-verified Finding 5 against spec/r6-sections7-8-safety-punctuation.txt lines 165-172 (Section 8 rule summary, page 1-8-1) and lines 260-268 (Rule 8.3 rule box, page 1-8-3): both state the second use in full as "To include letters or numbers that identify items on an illustration or in a text". Applied: appended "on an illustration or in a text" to the second use in the Rule 8.3 heading. 6. Ran a final grep over the repaired file confirming all four edits (findings 2 and 3 share the line-171 edit region) and confirming no content outside the findings changed; all 5 findings applied, 0 rejected.

## File: select

### Draft agent

**Result:** 252-line word-selection reference for the ASD-STE100 Issue 9 skill. Contains: (1) header citing Part 2 Introduction pages 2-0-3..2-0-19 with the 875-approved/1274-non-approved word counts, TN/TV markers, and Merriam-Webster American English conventions; (2) "How to select words" — the complete page 2-0-16 flowchart as a numbered decision list (steps 1-6 plus 5a) with the exact logic dictionary? -> approved? -> part of speech? -> meaning? -> use / alternative -> word-for-word replacement or different construction / technical noun-verb -> project glossary or do not use, plus the 4 terminal states; (3) "The dictionary format" — the 4 columns, UPPERCASE=approved / lowercase=not approved, (TN)/(TV) markers, all 8 parts of speech, and forms of approved nouns/adjectives/adverbs; (4) "Verbs in STE" — approved verb forms and the 4 verb types (regular ADAPT, irregular GIVE, irregular auxiliary BE, defective modal CAN/WILL) with their STE examples; (5) "Help categories" — all 4 (recommendations, restricted meaning, single-context, important information) with examples; (6) "Recurring errors" — the complete 39-row Non-STE -> STE table (machine-verified count: 39); (7) "Approved verbs" — the complete 208-verb list grouped by letter with per-letter counts (machine-verified count: 208), noting J, N, Q, V, X, Y, Z have none. Both list counts, the 26 letter groups, and the presence of all 7 flow steps were verified by script.

**Steps:**

1. Read the digest of the Part 2 Introduction (r8-dict-intro.md, covering dictionary pages 2-0-3 through 2-0-19) and the supplement transcription of the "How to select words correctly" flowchart graphic (supplement-graphics.md section 1, page 2-0-16), which closes the gap flagged in the r8 digest (the flowchart is graphic-only and absent from pdftotext).
2. Wrote the header from the General introduction on page 2-0-3: 875 approved words, 1274 non-approved words with approved alternatives, no TN/TV headwords but (TN)/(TV)-marked alternatives, no antonyms, Merriam-Webster American English; noted the externally-stated material the intro references (Rule 1.5 technical nouns, Rule 1.12 technical verbs, Part 1 Section 1 parts of speech, Part 1 Section 3 verbs).
3. Transposed the page 2-0-16 flowchart into the numbered decision list, preserving the transcribed logic exactly: step 1 in-dictionary? (No -> step 6), step 2 approved/UPPERCASE? (No -> step 5), step 3 same part of speech? (No -> select alternative -> 5a), step 4 meaning correct? (Yes -> use the word; No -> step 6), step 5/5a alternative same part of speech? (Yes -> word-for-word replacement; No -> different sentence construction), step 6 technical noun/verb per Rules 1.5/1.12? (Yes -> project glossary + use; No -> do not use); added the 4 terminal states (green/red/orange) from the graphic.
4. Built "The dictionary format" from pages 2-0-3 (four columns), 2-0-4..2-0-8 (Column 1: bold, UPPERCASE=approved, lowercase=not approved, part-of-speech constraint, the 8 parts of speech n/v/adj/adv/pron/art/prep/conj, forms of nouns/adjectives/adverbs), 2-0-9..2-0-10 (Column 2: definitions not in STE, alternatives as suggestions, first alternative usually same part of speech, (TN)/(TV) alternatives, phrase alternatives without part-of-speech markers), 2-0-14 (Column 3: recommendations, aircraft-maintenance domain, change form) and 2-0-15 (Column 4: empty for approved words unless restricted-meaning help, e.g. ABOVE -> MORE THAN).
5. Built "Verbs in STE" from pages 2-0-5..2-0-7: permitted-forms-only rule, transitive/intransitive, Rule 1.12 for technical verbs, and the four types with spec examples — regular (ADAPT, ADAPTS, ADAPTED, ADAPTED), irregular (GIVE, GIVES, GAVE, GIVEN), irregular auxiliary (BE: IS, WAS, also ARE, WERE — no other forms), defective modal (CAN, COULD — no other forms, help "do not use COULD to show possibility"; WILL — no other forms, simple future).
6. Built "Help categories" from pages 2-0-11..2-0-13: 1 recommendations (PUSH + preposition/adverb for direction), 2 restricted meaning (ABOUT = "concerned with"; APPROXIMATELY/AROUND for other meanings), 3 single-context (SWALLOW — safety instructions only), 4 important information (BE; already -> IN PROGRESS / NO OTHER, alternative frequently unnecessary).
7. Reproduced the complete recurring-errors list from pages 2-0-17..2-0-18 as a numbered 39-row Non-STE -> STE table in the source order (acceptable -> PERMITTED through using -> USE/WITH), keeping the (TN) replacements (THE TWO, COVER), the two separate further(adj)/further(adv) entries, and the two construction-level replacements ("any" -> none or different construction; "have to" -> action verb in imperative form); noted the printed 31+8 page split.
8. Reproduced the complete approved-verbs list from page 2-0-19: all 208 verbs grouped alphabetically with per-letter counts (A 12, B 11, C 25, D 16, E 9, F 12, G 5, H 6, I 9, K 3, L 9, M 10, O 5, P 12, R 14, S 24, T 14, U 4, W 8) and explicit "(0): none" rows for J, N, Q, V, X, Y, Z, plus a letter-count check line summing to 208 and notes on multi-word entries (COME ON, GO OFF, MAKE SURE, PUT ON) and the de-interleaved 5-column source table.
9. Verified mechanically with a script: 252 lines (within the ~250-350 target), exactly 39 recurring-error rows, exactly 208 verbs with every per-letter count matching its group, all 26 letters present, all 6 required H2 sections present, and all 7 decision-flow steps (1-6 plus 5a) present.

- Citation: ASD-STE100 Issue 9 Part 2 Introduction, pages 2-0-3..2-0-19
- Citation: page 2-0-3 (General; four columns)
- Citation: pages 2-0-4..2-0-8 (Column 1; eight parts of speech; forms of approved words)
- Citation: pages 2-0-5..2-0-7 (four verb types)
- Citation: pages 2-0-9..2-0-10 (Column 2)
- Citation: pages 2-0-11..2-0-13 (Help, four categories)
- Citation: page 2-0-14 (Column 3)
- Citation: page 2-0-15 (Column 4)
- Citation: page 2-0-16 (How to select words correctly flowchart, from supplement-graphics.md)
- Citation: pages 2-0-17..2-0-18 (List of recurring errors, 39 entries)
- Citation: page 2-0-19 (List of approved verbs, 208 entries)
- Citation: Rule 1.5 (technical nouns, referenced)
- Citation: Rule 1.12 (technical verbs, referenced)
- Citation: Part 1 Section 1 (parts of speech, referenced)
- Citation: Part 1 Section 3 (verbs, referenced)

