AI uses for records
Context graph temporal reasoning: testing who knew what, and when
By SourceX Editorial · Reviewed by Noah Loul ·
9 sections
Short answer
Represent each business fact as a dated, sourced claim with three clocks: when it was true in the business (valid time), when a system recorded it (record time) and when each person or agent could have seen it (exposure time), plus the authority and permission in force. Then test agents with as-of questions whose correct answer changes with the clock and the viewpoint the question names.
Agent run
Context graph temporal
Key takeaways
- Who-knew questions need three clocks: valid time, record time and exposure time; most graphs store only the first two.
- Delivery or access proves exposure, not knowledge; only a reply, approval or action shows acknowledgment.
- Each task item fixes an as-of moment and a viewpoint, and its gold answer cites edges valid and visible at that moment.
- Corrections invalidate rather than delete, erasures remove content but keep a tombstone, and expired history makes an item unknowable.
- Score answers, citations, leakage and permission separately, and report results per task family.
Who knew what, and when: which three clocks does a business fact need?
A business fact needs three clocks to answer a who-knew question: when the fact was true in the business (valid time), when a system first recorded it (record time), and when a particular person or agent could have seen it (exposure time). Most context graphs store the first, some store the second, and almost none store the third.
The three clocks disagree all the time in real records. A price change can be effective on the first of the month, keyed into the ERP two days later, and announced to the sales team a week earlier. Ask "what was the price on the second?" and the answer depends on which clock you mean. Ask "did the account manager know?" and none of the first two clocks can answer it.
Exposure is not knowledge. A delivered email, an assigned ticket or a meeting invitation shows that a fact reached someone's workspace. It does not show that they read it, understood it or acted on it. A task set that treats delivery as knowledge will reward agents for blaming people unfairly, so the gold answers below always state which kind of evidence they rest on.
| Clock | Question it answers | Typical evidence | Common mistake |
|---|---|---|---|
| Valid time | When was this true in the business? | Effective dates on contracts, price books, policies and org changes | Using the date a record was edited as the date the change took effect |
| Record time | When did a system of record hold this value? | Field history, audit logs, system-versioned tables, ingestion timestamps | Assuming the current row was always the value |
| Exposure time | When could this person or agent have seen it? | Message received times, ticket assignment, meeting attendance, access logs | Treating delivery or access as proof that the person knew |
| Acknowledgment | When did the person show they knew? | A reply, an approval, a signed form, an action that used the new value | Inferring acknowledgment from silence |
Why does a validity window alone fail a who-knew question?
A validity window answers what was true; it says nothing about who had access to that truth. Current temporal graph tools are clear about which clocks they keep, and reading their definitions shows where the gap sits.
Graphiti, an open-source framework for temporal context graphs, gives each fact a validity window and says that when information changes, old facts are invalidated rather than deleted, so you can query what was true at any point. Its edge model separates valid_at (when the fact became true), invalid_at (when it stopped being true), expired_at (when the edge itself was invalidated) and reference_time (the timestamp of the episode that produced the edge). The accompanying Zep paper describes this as a bi-temporal model: one timeline for the order of events and one for the order in which data was ingested.
Relational systems keep the record clock. In SQL Server, a system-versioned temporal table has two period columns that the database fills from the start time of the modifying transaction, in UTC, and every prior row version moves to a history table. A FOR SYSTEM_TIME AS OF query returns the rows that were current in the database at that moment. That is record time: the moment the system held the value, not the moment the business change took effect and not the moment anyone was told.
So a who-knew question needs a third kind of edge that neither model supplies by default: an exposure edge linking a fact, or the message that carried it, to a person or agent, with its own timestamp, its own evidence and its own strength.
What does a time-scoped context graph task set contain?
Each item in the task set is one question asked at one moment from one point of view, with a gold answer that cites the edges valid at that moment and visible to that viewpoint. The schema below is the asset; a team can copy it as is and fill it from its own records.
At a glance
What does a time-scoped context graph task set contain
| Field | What it holds | Illustrative value | Why it matters |
|---|---|---|---|
| as_of and viewpoint | The moment of the question and whose knowledge is being asked about | 2026-03-02 10:15 UTC; account manager AM-17 | The same fact gets different answers at different moments and for different people |
| clock_tested | Valid, record, exposure or acknowledgment | exposure | Stops graders from accepting a right answer to the wrong question |
| gold_answer and answer_type | The answer plus whether it is a value, yes, no or unknowable | yes, exposure only (delivered, read state unknown) | Unknowable is a legitimate gold answer when the history is missing |
| supporting_edges | Edge ids with valid, record and exposure times and the source record for each | price edge P-2; email edge E-9 to AM-17 | A correct answer must cite these, not a later edge |
| distractor_edge | A plausible edge that is wrong for this as_of or viewpoint | the ERP price row recorded on 2026-03-03 | Measures whether the agent leaks future records into the past |
| authority and permission_at_as_of | Which system or role may assert the fact, and whether the viewpoint could see it then | price book owner: pricing team; AM-17 on sales list | An answer built from a record the viewpoint could not see is wrong even if true |
| correction_lineage | Revision links to earlier and later versions of the fact | P-1 revised by P-2 | Lets the grader tell a correction from a new fact |
| erasure_status | Whether any supporting content was later erased, with a content-free tombstone | none | The task must still be gradable after lawful deletion |
| outcome_link | The record that shows what happened next | quote Q-44 revised; customer approved | Separates the decision from a plausible final message |
Which task families make temporal shortcuts fail?
