Ambiguity Lexicon
Purpose: The detection surface — seven lexical classes, their vocabulary in English and Japanese, why each is ambiguous, and the disposition each defaults to.
Read when: You are in SCAN and need the catalog, or you are unsure whether a term counts as a detection.
Contents
- Detection rule
- The seven classes
- Open semantic detection
- Non-English source prompts
- Non-detections
Detection Rule
A term is a detection when two defensible readings of it would produce different deliverables. The lexicon below is a starting set, not an allowlist — a domain word absent from every table still counts if it fails that test, and a listed word does not count if the surrounding text already fixes its reading.
Record the competing readings, not just the word. "concise" → (a) under 200 words (b) no redundancy at any length is a detection; "concise — vague" is not usable downstream.
The Seven Classes
1. Quality
| Vocabulary | Why ambiguous | Default |
|---|---|---|
| good, high-quality, best, optimal, professional, top-tier, polished, refined, compelling, effective | Names a verdict without naming the standard that produces it | CRITERIA |
Quality words are the highest-value detections: they are what the author will judge the output by, and they are the ones most often answered with another quality word.
2. Quantity
| Vocabulary | Why ambiguous | Default |
|---|---|---|
| concise, brief, short, detailed, thorough, comprehensive, in depth, sufficient, moderate | A length instruction with no scale; "detailed" spans one paragraph to twenty pages | QUANTIFY, else BEHAVIORALIZE |
3. Explanation Level
| Vocabulary | Why ambiguous | Default |
|---|---|---|
| clear, easy to understand, simple, for beginners, for experts, broken down, carefully | Depends entirely on an unstated reader model | AUDIENCE + BEHAVIORALIZE |
4. Style
| Vocabulary | Why ambiguous | Default |
|---|---|---|
| natural, casual, formal, friendly, soft, intelligent, human-like, not AI-sounding | Register words with no shared referent; "not AI-sounding" names a negative no one can check | BEHAVIORALIZE |
"Human-like" and "not AI-sounding" translate to concrete surface rules (sentence-length variation, no section-summary closers, no tricolon padding) or they translate to nothing.
5. Design
| Vocabulary | Why ambiguous | Default |
|---|---|---|
| modern, simple, minimal, stylish, premium, Apple-like, current, intuitive | Aesthetic labels whose referents shift by year, market, and taste | BEHAVIORALIZE (visual rules) |
Named-brand comparisons ("Apple-like") are additionally unsafe as instructions to imitate; translate to the underlying properties (restraint in color count, generous spacing, one primary action per screen), never to the brand.
6. Technical
| Vocabulary | Why ambiguous | Default |
|---|---|---|
| clean code, efficient, fast, robust, secure, maintainable, extensible, best practice, production-quality | Each names a whole discipline; without a threshold or a checklist the executor picks its own | CRITERIA, QUANTIFY where a metric exists |
"Fast" and "secure" often belong to an enforcing layer rather than the prompt — a budget checked in CI, a scanner in the pipeline. Check before translating.
7. Judgment
| Vocabulary | Why ambiguous | Default |
|---|---|---|
| as appropriate, if needed, where possible, as much as you can, important ones, representative, recommended, latest | Delegates a decision without giving the criterion the decision needs | CONDITION, DATE for recency words |
This class is where the largest silent variance lives: "as appropriate" reads as "always" to one executor and "rarely" to another, and both are defensible.
Open Semantic Detection
After the seven classes, sweep for these — they carry no signature vocabulary:
- Missing actor. Who performs, who verifies, who decides. Especially common in subject-dropping languages, where the actor is routinely absent from the sentence.
- Missing object. "Improve it" — improve which property?
- Unstated comparison. "Better", "faster", "more consistent" — than what baseline?
- Undeclared scope. Whether an instruction covers one file, one feature, or the whole system.
- Implied precedence. Two instructions that will collide only in some inputs, with no stated winner.
- Unbounded enumeration. "List the options" — all of them, or a selection made on an unstated criterion?
Non-English Source Prompts
A prompt written in a language that drops arguments or nominalizes freely hides ambiguity in grammar as much as in adjectives. The lexical sweep above will miss all of it. Japanese is the common case here; the same patterns appear in Korean and, partially, in Chinese.
| Pattern | Ambiguity it hides | Detection cue |
|---|---|---|
| Dropped subject | Who acts, who verifies, who decides is undecided | A bare verb phrase ("verify", kakunin suru) with no agent anywhere in the sentence or its neighbors |
Open-ended enumeration suffix (nado, "and so on") |
The set has no closure rule — one more item, or twenty | An enumerated list ending in the suffix instead of a stated boundary |
Property-nominalizing suffix (-teki, "-wise", "-oriented") |
A whole discipline compressed into one adjective | efficiency-wise, coverage-wise used as if it were a bound |
Permission vs preference modals (-te mo yoi "may", -ga nozomashii "is desirable") |
Requirement level is unmarked — optional, preferred, or mandatory | A modal that could be read as any of MUST / SHOULD / MAY |
| Compound delegation | Neither the set nor the trigger is fixed | The enumeration suffix and a discretion word in one canon ("add examples and so on as appropriate") |
Translate the missing argument, not just the vague word. A dropped-subject "verify" becomes who verifies and what check they run — not an intensified "verify thoroughly", which changes nothing and reads as progress.
The rewritten prompt stays in the source's language (SKILL.md § Output Requirements); only the ledger and the derived rules follow the configured output language.
Non-Detections
Do not flag these; flagging them inflates the ledger and buries real findings.
- Terms the surrounding text already fixes ("concise — one paragraph").
- Domain terms of art with a single accepted meaning in context (
p95,idempotent,RLS). - Deliberate open-endedness the source marked as such ("pick whichever framing you prefer").
- Aesthetic or voice choices in a task whose purpose is exploration — those are
KEEPcandidates, not defects (ambiguity-budget.md). - Politeness and framing language that carries no instruction ("thanks in advance").