Domain evidence discipline

Domain Evidence Discipline

This skill's evidence-control loop stays constant across domains; only the nouns change. For any non-code deliverable, name the four nouns BEFORE producing anything — in your working notes or the report, where a reviewer can see them — and hunt the same four afterward. Under this skill, a deliverable whose four nouns were never named is incomplete and unauditable.

Red-line gate — run before anything below. Where the deliverable would substitute for individualized, materially high-stakes professional or regulated judgment — a medical/clinical decision, legal advice, a buy/sell financial call, mental-health treatment, safety-critical engineering sign-off — refuse the deliverable and route to a qualified human: an evidence checklist supplies structure, never the judgment, and its presence invites trust it cannot back. General work that merely touches money or health (a budgeting spreadsheet, fitness copy) is not red-line; the line is substituting for the professional's individualized call. Adjacent work that is not substitution (compliance research, tooling for practitioners) may be drafted under the matching instance, labeled as not the professional's individualized call — and when the user would reasonably treat it as professional work-product, it ships only with a qualified human's review; unreviewed, it stays a labeled draft. This is the execution-side sibling of skill-authoring §5's authoring gate (that rule stops the checklist being WRITTEN and gates adjacent authoring; this one gates the deliverable being USED as the professional's answer).

1. The four nouns

1. Minimum evidence set (binding). The sources that must actually be OPENED before any output: the governing document (brand rules, spec, data dictionary) and the subject's own facts (the product's real materials, the raw dataset) — every time. Before declaring a governing document absent, search the workspace and the supplied materials for it and say where you looked; only then state the assumption you will work under. An OUTWARD claim — an externally checkable assertion about third parties or the current world state, even when repeated from supplied material — needs a live source fetched now, with enough sources to cover each materially distinct outward claim; a deliverable with no outward claims owes no fetch (evidence is claim-triggered, not ceremony). If a required source cannot be opened (the dataset is missing, a referenced brand guide was never supplied), stop and request it, or narrow the deliverable and label it explicitly unverified — never proceed as if verified.
2. Authority order (typed). Different questions have different owners; one flat chain does not fit all. The client/user owns objectives and authorized creative choices; mandatory rules (brand bans, legal disclaimers, policy) own compliance; current primary evidence owns factual assertions; the documentation owns what a field MEANS (raw values never redefine semantics); your taste and memory come last. An instruction never overrides observed facts, a mandatory constraint, or the red-line gate — surface the conflict instead of silently picking a side ("punchy copy" in a brief never overrides a brand rule banning superlatives).
3. Verification by observation. What "observed" means here: every factual claim traces to a source you opened; names, prices, dates, versions exact; arithmetic recomputed, not transcribed; rendered surfaces actually rendered and looked at. A claim you could not verify is removed or explicitly flagged — never left standing as fact.
4. Fraud table. The domain's classic fabrications, written down so a reviewer can hunt them by re-opening sources, re-fetching, and recomputing — the non-code counterpart of re-running the tests: it hands delegation-and-review §3's reviewer the domain's rubric.

2. Worked instances

Marketing / copy. Evidence set: brand rules (or the search-then- assumption path above), the subject's own capability facts; live fetches per the outward-claim rule (competitor, market-position, and price claims are outward even when the brief asserts them). Authority (typed): the client owns the objective and voice choices within the brand guide; the brand guide owns compliance; the subject's materials own capability facts; taste last. Observed: line-by-line check against brand rules; every capability claim traces to the subject's materials; prices/names/dates exact; the rendered page or email actually rendered. Frauds: fabricated statistics ("studies show" with no opened source), fake social proof (invented testimonials, counts, awards), unverifiable superlatives ("#1", "guaranteed"), stale prices or dead offers presented as current, missing required disclaimers, spec betrayal (copy violating a written brand rule while claimed on-brand).

Research / report. Evidence set: the sources the claims actually rest on, opened this session with access dates — fetched live when the claims are outward; an internal-sources-only synthesis owes no external fetch; one counter-source deliberately sought when the report advances a contestable outward central claim. Authority (typed): primary source owns facts; reputable secondary > aggregator for context; memory last and labeled. Observed: every figure carries a link and access date; arithmetic and unit conversions recomputed; quotes checked against the page they came from; currency of volatile facts (prices, policies, versions) checked against the source's date, not assumed. Frauds: uncited figures, stale figures presented as current, phantom precision (exact-looking numbers with no computation behind them), cherry-picked date ranges, a "sources" list whose entries were never opened.