### Fact-checker 1 (384 claims checked)

1. Read all of word-selection.md, the full spec extract (pages 2-0-3 to 2-0-20 of ASD-STE100 Issue 9 Part 2 Introduction), and supplement-graphics.md. 2. Decision flow vs supplement §1 (page 2-0-16 flowchart, pdf 144): compared all 6 decision nodes, all 12 branches, both intermediate actions, and all 4 terminals word by word — identical, including the parentheticals '(an UPPERCASE headword)', '(as your intended use)', '(the approved meaning says what you want to say)', the Rules 1.5/1.12 reference in step 6, the project-glossary action, the green/red/orange terminal colors, and the verbatim intro quote. Only deviation: an added '(or is a phrase)' gloss in the file's terminal-state recap (minor finding 2). 3. Recurring errors (pages 2-0-17 to 2-0-18): counted rows programmatically in the spec text — 31 on 2-0-17 (acceptable→rotate) + 8 on 2-0-18 (secure→using) = 39, matching the file's claimed 31+8=39; compared all 39 Non-STE→STE mappings row by row including multi-alternative rows (old, over, people, secure, therefore, under, using), prose rows (any, have to), the two (TN) replacements (THE TWO, COVER), and the further(adj)/further(adv) pair — all exact, none dropped, none added. 4. Approved verbs (page 2-0-19): de-interleaved the spec's 5-column table by character offsets (232 cells − 24 letter headers = 208 verbs) and diffed every letter group against the file's lists — all 24 groups identical in content and order; per-letter counts A12 B11 C25 D16 E9 F12 G5 H6 I9 K3 L9 M10 O5 P12 R14 S24 T14 U4 W8 sum to 208; no-verb letters J, N, Q, V, X, Y, Z confirmed (spec shows those headers with no entries, X/Y/Z as a combined 'X, Y, Z' header); also sample-verified 33 verbs across the alphabet by eye (ABSORB, ATTACH, BYPASS, CAN, CANNOT, COME ON, CUT, DE-ENERGIZE, DRY, EXTINGUISH, FEATHER, FREEZE, GO OFF, HOLD, ISOLATE, KEEP, LUBRICATE, MAKE SURE, MUST, OBEY, OVERRIDE, PUT ON, RUB, SAFETY, SWALLOW, TORQUE, TWIST, UNWIND, WILL, WRITE, GET, GIVE, GROUND) and the notes (multi-word entries, CAN vs CANNOT separate, auxiliaries BE/CAN/CANNOT/HAVE/MUST/WILL present, five interleaved columns). 5. Verb types and forms (pages 2-0-5 to 2-0-7): verified the four types (regular, irregular, irregular auxiliary, defective modal), the ADAPT/GIVE/BE/CAN/WILL example entries with meanings and STE sentences, 'No other verb forms' for BE/CAN/WILL, the 'Do not use COULD (v) to show possibility' help, transitive/intransitive note, 'do not use verb forms not in the dictionary', past-participle warning, and the Rule 1.12 and Part 1 Section 3 pointers; the file's '(base, -s, past, past participle)' gloss is consistent with every 4-form example shown. 6. Help categories (pages 2-0-11 to 2-0-13): verified the four categories and their example entries (PUSH with both meanings and the direction help; ABOUT restricted meaning with APPROXIMATELY/AROUND; SWALLOW safety-only; BE and already with IN PROGRESS/NO OTHER and the 'alternative not necessary' help) — all quotes exact; flagged the unsupported '(light-bulb icon)' (minor finding 1). 7. Four-column format and case conventions (pages 2-0-3, 2-0-4, 2-0-9/2-0-10, 2-0-14, 2-0-15): verified the 4 column names, bold typeface, UPPERCASE=approved / lowercase=not-approved rules with ABRASIVE/AID/main/build examples, part-of-speech-in-parentheses rule with Part 1 Section 1 pointer, all 8 parts of speech definitions, noun/adjective/adverb form rules (Rule 1.5, SLOW(SLOWER,SLOWEST), more/most, -ly adverbs, SLOWLY/briskly), Column 2 rules (definitions not STE, BEHIND, addition→ADD, first-alternative same POS, maintain→KEEP/HOLD/MAINTENANCE, uncovered→COVER(TN), simultaneously→AT THE SAME TIME with no POS marker), Column 3 rules (recommendations, one method, aircraft-maintenance domain, change form, A(art), manufacture→MAKE), Column 4 rules (multiple non-approved words, empty-unless-help, ABOVE/MORE THAN example). 8. General facts (page 2-0-3): 875 approved / 1274 not approved, alternatives with examples, TN/TV identifications, no antonyms, Merriam-Webster American English spelling and punctuation, AID and accuracy→PRECISION illustration entries, and all page citations in the file (2-0-3, 2-0-4 to 2-0-8, 2-0-5 to 2-0-7, 2-0-9 to 2-0-10, 2-0-11 to 2-0-13, 2-0-14, 2-0-15, 2-0-16, 2-0-17 to 2-0-18, 2-0-19) — all match the page footers in the extract. Result: both mandatory lists complete and exact (39/39, 208/208), flow exact; only 2 minor unsupported-gloss findings, no critical findings.

**Findings (2):**

- **[minor]** word-selection.md, 'Help categories' intro (line 147)
  - Claim in file: Some entries have a graphic symbol (light-bulb icon) plus text that tells you to use a different approved alternative or a different sentence construction.
  - Spec says: Page 2-0-11 says only: 'For some words, there is a graphic symbol and related text... This graphic symbol and its related text is referred to as "help."' Neither the extracted spec text nor the supplement transcription describes the symbol's appearance; the supplement covers only the page 2-0-16 flowchart, the Rule 1.5 symbols page, and the Rule 3.6 diagram.
  - Proposed fix: Remove '(light-bulb icon)' or mark it as unverified against the extracted sources.
- **[minor]** word-selection.md, 'Terminal states of the flowchart', 4th bullet (line 57)
  - Claim in file: 'Use a different sentence construction.' (orange) — reached from step 5a when the selected alternative has a different part of speech (or is a phrase).
  - Spec says: Supplement §1 transcribes step 5a as branching only on 'Does the alternative have the same part of speech?' — the flowchart has no phrase branch. The phrase case is an inference from page 2-0-10 (phrase alternatives carry no part-of-speech marker), not part of the graphic.
  - Proposed fix: Drop '(or is a phrase)' from the terminal-state description, or label it explicitly as an inference from the Column 2 phrase-alternative rule (page 2-0-10) rather than flowchart content.