Build the set from families that each defeat one shortcut, and balance them so a model cannot score well by always answering from the latest snapshot. The families below are the minimum; a team working in one domain will add its own.
| Task family | Question form | Shortcut it defeats | What a correct answer cites |
|---|---|---|---|
| As-of truth | What was X on date D? | Answering from the current value | The edge whose valid window contains D |
| As-recorded | What did system S show on date D? | Confusing effective dates with entry dates | The row version current in S at D |
| As-known-by | Could person P have known X by date D? | Treating whatever is true as known to everyone | An exposure edge to P before D, or a stated absence |
| Late correction | Was the earlier answer wrong at the time, or only in hindsight? | Rewriting the past with a correction | Both versions, their record times and the revision link |
| Permission at time | Could agent A use X on date D? | Ignoring access changes | The permission edge valid at D for A |
| Erased fact | What can be said about X after erasure? | Reconstructing erased content from derived facts | The tombstone and the rule that removed the content |
| Unknowable | Who knew X on a date older than retained history? | Inventing an answer where the evidence has expired | The retention boundary and the answer unknowable |
Worked case: did the account manager know about the price change before she quoted? (illustrative)
This case is illustrative; the people, dates and percentages are invented to show the structure. Starting state: a distributor raises list prices on one product line by 6 percent, effective March 1. The pricing team emails the sales distribution list on February 20. The price book in the ERP is updated on March 3, two days after the effective date. Account manager AM-17 issues quote Q-44 on March 2 at the old price.
Linked evidence: the pricing memo (valid time March 1, authority pricing team), the email message with a sent and received time on February 20 and AM-17 among the recipients, the ERP field history showing the price change recorded March 3, and the quote record with its creation time on March 2. The message metadata gives received time, but the read flag is a current yes-or-no state with no time attached, so it cannot show when, or whether, she read it before March 2.
Applicable rule and decision: the item asks, as of March 2 from AM-17's viewpoint, whether she could have known. The gold answer is "exposed, acknowledgment unknown": the memo reached her mailbox ten days earlier, but nothing she did shows she read it, and the system she quoted from still showed the old price. The decision that follows is a process finding about the ERP lag, not a finding about the person.
Outcome check: the quote record shows a revision on March 4 at the new price, and the customer's approval references the revised quote. That outcome link is what lets a grader confirm the decision held, rather than trusting a closing message in a thread.
Counterexample: an agent working from a current snapshot sees the new price, the email in her mailbox and an old-price quote, and concludes she ignored the change. It used today's price book, treated delivery as knowledge and skipped the record clock. Each error maps to a task family above, which is why the set needs all of them.
Correction, expiry or erasure: which one changes the answer to an as-of question?
A correction changes later answers but must not change what the system recorded at the time; an expiry ends the valid window; an erasure removes the content itself. An agent that handles all three the same way will either rewrite history or keep data it no longer may hold.
Corrections should invalidate, not delete. Record the corrected fact as a revision of the earlier one, keep both record times, and let as-recorded questions still return the earlier value. W3C PROV-O gives a vocabulary for this: generatedAtTime and invalidatedAtTime bound an entity's existence, wasAttributedTo ties it to the responsible agent, and wasRevisionOf links a revised version to the one it replaced.
Erasure is different. Under GDPR Article 17, a data subject can obtain erasure of personal data concerning them without undue delay when one of the listed grounds applies. A task set built from records that include personal data must be able to drop content while keeping a content-free tombstone, so that the item either becomes an erased-fact item or is retired. Whether erasure applies to a given record is a question for counsel on the specific facts; the task set only has to support the outcome.
Retention limits create a fourth case: the history simply is not there. Salesforce documents that field history is retained for 18 months in the application and up to 24 months through the API unless the org buys a longer audit trail. An as-of question older than that boundary, asked of a source without an archived copy, has the gold answer unknowable.
Which records prove what a person or agent could have known?
Rank evidence by what it actually proves, and record the rank in the item so the grader knows how strong the gold answer is. Delivery evidence proves exposure at best; only an action or a reply proves acknowledgment.
