AnthropicClaude Certified Architect – Foundations2026-105.6 Preserve information provenance and handle uncertainty in multi-source synthesis

Compare findings against their observation periods

Completed by the knowledge check

Learning objective

Attach publication and collection dates so temporal changes are not mistaken for conflicting evidence.

Core idea

Two credible sources can report different values because they describe different moments. Before declaring a disagreement, determine when each source appeared and when its evidence was collected. A report published this year can describe a survey conducted several years earlier.

For a multi-agent research system, dates belong in the evidence passed to synthesis. They must survive compression with the source and claim, rather than remaining in a specialist’s private notes. See 5.6.1 for preserving those relationships across handoffs. See 5.1.6 for the date fields in the upstream finding record; this lesson owns how synthesis interprets them.

How it works

Separate the clocks

Use explicit application-defined fields for the dates your report needs. These are fields in your evidence contract, not automatic date metadata supplied by Claude. The collection_period and reference_period fields refine 5.1.6’s observation_period by distinguishing when observations were gathered from the period the reported value describes. A useful distinction is:

Field	Meaning	Comparison error it prevents
`published_at`	When the source was published	Treating a later publication as a later measurement
`collection_period`	When observations were gathered	Comparing surveys from different periods as simultaneous
`reference_period`	The period a reported value describes	Confusing a release date with the year of a financial result
`retrieved_at`	When your system accessed the source	Treating access time as evidence that its content is up to date

Some sources provide a single collection date; others provide an interval or only a year. Preserve that precision. If a source gives only a reporting year, do not invent a day. If it gives no measurement period, return an explicit unknown with the source location inspected. Missing temporal information is a limitation, not permission to use the publication date as a substitute.

Distinguish an observation from a forecast as well. A forecast issued in one year about another year needs both the issue date and the target period. A forecast and a measured outcome can share a reference year without describing the same evidential state.

Require temporal metadata in the specialist’s output

Use AgentDefinition.prompt in the Agent SDK to instruct the research specialist to return dates with the finding. Specify that dates come from the provided source and that unavailable dates remain unknown. Anthropic’s prompting guidance recommends separating document content and metadata with <document>, <document_content>, and <source> tags when presenting multiple documents. You can include the extracted time fields alongside that source metadata.

The same prompting guidance recommends extracting relevant quotes before analyzing long documents. Include the passage establishing the date or measurement period, especially when a cover date differs from a table’s reporting period. A reviewer must be able to check which clock a date describes.

If you use Messages API document citations, context can hold metadata as text or stringified JSON. However, title and context are not themselves citable content; citations point into the document’s source content. Put the underlying dated passage in the source and retain the date extraction’s evidence reference. Metadata helps interpretation without becoming a replacement for evidence.

Test comparability before calling a conflict

Ask whether the records describe the same measure, population, geography, and period. If only the periods differ and the measurement basis is comparable, the values can describe temporal change. If the periods match but the definitions differ, label the methodological distinction. Dates alone do not make incomparable measures into a time series.

A later publication can also revise an earlier estimate for the same period. Preserve both publication dates and the source’s revision statement. The coordinator can decide to use the revised estimate for the requested question while retaining the superseded value’s history. Do not infer revision merely because one publication appeared later.

When the relevant date is missing, avoid silently selecting the newest-looking source. Anthropic’s hallucination guidance supports explicitly admitting uncertainty and restricting analysis to supplied documents. Apply that instruction to temporal claims: state that comparability remains unresolved, and identify the date or definition needed to resolve it.

Only a difference that survives the basis, period and revision checks is a conflict.

Key distinctions and trade-offs

Publication recency versus measurement recency: A fresh article can quote an old observation. Rank evidence by the period relevant to the question, not by the apparent freshness of its wrapper.

Temporal change versus contradiction: Different years can explain different values. They establish a trend only when the measurement basis is sufficiently comparable. A difference between registrations and shipments remains a definition issue even after dates are supplied.

Retrieval timestamp versus source update: Accessing a page establishes when the system collected it. It does not establish when the page’s figures changed or whether a statement remains valid.

More date fields make the handoff larger, but collapsing distinct clocks can produce a confident false conflict. Keep the clocks that affect the decision and preserve unknowns rather than manufacturing precision. See 5.6.3 for genuine incompatible values that remain after comparison.

Architecture / decision scenario

A coordinator researches a fictional software market. Specialist A finds a report published in March 2025 stating a market value of USD 1.9 billion for 2021. Specialist B finds a report published in January 2026 stating USD 4.6 billion for 2024. Both examples are hypothetical, not market estimates.

The specialists return publication dates, reporting years, source passages, and definitions. The coordinator checks that both reports measure the same region and revenue category. Synthesis presents the values as observations from different years rather than a disputed estimate for one year. It does not label the first value a 2025 result because of the report’s publication date.

A third source is retrieved in October 2026 but gives no measurement period. The coordinator retains the retrieval time and marks the period unknown. It excludes that record from the trend comparison until a targeted follow-up finds the underlying dated table. The report can still describe the supported historical change while making the third source’s temporal limitation visible.

Common exam traps

Choosing the most recently published source without checking the data period.
Treating a retrieval timestamp as the date of measurement.
Calling different years contradictory before testing comparability.
Filling an unknown collection date with the publication date.
Treating a forecast for a year as an observation from that year.
Assuming a later publication automatically supersedes an earlier estimate.

What to remember

Attach publication and collection dates to the finding, preserving each date’s meaning and precision. Compare reference periods and methods before deciding whether values disagree. A temporal difference can explain a change; an unknown date should remain a visible limitation rather than an invented fact.