Source-required research card

A one-page gate for any large language model research job. Every claim needs a URL, a verbatim quote, a retrieval date, and a source type. No citation, no ship.

Tick this before a research workflow feeds anything a client, a deck, or a campaign will use. If a box is unticked, the output stays internal.

The model proposes claims. A script checks quotes. A person approves. Do not skip the person because the script passed.

flowchart LR
    F["Fetch"] --> E["Extract claims"]
    E --> V{"Schema + quote check"}
    V -->|Fail| D["Drop and log"]
    V -->|Pass| R["Human review"]
    R --> S["Ship with URLs"]

Name the job

Research question (one sentence)
Where the output will be used (deck, brief, message, internal only)
Reviewer (one named person)
Context pack version loaded, if any

Every claim is a record

Field	Rule	Fail closed if
claim	One statement, plain words, capped length	It bundles two claims or contains a dollar figure with no quote
url	https only, a specific page	It is a homepage, a search result page, or empty
quote	Verbatim text from the fetched page	It does not appear in the text the workflow downloaded
retrieved_on	Date the workflow fetched the page	Missing or in the future
source_type	primary, vendor, press, forum, unknown	Outside the list

Records are validated against a JSON Schema before anything else runs
Page text comes from our fetch step, never from text the model supplied

Checks, in order

Schema check passes or the record is rejected and logged
Quote check passes or the claim is dropped and logged
Reviewer confirms the quote supports the claim, not something narrower or older
Reviewer confirms the source type fits the claim

Source-type weights

Type	Can carry	Needs a better source for
Primary	Rules, definitions, limits, legal wording	Opinions on real-world performance
Vendor	What the company says it offers today	Independent comparisons or results
Press	Launches and announcements with outlet and date	Product detail the article may have simplified
Forum or social	Leads to investigate	Anything printed or sent
Unknown	Nothing	Everything. Route to a human to classify.

Drop rules

Do not repair a claim by asking the model to find support for a sentence it already wrote. Drop it, log it, move on.

Quote not found: dropped and logged with the run id
No URL: dropped
Weak source type for the claim: parked with a flag for the reviewer
Dropped claims stay in an unsupported list. They never enter the output.

Fetch and safety

Public pages only. No logins, no accounts that are not ours.
robots.txt respected. Site terms read.
Fetched text is treated as data, never as instructions
No send, post, or write tool in the same step as the model
Fetched text stored with the run so the page can be checked later

Before it ships

Each shipped sentence keeps its URL and retrieval date next to it
Drop rate logged. A sudden rise means the prompt or model changed.
A golden set of known questions reruns after any prompt or model change
One flag pauses the workflow