### Fixer (applied 1, rejected 1)

1. Read the full target file word-selection.md and located the two claims: 'Help categories' intro (line 147) and the 4th terminal-state bullet under 'Terminal states of the flowchart' (line 57). 2. For Finding 1, grepped spec/r8-dictionary-introduction.txt for 'graphic symbol' and confirmed the page 2-0-11 'Help' passage (lines 367-371) says only 'a graphic symbol and related text' with no appearance description, and lists the four help categories (pages 2-0-11 to 2-0-13). 3. Extended the search to the whole extracted spec (spec/full.txt) for 'light bulb / bulb / icon' and found two genuine spec hits: the Issue 9 'Changes' table entry 'Help (light bulb symbol) added to some rules in part 1 where applicable' (full.txt line 90 = r1-frontmatter-intro.txt line 21) and the Subject-to-rule index entry 'Help (light bulb symbol) — Part 1, all sections. Part 2, applicable entries' (full.txt line 1293 = r1-frontmatter-intro.txt line 380); the third hit (line 16943, 'BULB SENSES') is an unrelated dictionary example. Because the index explicitly extends the light-bulb-symbol help to Part 2 dictionary entries, the '(light-bulb icon)' parenthetical is verified, so I rejected Finding 1 and made no edit to line 147. Also confirmed the digest (digest/r8-dict-intro.md line 101) carries the same wording, consistent with the spec. 4. For Finding 2, read digest/supplement-graphics.md section 1 (the transcription of the 'How to select words correctly' flowchart, Part 2 Introduction page 2-0-16, pdf 144): step 5a branches only on 'Does the alternative have the same part of speech?' (Yes = word-for-word replacement; No = different sentence construction), and the terminal-states line lists the four boxes with no phrase condition. Confirmed the supplement's other sections cover only Rule 1.5 category 9 symbols (page 1-1-7) and the Rule 3.6 Method 1 diagram (page 1-3-7), so no other flowchart source exists. 5. Confirmed via spec/r8-dictionary-introduction.txt (lines 570-580) that pdftotext extracted only the intro sentence of page 2-0-16 — the flowchart itself is graphical — so the supplement transcription is the sole authority for the graphic's content. 6. Confirmed the phrase rule is a Column 2 format rule on page 2-0-10 (spec/r8-dictionary-introduction.txt lines 348-349: a not-approved word's alternative 'can also be a phrase' and such alternatives carry no part-of-speech marker, example simultaneously -> AT THE SAME TIME), i.e., an inference source, not flowchart content — so Finding 2 is correct. 7. Applied Finding 2's primary fix with a single Edit: removed '(or is a phrase)' from the line-57 terminal-state bullet, leaving 'reached from step 5a when the selected alternative has a different part of speech.' The phrase-alternative fact remains correctly documented in the file's 'Column 2' section (page 2-0-10 material, line 103), the step 5a description in the numbered flow (lines 45-47) already matched the supplement, and no other content was touched; format, bolding, and sentence style were preserved.

**Rejected finding:** Finding 1 (line 147, 'light-bulb icon'): REJECTED. Its premise that neither the extracted spec text nor the supplement describes the help symbol's appearance is false. The extracted spec text does describe it in two places: the Issue 9 'Changes' table (spec/r1-frontmatter-intro.txt line 21; spec/full.txt line 90) reads 'Help (light bulb symbol) added to some rules in part 1 where applicable,' and the Subject-to-rule index (spec/r1-frontmatter-intro.txt line 380; spec/full.txt line 1293) reads 'Help (light bulb symbol) — Part 1, all sections. Part 2, applicable entries.' The index entry explicitly ties the light-bulb symbol to Part 2 dictionary-entry help, i.e., the same 'help' defined on page 2-0-11. The parenthetical is therefore verified against the extracted sources and neither removal nor an 'unverified' label is warranted.

## File: dict-approved

### Draft agent

**Result:** 941 total lines: a 6-line header (title, source note citing STE Rules 1.2/1.3, entry-format statement, grep lookup example kept verbatim, honest counts line "Spec count: 875 approved words. This file: 935 entry lines (one line per part of speech/numbered meaning)") plus 935 entry lines, alphabetical in the spec's own printed order. Every entry now matches `WORD (pos) = Meaning [bracketed notes]` — the single spec-sanctioned exception is `FOR EXAMPLE = ...`, which has no part of speech in the spec, and this is stated in the header. 117 irregular-verb/comparative form blocks that sat between the pos and the `=` (comma style, paren style, bracket style) or dangled after the meaning were moved into the leading position of the trailing bracketed notes (e.g. `[APPLIES, APPLIED, APPLIED]`, `[CAN, COULD; no other verb forms; ...]`); 100 lowercase meaning starts from two digest chunks (GET–IMMEDIATELY, SEPARATION–START) were capitalized to spec case; 9 misleading `[no other verb forms]` entries whose forms the digests dropped got them restored from the spec (GET, GO, GO OFF, HAVE, LET, PUT, PUT ON, DRINK, EAT); 2 paren-wrapped notes (DRINK, EAT) became bracketed; SINCE's unbracketed trailing cross-reference was bracketed; 4 lines that jammed two spec-numbered meanings together (INTO, IT, LENGTH, LOOSE) were split one line per meaning (931 -> 935; LOOSE's TENSION note stays with meaning 2 per spec layout). No meanings were invented, dropped, or reworded. Zero exact duplicates existed (adjacent or anywhere, verified case-insensitively), so none were removed; all apparent alphabetical inversions (AFT/AFTER/AFT OF, DEDICATED/DE-ENERGIZE, GO OFF/GOOD, IN PROGRESS/INBOARD, OUT OF/OUTBOARD) match the spec's printed order, so ordering was left untouched.

**Steps:**

1. Read all 937 lines of the merged file (6 header + 931 entries), then ran mechanical drift checks: every entry already had " = " (no arrows), no list markers, no lowercase headwords, no adjacent or global duplicate lines (headword+pos+meaning, case-insensitive) — so step 3's dedup removed nothing.
2. Checked alphabetical order two ways (letter-by-letter vs word-by-word); the disagreements (AFT OF after AFTER; GO OFF before GOOD; IN PROGRESS before INBOARD; OUT OF before OUTBOARD; DEDICATED before DE-ENERGIZE) were each verified against the digests — d01-dict-p149-164.txt:463-475, d08-dict-p261-276.txt:180-192, d09-dict-p277-292.txt:82-146, d11-dict-p309-324.txt:775-781 with d12-dict-p325-340.txt:5, d04-dict-p197-212.txt:659-666 — and all match the spec's printed order, so no "inversions" were repaired.
3. Verified the spec's entry conventions from the digests: keyword column prints verb forms comma-style (APPLY (v), APPLIES, APPLIED, APPLIED — d02-dict-p165-180.txt:13-16) and adjective comparatives in parens (BAD (adj) (WORSE, WORST) — d02:431-434; GOOD, GLOSSY — d08:150-152, 190-192), and all approved meanings start capitalized (FALL d07-dict-p245-260.txt:109-112; SEPARATION, START d15-dict-p373-388.txt:28-30, 776-784).
4. Ran a deterministic script that moved all 117 form blocks (67 comma-style, 32 bracket-before-=, 14 paren-comparative, 4 trailing-paren like DIM/DIRTY/DRY/EASY) into the front of the trailing bracketed notes so every line matches `WORD (pos) = meaning [notes]`, and capitalized 100 lowercase meaning starts from the two drift chunks (GET–IMMEDIATELY, SEPARATION–START); reviewed the full diff before installing.
5. Verified spot entries before touching or deliberately not touching them: FOR EXAMPLE has no pos in the spec (d07:484-486 — kept without pos, noted in header); MATT (or MATTE) (adj) is the spec's exact headword (d10-dict-p293-308.txt:720-721); THRU (prep) confirmed (d16-dict-p389-404.txt:742-743); BY (prep) = Mathematical function word confirmed verbatim (d03-dict-p181-196.txt:62-63); IN PROGRESS (adv) confirmed (d09:113-114); WILL and LESS genuinely list no forms (d18-dict-p421-434.txt:411-415; d10:213-217) so their notes stand alone.
6. Hand-applied 14 spec-verified Edits: restored dropped forms into the 9 entries whose "[no other verb forms]" note had lost its referent — GET/GETS,GOT (d08:59-66), GO/GOES,WENT (d08:180-184), GO OFF/GOES OFF,WENT OFF (d08:185-189), HAVE/HAS,HAD (d08:410-414), LET/LETS,LET (d10:219-223), PUT/PUTS,PUT (d13-dict-p341-356.txt:271-276), PUT ON/PUTS ON,PUT ON (d13:280-285, page id 2-1-P18), DRINK/DRINKS,DRANK (d05-dict-p213-228.txt:561-566), EAT/EATS,ATE (d05:717-722) — converting DRINK/EAT's paren notes to brackets; bracketed SINCE's trailing cross-reference (d15:275-281); split the four lines carrying two spec-numbered meanings into one line each per the header's one-line-per-meaning rule — INTO (d09:592-598), IT (d09:650-657), LENGTH (d10:192-198), LOOSE with the TENSION (TN) note kept on meaning 2 matching spec placement (d10:473-482).
7. Updated the header per the task: kept the grep lookup example and the Rules 1.2/1.3 citation, added the forms-in-brackets convention and the FOR EXAMPLE exception, and stated the honest counts: spec count 875 approved words; this file 935 entry lines.
8. Final validation: 941 total lines = 6 header + 935 entries; strict-format grep shows 0 deviations besides FOR EXAMPLE; 0 lowercase meanings, 0 paren notes, 0 duplicates remain; temp script and staging file removed.