Data analysis. Evidence set: the actual dataset (not a described version of it), its dictionary/provenance, the requester's definition of the metric in writing. Authority (typed): the data owns the values; the dictionary owns what a field means (raw values never redefine semantics); the requester's written metric definition owns the metric; recollection last. Observed: headline numbers recompute from raw in one command; dedup, exclusions, and imputations declared (operational-rigor §4 data-path integrity: no silent defaults, unmatched records surfaced); the plain-language summary agrees with the computed table it sits above. Frauds: silent data cleaning, dropped unmatched records, totals that do not reproduce, a "verified" label on arithmetic never recomputed, precision the data cannot support.

Hybrids and routing. A deliverable claiming in several domains (a market-research report for a campaign) takes the union: every domain it actually makes claims in contributes its evidence set and fraud table, and authority stays typed per claim. Business/ops documents route by primary claim type — financial and metric claims → Data analysis; market and external claims → Research; persuasive copy → Marketing. Applying the four nouns task-locally, alone or in union, is NOT minting an instance and carries no fixture duty.

3. Minting a new instance

• Mint a reusable instance only when at least one noun materially differs from every existing instance — name the differing noun in the instance. If none differs, the sector needs no instance: route to the nearest one and note the boundary.
• A sector that is coding in disguise (IaC, script logic, pipelines) routes to operational-rigor. An analysis PRODUCED through a runtime (SQL, a notebook, spreadsheet formulas) but DELIVERED as claims uses both skills: operational-rigor governs the execution and the data path; this skill governs the evidence, the claims, and the review.
• The red-line gate above applies to minting exactly as to producing: no instance for red-line judgment; a sector adjacent to one follows skill-authoring §5's qualified-reviewer requirement.
• Every regulation, policy, threshold, or figure the instance names carries a source and access date, fetched now (operational-rigor §4: a clue about external data is a map, not a schema; memory is labeled).
• A NEWLY minted instance without its own trap fixture is not done — the README covenant applies to instances exactly as to rules (the fixture's trap is the sector's central fraud; its safe outcome obeys ground-truth-gates rule 2's trap-armed clause). The three founding instances above carry that debt explicitly — see Provenance.

When NOT to use this skill

• Code, infra, or pipeline deliverables → operational-rigor (+ ground-truth-gates). An analysis produced through a runtime but delivered as claims uses both skills (§3).
• Substantial open-ended source discovery → the environment's research or deep-research tooling does the gathering; this skill is the evidence-control overlay on what that workflow produces, not a replacement for it (and it does not replace these gates).
• "What should we build next?" → product-roadmap. Personal goals → personal-goal-planning.
• Red-line professional judgment (the gate above) → no checklist; a qualified human.
• A one-off trivial lookup — quoting one number with its source needs no ceremony.

Provenance

Promoted 2026-07-18 from a 2026-07-16 staging draft (owner decision D1, option b: one condensed pattern skill, not eight ported adapters). The four-nouns pattern, the binding minimum-evidence-set idea, the per-domain fraud tables, and the instance-needs-its-trap rule adapt Sahir619/fable-method's domain-adapter schema (MIT; ideas only, no files copied; see README acknowledgements). Their round-9 result is the external evidence, cited as shape with its grade: with the marketing adapter a weak-tier executor found unmentioned source documents and caught planted frauds that a bare run praised (their measurement, their fixtures; their round-9a null — a prompt that names the evidence pre-solves discovery — is why the evidence set binds the OPENING and searching-out of sources, not their mention). The typed authority order and the adjacent-deliverable review clause came out of this PR's own dual-family review (a flat chain let an instruction outrank facts and mandatory constraints; adjacent PRODUCING had no runtime twin to §5's adjacent-authoring gate). The worked instances are this pack's own compressions, cross-checked against the pack's existing rules (operational-rigor §4; delegation-and-review §3; skill-authoring §5's authoring-side red-line, deliberately not duplicated verbatim — sibling gates with different jobs, each in its own words). unprobed in-house per the README covenant: no probe has run against these instances; the owed probes are three non-code trap fixtures, one per founding instance — marketing, research, and data analysis — each arming its instance's central fraud with the trap-armed check.