Message metadata is the most common exposure evidence. Microsoft Graph, for example, records sentDateTime and receivedDateTime in UTC for each message, plus an isRead flag. The flag is the current state, so a later read cannot be dated from it, and it can be toggled by the user. Treat it as weak evidence and never as a timestamped read.
| Evidence | What it proves | What it does not prove | Strength |
|---|---|---|---|
| System-of-record field history | The value the system held and when it changed | That any person looked at the field | Strong for record time |
| Message received time | The fact reached the person's mailbox | That it was opened or understood | Exposure only |
| Current read flag | Someone marked it read at some point | When it was read, or by whom on a shared mailbox | Weak |
| Ticket or task assignment | The person was responsible for the item from that time | That they opened the linked attachment | Exposure, moderate |
| Reply, approval or signed form | The person engaged with the content | Full understanding of every term | Acknowledgment |
| Action using the new value | The person or agent applied the fact | Why they applied it | Strong acknowledgment |
How should answers to as-of questions be scored?
Score the answer and its evidence separately, and fail any answer that cites an edge the viewpoint could not have had at the as-of moment. A right value supported by a future record is a leak, not a success.
Use four checks per item. Answer correctness under the clock the item tests. Citation correctness: every cited edge has a valid window, record time and exposure time consistent with the as_of. Leakage: any cited record with a record time after as_of fails an as-recorded or as-known-by item. Permission: an answer that relies on a record outside the viewpoint's access at the time fails regardless of truth.
Give explicit credit for abstention on unknowable items and penalize confident answers there. Report results per task family, not as one average, because a model can excel at as-of truth and still fail every as-known-by item. Version the task set and date each release, since a source schema change or a retention change can silently alter which items are answerable.
Which business records can carry all three clocks?
A temporal task set is only as good as the history behind it: field histories that were switched on before the period you test, message metadata that survived export, and effective dates kept separately from edit dates. Synthetic records rarely reproduce the lags and missing reads that make these items hard, which is why teams look for real operating records.
An AI team can ask SourceX to look for records like these at U.S. businesses, stating scope, volume, format, timing and licensing needs up front, including retained field history and message metadata. Nothing reaches the buyer before the supplying company has signed off on scope and terms, and what does arrive is de-identified. Access to records, permission under privacy law and a license to use them for AI are separate questions, and each one is settled in the agreement rather than assumed.
Frequently asked questions
What is temporal reasoning in a context graph?
It is answering a question at a specific moment using only the facts that were true, recorded or visible at that moment. A temporal context graph supports this by storing validity windows and record times on its edges. Who-knew questions also need exposure edges that link facts to the people or agents who could have seen them.
What is the difference between valid time and transaction time?
Valid time is when a fact was true in the business, such as a contract's effective date. Transaction time, also called record time, is when a database stored that value. SQL Server system-versioned tables, for example, keep transaction time automatically, while business valid time has to be modeled separately.
Does a read receipt or read flag prove that someone knew a fact?
No. A received timestamp proves the message reached a mailbox, and a read flag in systems such as Microsoft Graph is a current yes-or-no state without a time. Treat both as exposure evidence and look for a reply, an approval or an action that used the fact before calling it acknowledgment.
How should a context graph handle a fact that was later erased?
Remove the content and keep a content-free tombstone that records that an erasure happened and which rule required it. Items that depended on the content become erased-fact items or are retired. Derived facts must not be used to rebuild what was removed.
How is a who-knew task set different from a point-in-time training dataset?
A point-in-time dataset prevents future values from leaking into features for one prediction. A who-knew task set adds viewpoints: it asks what a specific person or agent could have known at a moment and grades the evidence chain, including permissions and corrections.
Sources
- Graphiti describes a context graph whose facts carry validity windows, with old facts invalidated rather than deleted so past states remain queryable. Source
- Graphiti entity edges carry valid_at, invalid_at, expired_at and reference_time fields with distinct meanings. Source
- The Zep paper describes a bi-temporal model with one timeline for event order and one for the order of data ingestion. Source
- SQL Server system-versioned temporal tables keep period columns set from the transaction start time in UTC and move prior row versions to a history table. Source
- Microsoft Graph messages record sentDateTime and receivedDateTime in UTC and an isRead Boolean showing whether the message has been read. Source
- PROV-O defines generatedAtTime and invalidatedAtTime to bound an entity's existence, wasAttributedTo for attribution and wasRevisionOf for revised versions. Source
- GDPR Article 17 gives data subjects the right to obtain erasure of personal data without undue delay when a listed ground applies. Source
- Salesforce retains field history for 18 months in the application and up to 24 months through the API without Field Audit Trail. Source
Tell us what data you need
Share scope, volume, format, timing and licensing needs. Datasets are sourced on request.