Three habits and the question behind each
| Habit | The question to ask | Warning sign |
|---|---|---|
| Check the source | Which filing is this from: which company, which form (10-K, 10-Q, 20-F), and is it the filing's own number or someone's paraphrase? | A figure with no document behind it, or 'according to reports' |
| Check the date | What date does the number describe, and when was it filed? Do all the rows share the same dates? | One 'current' figure assembled from documents filed weeks apart, with no dates shown |
| Respect a refusal | Did the answer say it could not find or compute something, and did it say why? | Asking the assistant to 'just estimate it', or reading an empty result as a zero |
Habit one: ask which filing
Captured 2026-09-27. Asked for a business summary of Costco (COST), the tool returned a paragraph plus labels saying where it came from: source_form 10-K, source_period 2025-08-31, and a note that the paragraph is a 'model-written paraphrase of the filing, not filing text'. Those labels answer the source question in full. The summary came from Costco's annual report for the fiscal year that ended August 31, 2025, and because it is a paraphrase, anything you want to quote as Costco's own words has to come from the 10-K itself. Notice the date too: on the capture day, that annual report described a year that had closed more than twelve months earlier, and it was still Costco's newest 10-K, because the report for the year that had just ended had not yet been filed. When an assistant gives you a figure without labels like these, ask for them: which company, which form, which period, and a link to the filing. If it cannot name a source, treat the figure as unverified, however precise it looks.
Habit two: every number carries two dates
Reading rows that span filing dates
Habit three: a refusal is an answer
Captured 2026-09-27. Asked for intrinsic-value estimates for Taiwan Semiconductor (TSM), the tool returned no numbers at all. Each of its three models came back empty with the reason no_us_gaap_annual_facts, and the summary explained why: companies that report under IFRS rather than US accounting rules 'are the usual cause -- the financials exist; these tools read us-gaap concepts only.' That is a good answer. It tells you where the tool's reach ends and that the numbers exist elsewhere. For a foreign company like TSM, that means its annual report on Form 20-F, which you can read directly on EDGAR. The weak response is to push back with 'just estimate it.' An assistant asked to fill a gap can produce a figure that looks exactly like a real one, and a number with no filing behind it is the thing habit one exists to catch. Treat a clear refusal with a reason as a sign of care, and follow its pointer instead of overriding it.
Habit three, continued: an empty answer is not a zero
Captured 2026-09-27. The same lender tool, asked about 'costco', returned zero rows and one sentence: 'not in the corpus under this borrower_norm as a funded debt position (equity/unfunded only, or a different key).' Read what it claims and what it does not. It says the private-credit dataset holds no funded Costco loan under that name. It does not say Costco has no debt; Costco's own 10-Q lists billions of dollars of long-term debt. The mistake to avoid is turning 'found nothing here' into 'there is nothing', whether you make it or your assistant does. When an answer comes back empty, ask two questions: what exactly was searched, and under what name?
Audit one figure you have already seen
Source, date, refusal
An assistant is only as trustworthy as the checks you run on its answers. Ask which filing a number came from, and whether it is the filing's own figure or a paraphrase. Ask what date it describes and when it was filed, and notice when rows mix filing dates. When an answer says it cannot compute something, or comes back empty, take it at its word and follow its pointer rather than asking for a guess. These are the same habits that make you a careful reader of filings; an assistant puts them to work more often.
Optional: practice in your own assistant
If you would like to practice these habits with the tools quoted above, you can connect any assistant that supports the Model Context Protocol (MCP), an open standard for letting an assistant call outside tools, by following the guide at oxfordledge.com/mcp.
Sit with the ideas.
You ask an assistant whether any private-credit funds lend to a large, well-known retailer. It replies that it found no rows under that name in its lender dataset, which covers funded loans only, and that the retailer may be listed under a different name. Which reading does that answer support?