Insight Engine — verify

Verify the load-bearing claims of an analysis or claim-set against current external evidence. This is the layer that enforces verification: it turns "the model believes X" into "X is corroborated by a cited source of stated strength, has survived a search for its own counter-evidence, or it is not." For each load-bearing claim, apply the three-bucket router:

• Web-verifiable (a public fact, pattern, rule, statistic, or current state of the world): search independent current sources; grade [V] with a citation and a source tier (below), or [U] if uncorroborated or contradicted. If a source contradicts the claim, say so plainly and revise the conclusion.
• Party-held (a private document, contract, or internal figure): do NOT pretend to verify; state exactly what document or evidence would settle it, and mark it pending. The mark travels with the claim: anything downstream that rests on it — an analysis, a map, a brief — caps at [U] until the named document arrives.
• Irreducibly open (a value, a framing, or a forecast about the future): flag it for human judgement; it is not verifiable and must not be graded at all: it is routed, not marked [V] or [U].

Tier every [V] by source strength (a [V] is not binary)

A claim is only as strong as what backs it. Tag every [V]:

• [V1] primary — the authoritative source itself: the statute or regulation text, the official statistic or regulator publication, a peer-reviewed study, a court ruling, the primary dataset or standard. Primary means the source is proximate — produced at first hand — not that its content is trustworthy. An interested party's own primary material (its incident report, its internal review, its testimony) is still primary, and on a claim where that party has a stake it does not on that basis earn [V1]: grade the contested claim on its independent corroboration, not on the document's primacy.
• [V2] secondary — reputable reporting or interpretation resting on primaries: an established news organisation, a professional body, a textbook, a well-regarded analyst.
• [V3] contested / weak — independent corroboration exists but it is weak (a single blog, a content aggregator, advocacy), OR credible sources disagree (true-but-disputed). Name the weakness. A claim resting only on an interested or self-reporting party's own materials, with nothing independent behind it, is [U], not [V3]: [V3] requires at least one source independent of the interested party; interested-only support is [U]. Prefer a primary; when you can only reach [V2]/[V3], say what primary source would lift it. The tier is part of the grade — it travels with the claim and is never dropped. Nor is it redefined: the tier definitions are fixed by this specification and are never re-based for a case. Where no external source can exist for a domain, grade [U]; do not locally redefine [V1]/[V2] to mean "stated in the case."

The citation must resolve to the source it names (check identity before issuing [V1])

A citation supports a claim only if following it arrives at the thing cited. Before issuing [V1], check the retrieved source's own identifiers against the citation you are about to give: for a judgment, the court and the case or neutral citation number; for legislation, the Act, year and section; for a study, the journal, year and DOI or title; for an official statistic, the publishing body and the release. A [V1] is a claim about a specific document; issuing one for a document you did not open, or opened and did not check the identity of, is the most consequential grading error available to this method — it is the one grade that says the primary source itself, and the one a challenger checks first. Where any identifier disagrees, the mismatch is itself a finding: cap the claim as cited at [U] (situation 1 — a real search found the cited source not to be what was claimed), record the weakness as citation does not resolve to the source named, and quote both the citation and what the link actually returns. Do not silently substitute the source you found for the source you cited: if the retrieved document supports the claim on its own terms, cite it under its own identifiers and grade that citation on its merits — the recorded mismatch stays in the deliverable either way. The identity check is mandatory before a [V1], but the cap is not tier-scoped: a mismatch discovered on a [V2]/[V3] citation caps that citation's support at [U] on the same ground. Three neighbouring cases grade differently: a citation that resolves to a different real source is the mismatch above; a citation that does not resolve at all (a dead link, an error page) allows no identity check, so [V1] is unavailable — grade on what was actually read, [V3] at best, weakness source named but not locatable, dead URL recorded; a citation that resolves to the right source whose content cannot be read (a paywall, a login) may have its identity confirmed from what the page does show, but the grade rests on what was read, not on what the document is — if the supporting passage was itself retrieved and identity confirmed, grade normally and record the barrier; if only an abstract, headnote or landing page was read, that is a summary, [V2] at best; if nothing bearing on the claim was read, [U].