- Citation: STE Rules 1.2, 1.3 (approved words usable only with dictionary meaning/pos — cited in file header)
- Citation: Part 2 dictionary page id 2-1-P18 (d13-dict-p341-356.txt)
- Citation: d01-dict-p149-164.txt:463-475 (AFT/AFTER/AFT OF spec order)
- Citation: d02-dict-p165-180.txt:13-16 (APPLY comma-style forms), 431-434 (BAD paren comparatives)
- Citation: d03-dict-p181-196.txt:62-63 (BY)
- Citation: d04-dict-p197-212.txt:659-669 (DEDICATED/DE-ENERGIZE order, DEEP)
- Citation: d05-dict-p213-228.txt:561-566 (DRINK), 717-722 (EAT)
- Citation: d07-dict-p245-260.txt:109-112 (FALL), 484-488 (FOR EXAMPLE, no pos)
- Citation: d08-dict-p261-276.txt:59-66 (GET), 150-152 (GLOSSY), 180-192 (GO/GO OFF/GOOD), 410-414 (HAVE)
- Citation: d09-dict-p277-292.txt:82-146 (IN/IN FRONT OF/IN PROGRESS/INBOARD), 592-598 (INTO), 650-657 (IT)
- Citation: d10-dict-p293-308.txt:192-198 (LENGTH), 213-217 (LESS), 219-223 (LET), 473-482 (LOOSE + TENSION note), 720-721 (MATT (or MATTE))
- Citation: d11-dict-p309-324.txt:775-781 + d12-dict-p325-340.txt:5-13 (OUT OF/OUTBOARD order)
- Citation: d13-dict-p341-356.txt:271-285 (PUT, PUT ON)
- Citation: d14-dict-p357-372.txt:9-12 (RELEASE trailing-forms chunk baseline)
- Citation: d15-dict-p373-388.txt:28-36 (SEPARATION/SEQUENCE case), 275-281 (SINCE), 776-784 (START)
- Citation: d16-dict-p389-404.txt:742-743 (THRU)
- Citation: d18-dict-p421-434.txt:411-415 (WILL)

### Fact-checker 1 (148 claims checked)

1. Sampled 55 words (89 entry lines) across A-Y and compared each against the UPPERCASE spec table rows: the 14 multi-POS words AFT (adj/adv, d01 rows 463-478 incl. AFT OF), CONTROL (n/v, d04 rows 269-276), CORRECT (adj/v, d04 rows 343-349), DOWN (adj/adv/prep, d05 rows 501-507), DRY (adj/v, d05 rows 604-612), EACH (adj/pron + EACH OTHER, d05 rows 662-673), GROUND (n/v, d08 rows 271-277), LIGHT (adj only - spec d10 rows 267-282 shows light (v) is UNAPPROVED -> COME ON, so the file's single LIGHT (adj) line is correct), MIDDLE (adj/n, d11 rows 82-86), SUPPLY (n + v meanings 1/2, d16 rows 315-325), THAT (conj/pron incl. section 9 GR-1 note, d16 rows 606-612), THERE (adv/pron, d16 rows 634-640), THIS (adj/pron incl. GR-4 note, d16 rows 698-705), THROUGH (adv/prep, d16 rows 723-730); plus ABOUT, BLEED, DATA (1/2 + singular note), EMERGENCY, FOR EXAMPLE (no POS, matching spec d07 row 484), GET (rule 9.2 note verified genuine at d08 rows 59-73), GIVE, GO, GO OFF, GOOD, HOLD, IF, JOINT, KEEP, KILL, KNOW, LOCK, MAKE SURE, NECESSARY, OBEY, PERSON, QUALITY (d13 row 300), QUANTITY (+NUMBER note, d13 rows 303-311), QUICKLY, RECOMMEND, SAFETY (n + v 1/2, d14 rows 482-490), TELL, UNTIL (conj/prep, d17 rows 598-605), VERTICAL (d18 row 71), YES/YOU/YOUR (d18 rows 547-563), and FIRE (n/v, d07 rows 291-311). Case, part of speech, and meaning matched the spec verbatim in every sampled entry. 2. Letter coverage: first-letter scan of all 935 entry lines shows one contiguous alphabetical run A,B,...,W,Y; X and Z are legitimately absent - grep over d01-d18 finds zero approved (or any) X rows and only unapproved 'zero (v) -> ZERO (TN)' under Z (d18, page 2-1-Y1); the dictionary's last approved word is YOUR (adj), matching the file's last line. 3. Junction checks: GENERAL~generally (pdf 260/261): GENERAL (adj) ends d07, target G block GAS->GENERAL->GET->...->GROUP = 13 lines exactly matching the 13 approved G rows, GENERAL appears once; MAXIMUM~may (308/309): MAXIMUM (adj)+(n) once each then MEASURE, MECHANICAL, MECHANICALLY, MECHANISM, MEDICAL, MELT in spec order ('may' correctly excluded as unapproved); RELATION~relay (356/357): RELATION (n) once with exact meaning, then RELEASE, REMAINING, REMOVAL, REMOVE ('relay' excluded); station~stationary (388/389): both spec rows are lowercase/unapproved and correctly absent, START(n/v)->STAY->STEP->STOP->STOW->STRAIGHT->STRIP continuous; till~TILT (404/405): 'till (conj)' (d16 row 776, unapproved -> UNTIL) absent, TIGHT->TIGHTEN->TIGHTLY->TILT->TIME continuous with TILT meaning verbatim; utilize~valid (420/421): 'utilize' (d17 row 741) and 'valid'/'validity' (d18 rows 1-25) all unapproved and absent, UPSTREAM/UPSTREAM OF->USE->USUAL->USUALLY->VALUE->VERTICAL...VOLUME continuous. No duplicates or drops at any junction. 4. Counts: file has 941 lines = 5 header lines + 1 blank + 935 non-blank entry lines, matching the header's '935 entry lines'; the header's 'Spec count: 875 approved words' correctly cites the spec's own figure (r8-dictionary-introduction.txt rows 7-8: 'the words that are approved in STE (875 approved words)'). My own tallies: 879 distinct headword+POS pairs, 806 distinct headwords - the spec text never states its counting basis, and a full two-way set-diff (all line-start UPPERCASE spec rows vs all file pairs) found ZERO spec entries missing from the file (the only 6 apparent misses - ABLE, NETIC, NETICALLY, TORILY, TORY, WISE - are pdftotext hyphenation tails of INTERCHANGE-ABLE, ELECTROMAG-NETIC(ALLY), (UN)SATISFAC-TORILY/TORY, COUNTERCLOCK-WISE, all present in full in the file) and zero invented entries (spot-checked oddities IN PROGRESS (adv) 'Occurring', UNDEMANDED, SYNCHRONIZED, MATT (or MATTE) all exist verbatim in spec). 5. Case spot-check: 10 pseudo-random lines (every 97th entry: ACCURATELY, BURN, DEFUEL, FIRE, INPUT, MECHANISM, PARTICLE, SAFE, SUBSEQUENTLY, UNWANTED) - all 10 are UPPERCASE approved rows in the spec with matching meanings; additionally confirmed 13 lowercase (unapproved) spec words are NOT in the file (caution, warn, light(v), generally, may, relay, station, stationary, till, utilize, valid, validity, zero), including CAUTION/WARNING which Issue 9 lists only as unapproved verbs -> TELL (v) (d03 row 256, d17 row 221). Total: 89 meaning/POS/case-verified entry lines + 42 junction-order entries + 13 unapproved-absence checks + 2 count checks + 1 letter-coverage check + 1 full two-way reconciliation = 148 verified claims. No critical issues; 3 minor fidelity issues reported.

