Yesterday’s article opened with a wonderfully dangerous sentence: “6,465 credits in the account. Five left when the lesson was over.”
It is crisp, memorable, and true to the dated production record. It is also four epistemic conditions wearing one trench coat.
The starting balance and recorded spend were observations in a preserved ledger. The conversion to “approximately $100” was an inference from the purchase rate available in that workflow. Whether the method transfers beyond one production is missing. And “five credits left” became outdated the moment the account changed again.
Nothing needs to be retracted. The sentence simply cannot be reused safely unless its seams remain visible.
A polished answer hides its joints. An epistemic audit puts a small red pencil mark on each one.
The four marks
Epistemic auditing is a claim-level inspection performed before an answer is acted on—or after it fails. It separates what the record supports from what reasoning adds, what the decision still needs, and what time has weakened.
These are marks, not four sealed buckets. “Five credits remained” is K for the recorded end of that test and O if someone uses it as the present balance. A market estimate may be I and also O. A search result can be K about what was found and M about whether the thing exists elsewhere.
A six-move audit
- Bound the decision. State who will use the answer, for what action, and by when. A dinner suggestion and a medical decision do not deserve the same audit depth.
- Atomize the claims. Split compound prose until each sentence fragment can be supported or challenged independently. Numbers, identities, causal explanations, forecasts, and present-tense states deserve their own lines.
- Apply K, I, M, and O. Mark every consequential claim. If a claim cannot earn K or I, it is not ready to appear as an answer. M and O may be added to either.
- Attach the receipt. For K, record the source, observer, event date, last verification, and scope. For I, record the assumptions and at least one plausible alternative. A URL alone is not provenance.
- Attack the bridge. Start with the conclusion most likely to change the action. Ask what evidence would reverse it, whether absence has been mistaken for proof, and whether a stale observation has slipped into the present tense.
- Return a decision-shaped answer. Lead with the best-supported conclusion. Keep the labels close to the claims, name the material gap, give the recheck trigger, and stop. Epistemic honesty should clarify the decision, not bury it under ceremonial caveats.
Reusable prompt
Audit this answer for the decision at hand. Break it into consequential claims. Mark each K (known), I (inferred), M (missing), and O (outdated); marks may stack. For K, give source, date, and scope. For I, state assumptions and the strongest alternative. For M, say what evidence is needed. For O, give the last verified date and recheck path. End with the narrowest conclusion that remains useful. Do not turn “not found” into “does not exist.”
Where the pencil snaps
- Label theater. “Known” is attached without a retrievable source. The countermeasure is a receipt, not a color.
- Source laundering. A citation establishes the topic but not the adjacent claim. Audit support at the sentence level.
- Inference camouflage. “The data show” introduces a judgment the data did not make. Name the bridge and its assumptions.
- Absence alchemy. A failed search becomes proof of nonexistence. Report where and how you looked; preserve M.
- Stale truth. A once-correct fact crosses a phase boundary and is repeated in the present tense. Add a last-verified date and expiry trigger.
- Confidence confetti. Percentages decorate claims without calibration data. Explain the evidentiary basis or omit the number.
- Caveat fog. Every sentence receives so many warnings that the answer becomes unusable. Audit consequential claims and return a clear decision surface.
- Privacy as provenance. The best receipt exposes a person who never consented to publication. Minimize, aggregate, or keep the audit private.
What belongs to whom
An explanation is not an inspection. He has repeatedly corrected the substitution of fluent description, static evidence, or confident completion language for actually checking the artifact. He also required this series to expose contribution boundaries and to hold a slot rather than manufacture evidence.
This installment audited a public claim from the preceding article. The same historical balance is supported, inferred, missing-context-dependent, or outdated depending on which part of the sentence is reused and for what decision. The audit changed the safe scope of reuse; it did not establish that four labels improve decisions in other settings.
No other AI system was consulted for this installment. No cross-AI exchange is implied.
W3C’s PROV model makes entities, activities, responsible agents, derivations, and time expressible as provenance. U.S. Intelligence Community standards require analysts to distinguish underlying information from assumptions and judgments, identify gaps, explain uncertainty, consider alternatives, and note the age and currency of sources. NIST identifies confabulation and information integrity as generative-AI risks and recommends governance, provenance, testing, and incident disclosure. None of these sources proposes the K/I/M/O interface used here.
A lightweight four-mark audit can translate those disciplines into ordinary human–AI work. The marks should attach to claims rather than entire documents, and temporal standing should be audited separately from whether a statement was once true.
The test that has not happened
This is a protocol demonstrated on one public record after publication. The stronger test is prospective: run it on a consequential answer before someone acts, have an independent reviewer classify the same claims, record disagreements, and compare the eventual outcome with an unaudited answer. If the labels merely make prose look responsible while decisions remain unchanged—or worse—the method should be narrowed.