STE — Simplified Technical English (ASD-STE100)

STE is a controlled language: ~53 writing rules plus a dictionary of 875 approved words where each word has exactly one meaning per part of speech. Dictionary compliance CANNOT come from memory — words you would swear are approved ("qualified", "ensure", "should") are not. Only the Verified level below claims compliance, and every claim in it comes from the checker script, not from intuition.

STE is defined for English only. The STE text is always written in English first. If the user asked in another language or asks for a translation, you translate FROM the STE English as a second pass — never draft directly in the other language.

The source can be anything: a document, a merge-request description, a procedure — or something just discussed in the conversation that the user wants again in STE. The workflow is the same; for conversational re-explanations the cheap levels (Marked, Clean) are usually the right default to offer first.

Step 1 — Choose the compliance level and output language (before anything else)

Ask both in ONE interaction (AskUserQuestion where the harness has it, a plain question otherwise) — but ask only what the request leaves open:

• Level: if the request names one, use it silently. A request that

semantically asks for compliance ("does this comply with ASD-STE100?", "quiero STE estricto") selects Verified automatically.

• Output language: if the request names a language, asks for a

translation, or was written in a non-English language (which implies English + that language), use that silently. The choice is "English only" or "English + <language>" — never translation-only: the STE English IS the deliverable, and the target language never changes Step 3 (drafting is always English; translation is always the second pass).

The levels:

• Verified STE — full dictionary compliance. Costs the most tokens:

replacements, TN/TV judgment, fix passes, and the manual checklist.

• Marked draft — STE-style draft with every unresolved word tagged so the

reader sees what a verified pass would resolve. Not verified.

• Clean draft — STE-style draft, no marks. The TN/TV annotation work

primes richer domain detail (measured: ~15% more content than Raw, at the price of slightly weaker general vocabulary and ~40% more tokens). Annotations are validated and stripped by --clean. Not verified.

• Raw draft — STE style from model training only. No annotations, no

dictionary, no checker, no marks, zero tool calls. The cheapest and fastest level; also the shortest, least detailed output.

Clean skips Step 2 (the --clean mode reads no dictionary). Raw skips Steps 2 through 5 entirely — write and deliver. If the dictionary cannot be downloaded, only Clean or Raw are possible: say so, and the note must say "not verified (dictionary unavailable)" with no flagged-word count.

Step 2 — Get the official dictionary (cached; Verified and Marked only)

C="${XDG_CACHE_HOME:-$HOME/.cache}/asd-ste100"
if [ ! -s "$C/ste100.txt" ]; then
  mkdir -p "$C"
  curl -sL --fail -o "$C/spec.pdf" "https://www.asd-ste100.org/assets/files/ASD-STE100_ISSUE9.pdf"
  pdftotext -layout "$C/spec.pdf" "$C/ste100.txt"
fi

-layout is required — the checker parses the dictionary columns and validates the cache structurally (Issue 9 marker, dictionary boundary, exactly 875 approved entries). A wrong cache is a controlled error with regeneration instructions, never silent misclassification. pdftotext comes with poppler; the checker requires gawk.

Step 3 — Draft in STE English, annotating TN/TV as you write

Draft from what you already know of the STE rules — do not read references/rules.md first; read it only when you are not sure about a specific rule. Shapes by text type: procedural — max 20 words per sentence, imperative; descriptive — max 25 words, no imperative, max 6 sentences per paragraph; safety instruction — WARNING (injury) or CAUTION (damage), then command or condition, then consequence.

While drafting (all levels except Raw), annotate every word you use as a technical noun or technical verb (rules 1.5 / 1.12) the moment you write it:

• Single word: webhook~tn or with its category webhook~tn19
• Technical verb: reboot~tv2
• Multi-word: [[tn6: pressure relief valve]]

The annotation moves the TN/TV judgment to the moment it is free and feeds the compliance-note inventory. The checker validates syntax, category code, and dictionary relationship; the semantic category membership is your judgment. Categories can be omitted while drafting (~tn) — the checker prints the numbered category list when any are missing, and the Verified label is forbidden while any declared word lacks a category.