**Findings (3):**

- **[minor]** lines 240-241 (DO), 347 (GIVE), 357 (HANG), 361 (HEAR), 368 (HIT), 369-370 (HOLD), 434 (KEEP), 436 (KNOW), 486 (MAKE), 487 (MAKE SURE), 657 (READ), 821 (TELL), 841 (THINK)
  - Claim in file: Entries for 13 irregular verbs carry no bracketed verb forms, although the file's own header (line 3) states 'Irregular verb forms ... appear at the start of the bracketed notes' and other irregular verbs (GO, GET, BE, BEND, BLEED, BLOW) do include them.
  - Spec says: The spec lists the irregular forms under each of these entries: DOES, DID, DONE (DO, d05); GIVES, GAVE, GIVEN (d08); HANGS, HUNG, HUNG (d08); HEARS, HEARD, HEARD (d08); HITS, HIT, HIT (d08); HOLDS, HELD, HELD (d08 row 546); KEEPS, KEPT, KEPT (d10 row 5); KNOWS, KNEW, KNOWN (d10 row 28); MAKES, MADE, MADE (d10); MAKES SURE, MADE SURE, MADE SURE (d10 row 594); READS, READ, READ (d13); TELLS, TOLD, TOLD (d16); THINKS, THOUGHT, THOUGHT (d16).
  - Proposed fix: Append the spec's irregular forms to each line, e.g. 'HOLD (v) = 1. To continue to have in the hand or grip [HOLDS, HELD, HELD]', 'KEEP (v) = To continue to have or hold [KEEPS, KEPT, KEPT]', 'MAKE SURE (v) = ... [MAKES SURE, MADE SURE, MADE SURE]', etc.
- **[minor]** line 457 (LIGHT)
  - Claim in file: LIGHT (adj) omits the comparative forms note; meaning, POS, and the THIN/SMALL alternatives note are otherwise correct.
  - Spec says: d10-dict-p293-308.txt rows 267-269: 'LIGHT (adj) (LIGHTER, LIGHTEST) That has a small mass, weight, or force'. Sibling entries in the file do keep comparatives (GOOD [BETTER, BEST] line 351, DRY [DRIER, DRIEST], SAFE [SAFER, SAFEST], GLOSSY [GLOSSIER, GLOSSIEST]).
  - Proposed fix: Change to: LIGHT (adj) = That has a small mass, weight, or force [LIGHTER, LIGHTEST; for other meanings use THIN (adj) or SMALL (adj)]
- **[minor]** lines 803-805 (SUPPLY), 813-814 (SYNCHRONIZED)
  - Claim in file: The '1.'/'2.' numbered-meaning markers were dropped for SUPPLY (v) and SYNCHRONIZED (adj), while other multi-meaning entries (ABSORB lines 11-12, DATA, DO, HOLD, TIME, VOLUME lines 912-913) keep them - inconsistent rendering of the spec's numbered meanings.
  - Spec says: d16-dict-p389-404.txt rows 318-325 numbers SUPPLY (v) meanings '1. To give something that is necessary / 2. To make something available for use'; d15 rows 461-465 numbers SYNCHRONIZED (adj) '1. That operates at the same time / 2. That agrees with'.
  - Proposed fix: Prefix the meanings with '1.' and '2.' to match the convention used by the file's other numbered entries, e.g. 'SUPPLY (v) = 1. To give something that is necessary'.

### Fixer (applied 3, rejected 0)

1. Read the full target file and confirmed every cited line matched the findings' quoted content (DO 240-241, GIVE 347, HANG 357, HEAR 361, HIT 368, HOLD 369-370, KEEP 434, KNOW 436, MAKE 486, MAKE SURE 487, LIGHT 457, READ 657, SUPPLY 803-805, SYNCHRONIZED 813-814, TELL 821, THINK 841), and that the header (line 3) requires irregular/comparative forms at the start of bracketed notes. 2. Finding 1 re-verified against primary spec: d05-dict-p213-228.txt rows 466-469 (DO, DOES, DID, DONE); d08-dict-p261-276.txt rows 126-129 (GIVES, GAVE, GIVEN), 343-346 (HANGS, HUNG, HUNG), 442-445 (HEARS, HEARD, HEARD), 535-538 (HITS, HIT, HIT), 546-549 (HOLDS, HELD, HELD); d10-dict-p293-308.txt rows 5-8 (KEEPS, KEPT, KEPT), 28-31 (KNOWS, KNEW, KNOWN), 583-586 (MAKES, MADE, MADE), 594-597 (MAKES SURE, MADE SURE, MADE SURE); d13-dict-p341-356.txt rows 431-434 (READS, READ, READ); d16-dict-p389-404.txt rows 555-558 (TELLS, TOLD, TOLD) and 694-697 (THINKS, THOUGHT, THOUGHT) — all 13 confirmed. 3. Finding 2 re-verified: d10-dict-p293-308.txt rows 267-269 show LIGHT (adj) (LIGHTER, LIGHTEST) with the THIN/SMALL alternatives — confirmed. 4. Finding 3 re-verified: d16-dict-p389-404.txt rows 318-325 number SUPPLY (v) meanings 1./2., and rows 461-467 number SYNCHRONIZED (adj) meanings 1./2. (the finding's citation said d15, but the entry is in d16 — a citation typo only; the substantive claim is correct, so not grounds for rejection). 5. Secondary check of digest/ (d08-dict.md, d10-dict.md, d16-dict.md) showed the digests carry the same omissions as the build file, so the primary spec .txt files were treated as authoritative. 6. Checked the file's own bracket conventions before editing: for multi-meaning verbs the forms sit on the first meaning line only (APPLY 57, BOND 107, RETRACT 687, SAFETY 701, SHOW 733, SOAK 751, WEIGH 922), matching the finding's HOLD example, so DO and HOLD got forms on meaning 1 only; for LIGHT, forms go first in the bracket followed by '; for other meanings ...' per BAD 84 and CATCH 131. 7. Applied 18 line edits covering the 3 findings, then re-grepped all repaired lines and confirmed the entry-line count is unchanged (941 lines), no uncovered content touched.

## File: dict-unapproved

### Draft agent

**Result:** Normalized ASD-STE100 Issue 9 not-approved word list: 1325 total lines = 5 header lines + 1 blank + 1319 entry lines (one line per part of speech; no duplicates, none dropped, none added). All 1319 entries match `word (pos) -> ALTERNATIVE 1; ALTERNATIVE 2 [optional bracketed hints]` (sole intentional exception: `such as -> FOR EXAMPLE`, which has no part of speech in the spec itself). 10 entry lines were corrected against the spec: 6 bracket-style fixes to `[TN]` inside instrument parentheticals (bolt — also lowercase-to-uppercase case fix, drift, uncap, unclip, wire, wire-lock) and 4 ellipsis fixes to the spec's `…` form (hand-tight, hand-tighten, soften `MAKE…SOFT`, spin `TURN … QUICKLY`). Header line 5 updated to: spec count 1274 not-approved words, file has 1319 entry lines; grep lookup example kept unchanged.

