AI research time scope: separate known then from known now
An AI research answer needs an explicit information cutoff and separate dates for events, public disclosure, retrieval, and analysis. Without those distinctions, a memo can accidentally use later knowledge to explain an earlier decision. A timestamp in a filename is not enough to establish what was knowable at the time.
The W3C Date and Time Formats note describes a limited set of date-time representations and timezone notation. It is a useful formatting reference, not a guarantee of source timing accuracy. The research workflow still needs to establish what each timestamp means.
Create a four-clock record
Record when the underlying event occurred, when relevant information became public, when the researcher retrieved it, and the cutoff for the question. These clocks can diverge substantially. A quarterly period end is not the same as the filing date that made the numbers available.
Specify the timezone and precision of each field. If a source provides only a calendar date, do not invent an exact publication hour. A narrow historical question may remain unresolved if its answer depends on an unavailable intraday timestamp.
| Clock | Question |
|---|---|
| Event time | When did the underlying activity happen? |
| Public time | When was this information disclosed? |
| Retrieval time | When did the research process access it? |
| Question cutoff | What information is eligible for the answer? |
| Revision time | Did a later correction change the record? |
Keep the original source date alongside any normalized date. Conversion into a common timezone should not erase uncertainty or transform a date-only field into fabricated precision.
Hypothetical earnings cutoff
Assume a fictional quarter ends on March 31 and its results are published on May 10. A research question asks what was publicly known on April 15. The quarter’s results concern a period before the cutoff, but the May publication is later information and should not enter the eligible evidence set as if it was already available.
The memo can separately discuss what became known later, clearly labeled. That comparison may be useful for learning, but it must not rewrite the earlier information environment. If the AI is asked to analyze the April decision, give it the eligible packet and explicitly segregate later documents.
Now suppose a correction is published in June. A current factual summary may use the corrected figure, while a reconstruction of the April information set needs the version available then. These are different questions rather than conflicting standards. The scenario is hypothetical and does not describe a real issuer.
Filter sources before synthesis
Ask AI to list each source’s public date and eligibility status before generating the narrative. Review the list manually where timing affects the conclusion. A web page’s current update date may not establish when a particular statement first appeared, so distinguish first publication from later revision when evidence allows.
For continuously updated datasets, preserve a version or snapshot. A current historical series can incorporate revisions unavailable to an earlier researcher. A time-based analysis should say whether it uses the original release or today’s revised history.
When a source has uncertain timing, assign a status such as eligible, later information, or publication time unresolved. Do not let the model infer eligibility merely because an event date is early. The clock controlling eligibility is the one defined by the research question.
Keep hindsight visible
Write separate sections for the evidence available by the cutoff and the later outcome. Compare them only after each is established. This allows useful reflection without turning knowledge of the result into proof that the original decision should have been obvious.
The method cannot establish exact historical availability when archives or release records are missing. Narrow the conclusion in that case. Common mistakes include using period end as disclosure date, mixing timezones, accepting current revised data as original release data, and reporting a precise hour from a date-only source. A defensible memo states its cutoff and names the evidence excluded for timing reasons.
FAQ
Can a document about an earlier quarter be used before its publication?
Not in a public-information reconstruction unless evidence establishes the relevant information was already available through another eligible source.
Should later corrections always be ignored?
No. Use them for current factual understanding, while keeping them separate when reconstructing what was known at an earlier cutoff.