How can you tell a sourced fact from an agent’s inference?

Split the sentence into claims

Consider an invented draft: Company A reported higher revenue, so demand is durable and its strategy is working. That sentence combines a reported measure, a business interpretation, and a causal claim. A link to a revenue table may support the first part while leaving the other two unresolved.

Label each clause before evaluating the whole paragraph. Useful labels include reported fact, calculation, interpretation, hypothesis, and unknown. They describe the claim’s relationship to evidence, not its importance. A calculation can be arithmetically correct while using incomparable inputs.

Make a claim card

For the hypothetical revenue statement, record the document title, issuer, reporting period, publication date, table or section, units, and accounting basis. Preserve the exact passage needed to verify the claim in your private working notes. Then state precisely what that passage does and does not establish.

A separate calculation card should identify both inputs and the formula. If a number is rounded in the original disclosure, do not imply extra precision in the result. An inference card should identify at least one plausible competing explanation and the evidence that would help distinguish it.

Check the context behind the link

Open the original source rather than accepting a generated citation label. Verify the issuer, period, units, and whether an amendment changes the relevant figure. A real document can be cited for a statement it does not contain. A source published after a historical decision cannot establish what was known at that earlier time.

The SEC’s EDGAR API documentation explains that company data includes filing histories and extracted XBRL facts. It also describes units and aggregation across filings. For review purposes, treat a machine-readable value as an entry to investigate with its filing context, not a complete explanation of a business outcome.

Run a disagreement check

Give another reviewer the claim card without the proposed investment conclusion. Ask whether the passage supports the literal statement, which qualifications are missing, and which conclusion remains an inference. If the reviewers disagree, preserve both readings until the underlying definition is resolved.

Try a deliberate negative case: attach a valid source about a different fiscal period. A useful check should reject the mismatch. Also test a passage containing a forward-looking target rather than a completed result. The exercise evaluates citation handling and temporal reasoning on those cases only.

Write the answer with boundaries intact

A clearer hypothetical conclusion is: the issuer reported this measure for this period; the result is consistent with one explanation, but does not establish that explanation. Follow it with the unresolved question. This gives the reader a route to verify or challenge the inference.

In Thesis Lab, review the displayed source and its date alongside the simulated decision. Current SEC evidence uses dated snapshots for Microsoft, NVIDIA, and Apple. It is not a complete live filing corpus, and a displayed source does not prove that every sentence in a generated answer is supported.

Method, sources, and limits

The claim cards and mismatch exercises are original hypothetical examples. SEC documentation and the Thesis Lab release were checked on 26 October 2026. This method helps inspect evidence; it does not validate an issuer’s financial statements, establish causation, or provide an investment recommendation.