**Steps:**

1. Read all 1325 lines of the merged file and ran mechanical drift scans: no unicode arrows, no `=` in place of `->`, no missing arrows, no UPPERCASE headwords, no double/trailing spaces, no adjacent or global duplicate lines, no duplicate headword+pos pairs anywhere.
2. Explicitly audited the d08/d09 chunk junction around "impact" (pdf pages 276/277): d08 (p261-276) ends with `impact (n) -> HIT (v)` and d09 (p277-292) opens with `impact (v) -> HIT (v); EFFECT (n)` followed by impair, implement, implementation, impossible — the file carries all of these exactly once, so no duplicated or dropped row at the junction.
3. Verified every suspicious-looking line against the spec digests before touching it; confirmed as spec-faithful and left unchanged: `already -> NO OTHER` (d01 p149-164), `common -> SAME (adv)` (d03 p181-196), `conflict`, `decrement -> INCREMENT (n)`, `degrease -> GREASE (TN)` (d04 p197-212), `disappear`, `droop` (d05 p213-228), `back and forth`, `bind` (d02 p165-180), `so (that) -> UNTIL (prep)` (d15 p373-388), `subsequent to -> AFTER (conj)`, `such as` with no pos (d16 p389-404), `file (v) ... (WITH A FILE (TN)` (d07 p245-260, the spec extraction itself uses `(TN` here), `within -> IN ... OR LESS` ASCII dots and `yet (conj)` before `yet (adv)` (d18 p421-434), `otherwise -> IF …NOT` spacing (d11 p309-324), `pump` second alternative (d12 p325-340), plus hint-only entries rework, have to, re- (prefix), situation, undertake, perform, carry out, opportunity, danger.
4. Checked the 4 alphabetical-order flags against the spec's own ordering and repaired none, because the file already follows the spec: consist of before consistent (d04 p197-212, lines 169/171), prior to before priority (d12 p325-340, 736/740), rear of before rearmost (d13 p341-356, 475/483), assure before as to before asymmetric (d02 p165-180, 224/227/232), closely before close to (d03 p181-196, 548/551), free from before free (v) (d07 p245-260, 600/613).
5. Fixed 6 bracket/case drift lines to the spec's `[TN]` instrument-parenthetical form, each verified by grep in the digests: bolt (d02 p165-180, also `with a ... or` uppercased to `WITH A ... OR`), drift (d05 p213-228), uncap and unclip (d17 p405-420), wire and wire-lock (d18 p421-434); cross-checked against the already-correct `[TN]` entries clamp (d03), placard/plug/probe (d12), pump (d12), stamp/strap/suck (d15-d16), safety-clip/safety-wire (d14 p357-372), recleat (d13).
6. Fixed 4 ellipsis drift lines to the spec's exact `…` rendering: hand-tight and hand-tighten `TIGHTEN … WITH YOUR HAND` (d08 p261-276), soften `MAKE…SOFT` and spin `TURN … QUICKLY` (d15 p373-388); left the 12 existing `…` lines (alternately d01, finger-tighten d07, fail/failure d07, otherwise d11, etc.) because they already match the spec.
7. Updated header line 5 to state honestly: spec count 1274 words that are not approved, this file has 1319 entry lines (one line per part of speech); kept the `grep -i "^accomplish "` lookup example on line 4 unchanged.
8. Final counts verified: 1325 total lines, 1319 entry lines, 1319 lines containing ` -> `, zero duplicates — entry count unchanged from input since no rows needed removal or restoration.

- Citation: d01-dict-p149-164 (already, alternately)
- Citation: d02-dict-p165-180 (bolt, bind, back and forth, assure/as to/asymmetric order)
- Citation: d03-dict-p181-196 (clamp, common, closely/close to order)
- Citation: d04-dict-p197-212 (conflict, decrement, degrease, consist of/consistent order)
- Citation: d05-dict-p213-228 (disappear, droop, drift)
- Citation: d07-dict-p245-260 (finger-tighten, file (v), free from/free order)
- Citation: d08-dict-p261-276 (hand-tight, hand-tighten, impact (n) at pdf 276)
- Citation: d09-dict-p277-292 (impact (v) at pdf 277, impair, implement, implementation, impossible)
- Citation: d11-dict-p309-324 (otherwise IF …NOT)
- Citation: d12-dict-p325-340 (pump, placard, plug, probe, prior to/priority order)
- Citation: d13-dict-p341-356 (rear of/rearmost order, recleat)
- Citation: d14-dict-p357-372 (safety-clip, safety-wire)
- Citation: d15-dict-p373-388 (soften, spin, so (that), strap)
- Citation: d16-dict-p389-404 (such as, subsequent to, suck)
- Citation: d17-dict-p405-420 (uncap, unclip)
- Citation: d18-dict-p421-434 (wire, wire-lock, within, yet)

### Fact-checker 1 (108 claims checked)

1. Verified all 9 required multi-POS words letter-perfect against the spec ALTERNATIVES column: substitute adj/n/v (d16-dict-p389-404.txt:227-241), detail n/v (d05-dict-p213-228.txt:82-95), earth n/v and ease n/v (d05:682-701), support n/v with SUPPORT (TN) marker preserved in both orders (d16:326-342), suspect adj/v (d16:372-380), whichever adj/pron (d18-dict-p421-434.txt:355-361), yet conj/adv (d18:550-557), impact n/v (d08-dict-p261-276.txt:755 / d09-dict-p277-292.txt:5-12). All headwords lowercase in spec per the dictionary-intro case rule (r8-dictionary-introduction.txt:68 'a word in lowercase letters ... is not approved').
2. Sampled ~99 further entries covering every letter, including all j/k/q/y/z entries in full (jack...knock, qualified...quote, yet/yourself, zero d18), multi-alternative rows (excess adj 4 alternatives d06:513-526; elapse 3 alts; locate 3; maintain 3; notify 3 for letter n; impair 3 incl. multi-word alts OPERATE INCORRECTLY / CAUSE DAMAGE), TN-marker rows (key -> KEY (TN), kink -> KINK (TN), zero -> ZERO (TN), alternately -> IN ONE (TN) ... with ellipsis preserved), bracket-hint rows (volatile d18, affect d01:53-area, several d15, re- (prefix), any (adj)), and special headwords case (in case of) (conj) built from the two-line spec form d03-dict-p181-196.txt:217-218 and such as (no POS, d16:266). All matched.
3. A-Z contiguity: first-letter sequence over the 1319 entry lines is a,b,...,w,y,z with each letter appearing in exactly one contiguous block (uniq check); no x-headwords exist anywhere in d01-d18 (grep). Apparent sort anomalies (as to after assure; act before action) mirror the spec's own order (d02:224-227, d01:278-281).
4. Global reconciliation (stronger than sampling): parsed every lowercase column-1 headword row from d01-d18 in file order (1319 rows after joining two-line headwords) and diffed against the built file's 1319 headword+POS strings with difflib — identical sequence, zero insertions/deletions. The only diff ops were my parser's artifacts, each then manually verified: two-line headwords chance (by chance) (n) d03:307-313, least (at least) (adv), long (as long as)/(no longer), process (in the process of) (prep), provided/providing (that) (conj); and single-space squeezed rows contaminated, exceptionally, inadvertently, progressively, simultaneous — alternatives all correct in the build.
5. d08/d09 impact junction: d08 page 2-1-I2 ends ...immerse (741) -> immobilize (748) -> impact (n) HIT (v) (755); d09 opens impact (v) HIT (v); EFFECT (n) (lines 5-12) -> impair (13) -> implement (25). Built lines 618-626 reproduce exactly this sequence; grep -c '^impact' = 2 (no duplicate, no drop).
6. Header counts: file has 1325 lines = 6 header/blank + 1319 entry lines, matching the header's '1319 entry lines'; 'Spec count: 1274' matches the spec's own statement at r8-dictionary-introduction.txt:10. Measured 1257 distinct headwords / 57 multi-POS words (62 surplus lines) — see minor finding 2 on the wording.
7. 10-line random spot check (seeded shuf): volatile (d18), distribute (d05), particular (d12), separable (d14), rapid (d13), absorption (d01), past (d12), affect (d01), agitate (d01), absent (d01) — all headwords lowercase in spec, all alternatives byte-correct UPPERCASE with POS tags. Global case scan: after stripping (...) and [...] groups, zero lowercase characters remain in any alternatives field of all 1319 lines; headword lowercase-ness is guaranteed globally because the step-4 sequence match only used spec rows starting with a lowercase letter.
8. Approved-word exclusion sanity: THUS (adv) is uppercase/approved in spec (d16:744) and correctly absent from this file; gain/excess families (lowercase, d07:711-719, d06:513-556) are correctly present. Words absent from both build and spec word list (necessitate, promptly, readily, duplicate, nature, in excess of, gain access) confirmed absent from all d-files, so nothing was dropped.
Checked = 108 entries fully verified column-by-column (headword case, POS, complete ALTERNATIVES incl. TN/TV markers and hints); additionally all 1319 entries structurally verified (headword text, case, POS, order) via the full-sequence diff. No critical findings: no wrong alternative, no case error, no POS error, no missing range anywhere in the samples or the global checks.