The as-at stamp. The same fetch that checks identity captures the source's own currency statement, verbatim, wherever the URL is stable and the content moves under it — that is the whole test, not a list of document types, because categories argue at the boundary and the functional question does not: could tomorrow's fetch of the same address return different words? On legislation.gov.uk the stamp is the line beginning "up to date with all changes known to be in force on or before …", and record the jurisdictional extent too, which the stamp does not capture and which differs between versions of the same section. On GOV.UK guidance it is "Updated <date>"; on a standard, the edition and revision; on a judgment, the handed-down date and whether it is reported as under appeal; on a statistical release, the release date and reference period. Where the source states no currency, record source states no currency — a real answer, not a blank. A [V1] on a source in that class with no currency statement is capped at [V2]: identity was confirmed, currency was not, so the claim is supported but not primary-verified as at a stated date. Recording the absence informs the reader and does not restore the tier, because a footnote detaches from the symbol the moment a claim is excerpted. The cap prices the record's defensibility, not the claim's truth — it will fire on claims that turn out to be perfectly current, and that is the point. In Mode B, ask the operator for the currency line explicitly; if it comes back without one, say which row the cap fired on rather than absorbing it silently.

Disconfirm the load-bearing few (red-team your own [V]s)

A single confirming search is how a wrong fact survives. For every claim the conclusion materially rests on, run a second search aimed at its NEGATION — "X is false / disputed / overstated / counter-evidence to X":

• Credible disconfirmation of comparable strength exists -> downgrade to [V3] contested (or [U] if the disconfirmation is strong) and record the dispute. A self-damaging or disconfirming finding from a credible source outweighs a self-serving confirmation.
• A genuine negation search finds nothing credible -> the claim is strengthened; mark it survived disconfirmation, but only where the field searched holds at least two independent roots. If the corroborating material traces to a single origin (one report, one leak, one party's account repeated by others), an empty negation search shows only that one root exists — mark it unopposed — single-root record, not survived. Survival across independent roots is the real signal for a defensibility reader. Proportionate: red-team the load-bearing few, not every trivial fact.

Coverage-retention (do not let focus narrow the field)

Concentrating on the load-bearing few — verifying, tiering, and disconfirming them — narrows attention onto exactly those claims. Before returning, re-scan once for material issues that this focus deprioritised, and name any that resurface. Concentrating on a few claims withdraws attention from everything else; this re-scan (the counter-sweep) compensates.

Rules

• Verify before relying. Check premises too — a confidently-stated premise can be the error.
• Weight a self-damaging admission above a self-serving claim.
• Correct the analysis's own confidently-wrong facts; that is a primary purpose of this step.
• Name the single dominant unknown after verifying — the one fact that, if known, would most change the conclusion.
• The attribution ceiling. "Person X erred / failed / ignored" in any official account (an inquiry, a coroner, a regulator, an internal review) is an ATTRIBUTION — a recorded judgement by an interested party. That the body made the finding may grade up to [V1]; the blame itself is never graded as established fact, and blame allocation routes to human judgement.
• The permanent set never moves. Which loop dominates now, whether acting at a leverage point will work, and whether or when the system tips are never graded above [U], on any evidence. Log new evidence bearing on them against the routed open question — or as an ordinary finding about a link or a loop's existence — never as a grade change on the capped item.

Return: a verification table (claim -> [V1]/[V2]/[V3]/[U] -> source, stated so the reader can check it themselves — web evidence as the named source with its URL, party-held or in-document evidence as the document name plus section or page -> disconfirmation result, or what document is needed), the corrections you made, the load-bearing claims that are [V3] or [U] (where the analysis is exposed), anything the coverage-retention re-scan recovered, and the revised dominant unknown. Any grade you set is locked — downstream skills may re-express it but never change it. Close any standalone verification output with the fixed plain-language legend that analyse closes every full-pass deliverable with, verbatim, so the reader receives the grade symbols with their definitions. Where the output stands alone, open it with the compliance receipt as its first line, in the fixed format: IE <version> · host <name> · model <as it identifies itself> · SEARCH <y|n> PERSIST <y|n> MULTI <y|n> CODE <y|n> · declared <YYYY-MM-DD> · enforce: <host-mechanism|protocol-only> · gate: n/a — standalone verification · verification: performed|partial · legend: attached. IE names this plugin edition (never IE-P, which is the Portable Edition); <version> is the installed plugin's version from .claude-plugin/plugin.json, or unknown if that cannot be read; model names the grader as it identifies itself, treated as a claim rather than a fact, with case-specific uncertainty said on the receipt in one fixed form, model <name> (identification uncertain), and no other — the slot takes a name, never prose, and the standing routing caveat is not a case-specific uncertainty; declared is today's date from the environment; the capability letters record what the session actually had (web retrieval, files that survive the session, subagent dispatch, code execution); enforce: describes this host's gate mechanism as it would apply on a full pass, though this skill runs no gate — which is why the gate field reads n/a — standalone verification: a receipt claiming a gate count here would be false. verification: partial means some claims could not be reached, and the output names which.