Step 4 — Run the checker (one call per pass)

scripts/ste-check.sh --annotated --verified draft.md
scripts/ste-check.sh --annotated --marked draft.md
scripts/ste-check.sh --clean draft.md      # implies --annotated
scripts/ste-check.sh --entry <word>        # full dictionary entry lookup

Always pass --annotated for drafts you wrote (your drafts carry TN/TV annotations); never pass it when checking third-party text.

Exit codes: 0 automated checks clean, 1 findings remain, 2 invocation/cache failure. One run classifies every word (NOT APPROVED with extracted alternatives per part of speech, NO ENTRY, DECLARED TN/TV inventory, CATEGORY REQUIRED, REVIEW, CONTRACTIONS, IDENTIFIER REVIEW) and runs the mechanical checks (semicolons, sentence and paragraph limits with line numbers, "-ing" warnings, parenthetical sentences). --marked also emits <draft>.marked.md with every unresolved word tagged. --verified prints the manual rule checklist.

Step 5 — Resolve (Verified only; max 3 total checker passes)

• NOT APPROVED → replace with the extracted alternative. The extraction is

mechanical: when it does not fit your meaning or part of speech, read the entry first with --entry <word>. A lowercase word CAN stay if the context makes it a technical noun or technical verb (rules 1.6 / 1.12) — annotate it and justify it in the note; replacement is the default, TN/TV the documented exception.

• NO ENTRY → annotate as TN/TV with a category, or rewrite without the word.
• REVIEW / CATEGORY REQUIRED / CONTRACTIONS / IDENTIFIER REVIEW / mechanical

findings → resolve each one.

• Rerun the checker after fixing. "Automated checks clean" is not the end:

complete the printed MANUAL CHECKLIST before applying the Verified label.

• If findings remain after the third checker pass, the deliverable is NOT

Verified: deliver the marked artifact with a "verification incomplete" note.

Step 6 — Deliver

1. The STE English text (verified, marked, or clean).
2. The translation, if the user asked in another language or requested one:

sentence by sentence from the STE English, same source term → same target term everywhere, conditions before commands, warnings keep the risk word → command → consequence shape. Formal STE compliance exists only for the English text — say so.

1. The compliance note.

Compliance note (required)

Verified deliverables end with exactly these three items:

• TN/TV: every declared technical noun/verb with its category number and

name — taken from the checker's DECLARED inventory plus any justified REVIEW words (e.g., "residual current device — TN, category 6, systems and components").

• Replaced: unapproved words from the source or draft and what replaced

each one (e.g., "qualified → APPROVED").

• Unverified: words that could not be checked — or "none".

The other levels end with one line instead:

• Marked: the level and the checker's flagged count (e.g., "Draft level:

marked — 7 words flagged by ste-check; not dictionary-verified STE").

• Clean: "Draft level: clean — annotations validated and stripped; the

dictionary was not read." (--clean classifies nothing, so there is no flagged count to report.)

• Raw: "Draft level: raw — model training only; nothing was checked."

A marked, clean, or raw draft must never present itself as STE-compliant, and "all vocabulary is approved" without a checker run behind it is never written.

Economy mode (optional)

If the harness supports delegating to a subagent with a cheaper model tier — or an ste-writer agent type is available — you MAY delegate Steps 3–5 to it and only relay its deliverable and compliance note. The checker keeps the quality floor deterministic, so a cheaper drafting model is safe: its vocabulary errors are caught mechanically, and its judgment calls surface in the compliance note. Never delegate the level choice (Step 1).

Common mistakes

• Asserting compliance from memory. "qualified" is not approved (use

APPROVED); "display" is approved as a noun but not as a verb; "should" is not approved (MUST, or IF for conditions). Run the checker first, claim after.

• Trusting an extracted alternative blindly when it does not fit the

sentence — the columns are parsed mechanically; --entry shows the full entry with meanings and examples.

• Treating a Marked, Clean, or Raw draft as STE. Only Verified — automated checks

clean AND manual checklist completed — may claim compliance.

• Keeping a 4+ word noun cluster. Rule 2.2: write it in full once, then

hyphenate the unit ("pressure-relief valve") or give a shorter form.

• Using Issue 8 terminology: "technical name" no longer exists — Issue 9

says technical noun / technical verb.

• Drafting directly in Spanish (or any non-English language). The dictionary

only exists for English — draft in STE English, verify, then translate.