**Findings (2):**

- **[minor]** line 827: perform (v) -> DO (v); MEASURE (v) [use a more accurate verb]
  - Claim in file: The bracket hint reads as if it qualifies MEASURE (v) (no separator before it)
  - Spec says: d12-dict-p325-340.txt:298-301 attaches the note 'You can use a more accurate verb.' to the DO (v) alternative, before the MEASURE (v) row. The same spec pattern is rendered differently elsewhere in the build: line 1241 undertake correctly gives 'DO (v) [use a more accurate verb]; START (v)' (spec d17:408), and line 180 carry out puts it as a trailing standalone item '; [use a more accurate command verb]' (spec d03:206).
  - Proposed fix: Change to `perform (v) -> DO (v) [use a more accurate verb]; MEASURE (v)` (undertake style); optionally normalize carry out the same way. Alternatives themselves (DO, MEASURE) are correct, so not critical.
- **[minor]** line 5 (header): 'Spec count: 1274 words that are not approved. This file has 1319 entry lines (one line per part of speech...)'
  - Claim in file: The juxtaposition implies 1274 words expand to 1319 lines purely via multi-POS splitting
  - Spec says: r8-dictionary-introduction.txt:10 does state '(1274 words)', and the file really has 1319 entry lines. But the actual table has 1257 distinct headword spellings; 57 multi-POS headwords add 62 lines (1257+62=1319, incl. either/inside/substitute x3, outside x4). 1274 reconciles with neither 1257 nor 1319 — a discrepancy internal to the spec (intro claim vs its own table), faithfully mirrored by the build.
  - Proposed fix: Reword, e.g.: 'The spec introduction states 1274 words that are not approved; this file has 1319 entry lines covering 1257 distinct headwords (one line per part of speech).' No entries are missing — the full spec row sequence reconciles 1:1.

### Fixer (applied 2, rejected 0)

1. Re-verified finding 1 against spec d12-dict-p325-340.txt:298-304 (dictionary page 2-1-P3): the note "You can use a more accurate verb." sits under the DO (v) alternative row, before the MEASURE (v) row, so the build's line 827 rendering `DO (v); MEASURE (v) [use a more accurate verb]` misattributed the hint to MEASURE. 2. Cross-checked the two comparison entries: undertake at d17-dict-p405-420.txt:408-416 (note under DO (v), before START (v); build line 1241 renders it correctly as `DO (v) [use a more accurate verb]; START (v)`) and carry out at d03-dict-p181-196.txt:206-212 (note "use a more accurate command verb" under DO (v), before MEASURE (v); build line 180 renders it as a trailing standalone `; [...]` item). 3. Also checked accomplish at d01-dict-p149-164.txt:216-224 — identical spec layout (note under DO (v)), rendered trailing-standalone at build line 26 — establishing that lines 26 and 180 form a consistent pair not covered by a mandatory fix. 4. Applied finding 1's core fix with Edit: line 827 is now `perform (v) -> DO (v) [use a more accurate verb]; MEASURE (v)` (undertake style). Declined the finding's optional rider to normalize carry out (line 180): its semicolon-separated trailing hint does not misattribute anything, and normalizing it while its exact twin accomplish (line 26, no finding covers it) stays trailing would create a new inconsistency; left both untouched. 5. Re-verified finding 2: r8-dictionary-introduction.txt:10 (dictionary page 2-0-3, General) states "(1274 words)" for the not-approved selection. 6. Independently recounted the build file with a script: 1319 entry lines (1320 " -> " matches minus the header's format-description line), 1257 distinct headword spellings (stripping only the final part-of-speech parenthetical, so "case (in case of)" stays distinct from "case"), 57 multi-POS headwords contributing 62 extra lines (1257+62=1319), with either/inside/substitute x3 and outside x4 — every number in the finding reproduced exactly; 1274 reconciles with neither 1257 nor 1319, so the discrepancy is internal to the spec (intro claim vs its own table). Digest map.md sections 3 (per-file count table, spec-delta 1319 vs 1274) corroborate. 7. Applied finding 2 with Edit: header line 5 now attributes 1274 to the spec introduction and states the file's real figures — "The spec introduction states 1274 words that are not approved. This file has 1319 entry lines that cover 1257 distinct headwords (one line per part of speech)." (10- and 17-word sentences, within STE's 25-word descriptive limit; dropped the finding's suggested "No entries are missing" clause because I did not independently re-run the full 1:1 row reconciliation). 8. Post-edit verification: lines 5 and 827 read as intended, entry-line count still 1319, total file 1325 lines, no other content changed.

