docs / 01-claims.md
Claims
A claim is a sentence plus the words in a source that carry it. The draft, the analysis and the check are all built on claims.
What a claim is
The agent writes the sentence. The quoted words come from a model reading the page, and are kept only if they are found in the stored text, with their position in it. So anyone can open the page and read them in place, and the model cannot quote words the source does not contain.
What a claim says is never edited. When a second source says the same thing, it is added to the claim, and the claim then shows both.
A sentence in the draft and a claim are different things. A paragraph can rest on several claims, or on none. The claim is what the paragraph is checked against.
Three ways to ground a claim
Most claims quote a passage.
Some come from a calculation over a dataset. "Two ministries stopped appearing together in the news after 2023" is a pattern across thousands of articles, and no single one says it. The query is saved so the number can be produced again.
Some follow from other claims, and they come in two kinds.
- A synthesis is a broader fact built from several findings. "Three audits in a row found unspent capital budgets" is in no single report; it rests on three claims, one per audit. A reviewer checks it by reading those claims.
- A judgement is a reading. "Union resistance to reform is party resistance" is in no document. It rests on a claim about who leads the union and a claim about who appoints the commission. It is marked as its author's, and a reviewer argues with it.
A chain of claims resting on claims is normal. A chain that never reaches a passage or a calculation is flagged.
How sure, without a score
Three labels, read together and never added up.
| Label | Values | What it asks |
|---|---|---|
| Support | explicit · paraphrase · inferred | Does the source say it, or did we put it together? |
| Corroboration | single · corroborated · contested | How many publishers say it? |
| Standing | party · observer · authority | Does the publisher have a stake in this claim? |
No agent picks the first two, because a label a model picks drifts upwards.
Support starts at inferred, and a separate check raises it only when the
quoted words say what the sentence says. Corroboration is counted from the
sources the claim rests on, so it changes when a source is added.
A claim reads explicit · single · observer, not 0.72. A reviewer can check
each label against the source. A number has nothing to check it against.
Standing is set per claim. A ministry is the authority on what its budget line says, and a party to whether the money was well spent.
Following a citation back
The model picks a claim. Code writes the endnote by following the chain. The model never handles a URL, so it cannot invent one.
Each step is one click. Most readers stop at the passage, to see whether it really says that.
A gap is a result
When the agents look for something and cannot find it, they record a gap: what was wanted, where they looked, and who would know.
When the analyst knows the answer, they say so in chat. It becomes a claim with their name on it, and the gap records that a person closed it, not a search.
Counting gaps like this shows where open sources run out and interviews are still needed.
When sources disagree
Both claims are kept side by side, with a line saying what the disagreement is about. It may be an error, a change over time, or two people each describing the part they can see. The system never averages them. In political economy, the disagreement is often the finding.