--- name: verify-a-listing-against-its-source description: Check an entry against the page its subject publishes, not the aggregator that carried it, and write down only what that page actually said. runbook: schema_version: v1 semver: 1.0.0 profile: P1 min_runtime_profile: R1 domain: data-pipelines targets: - postgres trigger: manual execution: human-with-agent risk: reversible-write blast_radius: service duration: 5-30m capabilities: - cli:psql publisher: std lang: en license: CC-BY-4.0 relations: - kind: requires ref: std/enrich-from-pages-you-do-not-control@1.0.0 why: The fetch-level rules — what a refusal means, what a status code means, when a browser is needed — are there and are not repeated here. inputs: entry_id: type: string required: true description: the entry being verified source_url: type: string required: true description: the page published by the subject itself, not by an aggregator steps: - id: s1 kind: decision title: Is this the subject's own page? question: Was this page published by the thing being described, or by somebody carrying it? branches: aggregator: s2 default: s3 - id: s2 kind: decision title: Is there a page the subject publishes? question: Does the subject publish this itself anywhere, or does only the carrier have it? branches: carrier_only: s9 default: s3 - id: s3 kind: decision title: Is the entry still current? question: Does the page's title, or a direct statement in its body, say it is gone — rather than the words appearing somewhere in the text? branches: gone: s8 default: s4 - id: s4 kind: decision title: Is this page about only this entry? question: Does the page show other entries alongside — a menu, a catalogue, a list of neighbours? branches: shows_others: s5 default: s6 - id: s5 kind: action title: Take nothing from a page of neighbours risk: read-only expect: no field is filled from this page; a neighbour's words are not this entry's on_fail: escalate:s9 next: s7 - id: s6 kind: action title: Read the fields, and only what is stated risk: read-only expect: each value appears on the page as a statement about this entry, not as a word that matched on_fail: escalate:s9 next: s7 - id: s7 kind: action title: Write only the empty fields risk: reversible-write tool: cli:psql command: psql -Atc "update entries set checked_at = now() where id = '{{inputs.entry_id}}' and checked_at is null" expect: fields already filled are left alone, and nothing unstated is invented on_fail: escalate:s9 next: end:success - id: s8 kind: action title: Retire the entry, and say on what evidence risk: reversible-write requires_approval: true expect: the entry is withdrawn, with the page and the words that established it on_fail: escalate:s9 next: end:aborted - id: s9 kind: escalate title: Leave it as it is and ask to: team:platform source: url: https://runbooks.directory/r/std/verify-a-listing-against-its-source/source.md fetched_at: 2026-09-04 upstream_state: current --- ## When to reach for this An entry in your system describes something somebody else publishes — a listing, a product, an event, a filing — and you are about to act on it: show it to a reader, send somebody to it, or take it down. What follows is how to check it without being fooled by the page. The fetching rules — what a refusal means, what a status code means, when a page needs a browser — are in `std/enrich-from-pages-you-do-not-control` and are not repeated here. This is about reading the page once you have it. ## Judge by the source, not by the carrier The rule that changes the most outcomes: **the subject's own page decides.** An entry marked closed by an aggregator was open on the subject's own site — with details the aggregator never carried. Acting on the aggregator would have retired a live entry. And the reverse happens too: a page whose title looked perfectly healthy said, in its body alone, that it did not exist. Two automated checks passed it. An aggregator is a copy, made at some past moment, by somebody with no obligation to keep it current. ## What a page will make you get wrong Every one of these cost a separate investigation, and each was found on real data. **A neighbour's words.** A level was read as junior because a side menu listed other entries, one of them junior. The word was on the page; it was not about this entry. **Wording that reverses.** "Contract type: permanent" contains the word *contract*, and a matcher looking for it records the opposite of what the page says. **A number from the next line.** Two figures sit side by side on one page and belong to different requirements. The nearer one is not necessarily yours. **A page of neighbours.** Some sites show dozens of other entries, with their own figures, before the one you asked for. Nothing on such a page can be attributed — and by the rule above, it is not the subject's page anyway. **A form instead of a description.** The link goes straight to an application or enquiry form: fields to fill in, and nothing that describes the thing. **A figure that is not a range.** "Starting from" is one number, and a matcher looking for two never sees it. Absent is not the same as none. **Words about closure on a living page.** Accept *gone* only from the page's title or a direct statement about this entry in its body — never from the phrase appearing somewhere in the text. ## Write only what was stated Three rules, and the first is the one that decays a database quietly. **A guess looks exactly like a check.** A value put in by inference is indistinguishable afterwards from one that was verified, and there is no later moment at which somebody separates them. Not stated on the page means empty in the record. **Do not overwrite what is filled.** An existing value may have been checked by hand. A sweep that overwrites is a sweep that discards work nobody can recover. **Never correct a name automatically.** Across one batch, 56 of 114 name discrepancies were real and of entirely different kinds: an abbreviation expanded, an acquired subsidiary trading under its own name, the same organisation in another language. Each needs a different answer and none of them is a rename. The name is what the reader sees, so a wrong correction is a wrong statement about somebody else.