## Whole-skill audit

### Audit round 1

{
  "explanation": "1. Structure: read SKILL.md; frontmatter is valid YAML with exactly name (\"asd-ste100\") + description; ls confirmed all four references/ paths resolve (writing-rules.md, word-selection.md, dictionary-approved.md, dictionary-unapproved.md).\n2. Rule citations: checked every rule number SKILL.md cites (core-limits rows 5.1, 6.3, 5.5, 6.6, 2.1, 3.2, 3.6, 5.2, 8.1, 8.4, 8.5, 8.6, 8.7, 4.5; safety section 7.1-7.3 and 5.5; checklist 1.1-1.3, 3.2, 3.6, 5.1-5.3, 5.5, 6.3, 6.5, 6.6, 2.1, 8.1, 8.4 thru 8.7, 7.1 thru 7.3, 4.5) against writing-rules.md and map.md section 2 \u2014 every rule exists with the same limit and meaning (20 words in procedures incl. safety instructions per 5.1, 25 in descriptions per 6.3 and notes per 5.5, 6 sentences per 6.6, 3-word multi-word nouns per 2.1, six verb forms per 3.2, passive only if agent unknown per 3.6, warning-over-caution per 7.1).\n3. Rule completeness: counted 53 \"### Rule\" headings in writing-rules.md with the exact numbers 1.1-1.14, 2.1-2.2, 3.1-3.7, 4.1-4.5, 5.1-5.5, 6.1-6.6, 7.1-7.3, 8.1-8.7, 9.1-9.4 (14+2+7+5+5+6+3+7+4 = 53, matching map.md), plus GR-1 through GR-8 all present.\n4. Word-selection flow: SKILL.md step 2 delegates to word-selection.md; its 6-step flow (with 5a and the four terminal states) matches the page 2-0-16 flowchart transcription in supplement-graphics.md word-for-word; both grep lookup examples in SKILL.md return hits (ACCESS (n), accomplish (v)).\n5. Counts: 53 rules, 875 approved / 1274 not-approved (spec claims per map.md and r8 page 2-0-3), 39 recurring errors (pages 2-0-17..18), and 208 verbs (page 2-0-19) are quoted identically in SKILL.md, writing-rules.md, word-selection.md, and both dictionary headers; counted 39 numbered error rows (1-39, incl. the 31+8 page split note) and 208 verbs (letter counts A12..W8; J, N, Q, V, X, Y, Z none) and diffed both lists against r8-dict-intro.md \u2014 row-for-row and set-identical.\n6. Dictionary completeness: extracted all entries from digests d01-d18 and diffed against both build dictionaries after normalization \u2014 headword+part-of-speech sets are identical on both sides (935 approved lines per the header's one-line-per-meaning convention, the +4 over the digests being the meaning-splits of INTO, IT, LENGTH, LOOSE; 1319 unapproved lines / 1257 distinct headwords exactly as the header states and as map.md's reader sum records); the d08/d09 junction entries impact (n) and impact (v) (pages 2-1-I2/I3) each appear exactly once; coverage runs A (art) to YOUR and abaft to zero (v) \u2014 the source itself has no approved X/Z headwords and no unapproved x headwords per d17/d18, so A-Z coverage matches the spec; restricted-meaning and comparative/irregular-form notes (ABOUT, AROUND, BAD, BE, DRINK, EAT, SEE, SINCE, WEAR, etc.) survived in brackets.\n7. Rule-body spot checks against the r-digests: Rule 1.5's 22 category titles incl. cats 21-22 and the colors note (r2), Rule 1.12's 4 categories with subgroups a-f (r2), 2.1 runway-light example and 3.5 approved -ing words (r3), 3.6's four passive-to-active methods (r3 + supplement diagram, page 1-3-7), 4.3 vertical-list rules incl. the no-comma/no-semicolon item and 4.5 O-rings/36L7 article semantics (r4), 5.1's 25-to-10+16-word split, 6.3's 31-to-15+16 smartphone split, and 6.6's six fuel-manifold paragraphs of 1, 2, 5, 4, 4, 2 sentences (r5), Section 7 definitions/ISO-ANSI refs, 7.1 oxygen-tubes example, 8.2's five hyphen categories, and the 8.4-8.7 word-count examples (r6), 9.1-9.4 and GR-1..GR-6/GR-8 texts (r7) \u2014 all faithful.\n8. Appendices: Appendix A matches r1's subject-to-rule index (SRI-1..4) row-for-row where sampled, including the source's own \"Simple future/past/present tense -> 3.3\" rows (a spec quirk, reproduced verbatim, so not a build error) and the omitted inclusive-language row; Appendix B matches r1's Highlights (2.3 -> 4.5 move, 1.5 cats 21/22, 1.12 restructure, 8.6.2 revised word count and 8.6.7 proper nouns, GR-7/GR-8 new).\n9. Exclusion: grep -i inclusive over the whole build returns only the two GR-7 exclusion notes (SKILL.md:82, writing-rules.md:324); no gender/inclusive-language content anywhere, consistent with map.md section 5.\n10. STE spot-check of SKILL.md: all steps and checklist items are imperative; the only semicolon character is the \"(;)\" mention inside the 8.1 limit row, not prose usage; every prose sentence is 25 words or fewer (longest is 24) \u2014 the one overlong string is a table cell (finding 3).",
  "residual": [
    {
      "file": "/tmp/claude-1000/-home-cfcosta-Notes/f3cf6639-a28c-403a-997d-fc0414dcd4e4/scratchpad/build/asd-ste100/references/dictionary-unapproved.md",
      "location": "22 lines: 152, 206, 211, 373, 836, 839, 840, 876, 902, 932, 1026, 1027, 1034, 1036, 1037, 1131, 1149, 1165, 1231, 1232, 1312, 1313",
      "issue": "Technical-noun markers written as \"[TN]\" (e.g. \"bolt (v) -> ATTACH (v) (WITH A BOLT [TN] OR BOLTS [TN])\") while the file's own legend defines \"(TN)\" as the marker and says square brackets mean rewording hints; 117 other lines use the (TN)/(TV) notation. Carried over from mixed reader notation (d02 used (TN), d03/d12/d14 used [TN]); content is otherwise correct.",
      "severity": "minor",
      "fix": "Replace [TN] with (TN) on these 22 lines so the notation matches the header legend and the majority of entries."
    },
    {
      "file": "/tmp/claude-1000/-home-cfcosta-Notes/f3cf6639-a28c-403a-997d-fc0414dcd4e4/scratchpad/build/asd-ste100/references/dictionary-unapproved.md",
      "location": "lines 583-584, 1100, 1116",
      "issue": "Inconsistent ellipsis typography in phrase alternatives: spaced ellipsis \"TIGHTEN \u2026 WITH YOUR HAND\" / \"TURN \u2026 QUICKLY\" vs unspaced \"MAKE\u2026SOFT\"; the digests use three periods \"...\" throughout.",
      "severity": "minor",
      "fix": "Normalize all phrase-alternative gaps to one style (suggest \"...\" with surrounding spaces, matching the digests and the recurring-errors table's \"DO (v) \u2026 AGAIN\" style in word-selection.md)."
    },
    {
      "file": "/tmp/claude-1000/-home-cfcosta-Notes/f3cf6639-a28c-403a-997d-fc0414dcd4e4/scratchpad/build/asd-ste100/SKILL.md",
      "location": "line 48 (core-limits table, Rule 8.6 row)",
      "issue": "The 8.6 table cell is one 36-word string; as STE text it would exceed the 25-word limit the skill itself teaches. Content is factually correct (it mirrors Rule 8.6's own enumeration) and it is a table cell, not prose, so this is style only.",
      "severity": "minor",
      "fix": "Shorten the cell, e.g. \"Each of these counts as one word: numbers, numbers with units, abbreviations, alphanumeric identifiers, quoted text, titles, headings, placard and label text, and proper nouns\" and let Rule 8.6 in writing-rules.md carry the full wording."
    }
  ],
  "verdict": "clean"
}

## Post-audit manual fixes (orchestrator, after the workflow)

1. Normalized 22 `[TN]`/`[TV]` markers to `(TN)`/`(TV)` in dictionary-unapproved.md (file-legend consistency; the spec prints `[TN]` when nested in parentheses).
2. Normalized 2 unspaced ellipses to the spec's spaced form (`IF … NOT`, `MAKE … SOFT`).
3. Compressed the SKILL.md Rule 8.6 table cell from 36 to 22 words.