Overview

Judge a claim against HG data the way an auditor would. Claude splits compound statements into checkable claims, matches entity, scope and time period before comparing, marks a claim Contradicted only on direct positive evidence, never treats a missing row as proof of absence, and keeps a record that lets anyone reproduce each verdict.

Use cases

Fact-checks that hold up in the room
When a document says a company runs a product, uses a cloud or spends a figure, Claude checks it against the matching HG record and cites the tool, figures and dates behind every verdict, so a seller can defend each correction.
No false alarms from missing data
A product HG hasn't detected isn't proof it's absent. Claude reports missing, weak, stale or unavailable evidence as couldn't substantiate, so a document is never marked wrong for something HG simply can't see.

View full skill

HG Evidence Discipline

When to use

A workflow checks statements in a document (an account plan, brief, battlecard or email) against HG data.
A workflow classifies accounts or claims and each label must rest on evidence someone else can reproduce.
A reviewer asks "how do you know that's wrong?" and the answer must name a tool, figures and dates.

The protocol

1. Split into atomic claims

Break compound statements into claims that one lookup can check. "They run AWS and Snowflake and spend $40M on cloud" is three claims. Keep the original wording and where it appears so the verdict can quote it.

2. Match scope before comparing

Before comparing a claim with a record, match:

Entity: the same company, resolved by ID or domain, not by a similar name.
Hierarchy level: parent, subsidiary or a single site. An install at a subsidiary can make a parent-level "they don't use X" wrong. Use hg-subsidiary-handling to decide which level the claim is about.
Geography: global, a country or a region.
Product or vendor: the named product, not its category or another product from the same vendor.
Time period: the year, quarter or "current" the claim refers to.
Metric: installs, spend, headcount or revenue. Don't compare a modeled IT budget with a reported revenue line.

If a dimension can't be matched, the evidence is indirect.

3. Assign exactly one verdict

Confirmed: direct, scope-compatible evidence supports the claim.
Contradicted: direct, scope-compatible positive evidence is incompatible with the claim. Example: the claim says "no Google Cloud" and HG records a current Google Cloud install with a verification date.
Couldn't substantiate: evidence is missing, ambiguous, stale, unavailable, errored or only indirect.

4. A missing row is never a contradiction

HG not detecting a product, contact or figure does not show that it doesn't exist. Absence of evidence goes to Couldn't substantiate, with what the claim now rests on (for example "the document's own source only").

5. Freshness and signal strength

An install is current when it was last verified within 90 days.
A stale signal (last seen more than 180 days ago) or a weak signal (very low intensity for that company) cannot contradict a claim. Report it as "weak signal" or "stale signal" with the figures and dates, and use Couldn't substantiate.
Between 90 and 180 days, say "last verified [date]" and weigh it as supporting evidence, not proof.
See hg-recent-signals for date fields and placeholder dates.

6. Intensity is not dollars

Install intensity is relative usage within one company. Never convert it to spend and never compare it across companies. "One of their most-detected products" is fine; "heavier user than their peer" is not.

7. Label modeled figures

Spend figures from HG spend tools are modeled. A claim of "$40M cloud budget" checked against modeled spend is compared as an estimate and labeled "modeled". A modeled figure in a different range does not contradict a precise reported figure unless the gap is far outside the estimate's band. Follow hg-modeled-vs-observed for wording.

8. Conflicting evidence

When two sources disagree, prefer, in order: the source that is closer in scope (entity, level, period), then the stronger source type (company-reported filings and observed installs over modeled estimates), then the more recent one. Record the conflict and the reason for the choice next to the verdict.

9. Unavailable sources

Some tools depend on the organisation's entitlements: data-query access for hg_data_query, an SEC data entitlement for the SEC tools, and a connected contact provider for contact_search. If a tool is unavailable, returns an entitlement error or fails, mark the claims that needed it Couldn't substantiate ("source not available to this organisation"), and keep checking the other claims. Never stop the run for it and never guess the answer.

10. Keep a claim-to-evidence record

For each claim keep: the claim text and location, the tool called with its key inputs, the figures and dates returned, the verdict, and the reason. The record must let another person rerun the lookup and reach the same verdict. Cite it as hg-citation-discipline describes.

Common pitfalls

Calling a claim false because a product doesn't appear in the top results. Filter by product before concluding it's absent, and even then use Couldn't substantiate.
Checking the parent company when the claim is about a subsidiary, or the reverse.
Quoting an install with a 2000-01-01 first-seen date as an adoption year.
Treating a modeled spend figure as a reported number.
Marking a filing-based claim as wrong when the SEC tools weren't available.
Giving a claim two verdicts, or none.

Related skills

hg-citation-discipline for how to cite each lookup.
hg-recent-signals for recency windows and date fields.
hg-modeled-vs-observed for how much confidence each source supports.
hg-sec-research for checking financial results, risk language and disclosed events.
hg-subsidiary-handling for picking the right hierarchy level.

Skill: HG Evidence Discipline