verifier.md

The verifier

The brief is only useful if I can check it. verify_brief() in src/spend_radar/brief.py does that check.

What a claim is

Sections 2 to 6 of a brief are lists of claims. Each claim has three fields:

• text: what we say
• evidence_id: the evidence item it rests on
• quote: words copied from that item

Evidence comes from two places. Items from a provider (fixtures or live) are tagged with source and date. Items about our own contract, usage and spend are built from vendors.csv and the analysis, so a claim about our data is checked the same way.

What gets flagged

A claim is flagged if any of these is true:

• it has no evidence_id, or the id is not in the stored evidence
• it has no quote, or the quote is under 10 characters
• the quote is not found word for word in that item's title or snippet (white space is ignored, case is not)
• it contains a number that does not appear in our data or in any evidence item
• it contains a date that does not appear in our data or in any evidence item

Numbers are checked by type. A claim with $9 needs a $9 somewhere in the evidence. A 9% in the evidence does not count. This stops an invented price from slipping through on a coincidence.

One number is allowed without evidence: the uplift cap we ask for (3%). It is our own target, set in core_logic.py.

What happens to a flagged claim

It is removed from sections 2 to 6 and listed under Unverified, check before using, with the reasons. Nothing is dropped silently. The heading always appears. If nothing was flagged it says so.

What it does not check

• It checks that words and numbers come from the evidence. It does not check that the evidence is true.
• It does not check that a claim means what the quote means. A person has to read it.
• Fixture evidence is invented. Live evidence is a headline and a short snippet, so read the source before you quote it.