Skip to content

Systems and records

Missing timestamps dataset limitations: the claims each gap rules out

By SourceX Editorial · Updated

9 sections

Short answer

Missing timestamps limit what a dataset can claim, not only what it can compute. Before saying records support a temporal holdout, durations, step order, trends or a post-cutoff evaluation, check which timestamp each claim needs, how complete it is in every stratum, and whether gaps follow period, system or outcome. Where coverage fails, narrow the claim to the records that carry the time, and publish the gap with the dataset.

System export

Missing timestamps dataset

This page covers missing timestamps dataset limitations: the claims each gap rules out.Illustrative

Key takeaways

  • Each temporal claim needs a specific timestamp at a specific resolution, so a dataset is only as dated as the field a given claim depends on.
  • The chronology resolution ladder runs from unordered records to full event histories, and each rung unlocks a defined set of claims.
  • Measure completeness per stratum: an overall dated share can hide an undated migration era or an undated outcome class.
  • Gaps that follow the outcome, such as open cases without close dates, bias durations and cannot be treated as random.
  • Post-cutoff evaluation rests on trustworthy creation dates, which is why LiveCodeBench filters problems by release date.
  • The usual fix is a narrower claim stated with its denominator, not a larger dataset.

What do missing timestamps stop a dataset from claiming?

They stop any claim whose truth depends on when something happened: that a test set comes after the training data, that cases took a certain time, that steps happened in a certain order, that a rate changed, or that evaluation records post-date a model's training cutoff. Content claims, such as which kinds of cases a dataset contains, usually survive.

That distinction matters because most writing on timestamp gaps is about repair. The guide to missing timestamps and broken threads sorts damage into fixable problems and deal breakers, and the timestamp audit guide checks validity, order and time zones. This page starts after both: once the repairable gaps are repaired and the audit is done, some records still lack the time a claim needs, and the question becomes what you can still say.

The answer is decided claim by claim. A dataset can be fully dated for one purpose and unusable for another, because a creation date supports a date-based holdout but not a duration, and a duration needs both ends of a case recorded at the same resolution. Writing the claim first and then checking its field keeps a limitations statement honest.

The chronology resolution ladder: how much time does each record carry?

Place every record, and then every stratum, on one of five rungs according to the finest time information it reliably carries. A rung is earned only when the field is original or reconstructed from a documented source; an import date that replaced the original value counts as no date at all.

At a glance

The chronology resolution ladder: how much time does each record carry

Each rung unlocks the claims of the rungs below it plus its own. The rung that matters for a claim is the lowest one held across all the records the claim covers, so a test set with one undated stratum sits on rung 0 for that stratum's records, however well dated the rest are. The five rungs are listed below:

  • Rung 0, unordered: no usable time. Supports claims about what kinds of cases exist and how they are labeled, and nothing about order, recency or change.
  • Rung 1, period known: the record can be placed in a month, quarter or year, for example from the export batch it arrived in. Supports coarse era splits and period counts when coverage is stable across periods.
  • Rung 2, dated: a calendar date per record without a reliable time of day or offset. Supports date-based holdouts with a buffer and daily volumes, but not order within a day or durations shorter than a day.
  • Rung 3, timestamped: a full instant with a known offset, or a UTC instant. Supports order within a case and durations; claims about local working hours also need the local offset.
  • Rung 4, event history: a timestamped log of status and field changes. Supports time in each state, service-level compliance and which rule was in force when a decision was made.
A five-step staircase from unordered records to full event histories, each step labeled with the temporal claims it adds: content only, era splits, date holdouts, order and durations, and rule-in-force checks.
The chronology resolution ladder. Editorial framework; no measured data.

Which claims need which rung? The temporal claims ledger

Every common temporal claim has a minimum rung, a field it depends on and a typical way of failing. When a claim fails, a narrower version often survives, and that narrower version is what belongs in the dataset card.

Two rows fail silently. Durations computed only from cases with both an open and a close time describe the cases that finished, not the cases that arrived. Local working-hour patterns need the local offset, and a UTC stamp may not carry it: RFC 3339 reserved the offset -00:00 for a time whose UTC value is known but whose local offset is not, and RFC 9557 (2024) says Z should now be used in its place, so a timestamp ending in Z gives the instant without promising the local clock time.

Read the ledger from the claim you want to make, not from the fields you happen to have:

Which claims need which rung? The temporal claims ledger
Claim you want to makeMinimum rung and fieldHow missing time breaks itNarrower claim that may survive
Evaluation records post-date the model's training cutoffRung 2: original creation date on every evaluation recordUndated or import-dated records can predate the cutoff without anyone knowingPost-cutoff for the dated subset, with the undated count excluded and reported
Trained on the past, tested on what came laterRung 2: decision date on both sides, plus an outcome date for the gapMissing outcome dates let late results from the test period leak into trainingA holdout on strata where both dates exist, with excluded strata named
Cases took this long to resolveRung 3: open and close instants at one resolutionOpen, abandoned or migrated cases lack close times, so durations describe only finished casesDurations of closed cases, with open cases reported as censored rather than dropped
Steps happened in this orderRung 3 on each event, or a reliable sequence numberTies from batch imports or date-only events hide the order inside a dayOrder between days only, with the share of tied events stated
The rule in force shaped this decisionRung 4: decision time plus rule effective datesTruncated field history removes the status changes that fix the decision timeThe claim for the period whose history is complete
The rate changed over timeRung 1 with stable completeness across periodsIf older periods are less dated, the trend can reflect missingness rather than changeA trend within periods whose dated share is similar
Activity peaks in local working hoursRung 3 with the local offsetA UTC or Z stamp may not record the local offsetPatterns stated in UTC, or for systems with a documented time zone
The data represents current practiceRung 2 with recent periods coveredMissing or undated recent records leave the newest workflow outRepresents the stated period, with its end date

Is the gap random, or does it follow the outcome?

Assume it follows the outcome until a test says otherwise. Missing time is rarely spread evenly: open cases lack close dates by definition, migrated records often carry an import date instead of their own, and abandoned cases may never receive a final status.

Donald Rubin set out the statistical stakes in Biometrika in 1976: ignoring the process that caused data to be missing is justified only under stated conditions, the best known being that the data are missing at random. When missingness depends on the very thing being measured, such as whether a case was resolved, estimates computed from complete records are biased, and adding records does not remove the bias.

System limits create structured gaps as well. Salesforce's help documentation says that with Field Audit Trail turned off, field history is kept for up to 18 months, or 24 months when read through the API. An org exported years later can carry full creation dates but status histories only for its most recent stretch, so rung 4 claims hold for the newest period and fail for everything before it.

Three comparisons separate a random gap from a structured one: the dated share across outcome classes, across systems or migration eras, and across periods. A difference between any two groups that is larger than the sampling interval means the gap is structured, and every claim touching those groups needs the narrower wording from the ledger.

The sampling worksheet: measuring completeness where it can differ

Measure completeness on a stratified sample of cases, not as one overall share of rows. The case is the unit because claims are made about cases; a row count lets one long case with hundreds of events outweigh many short ones.

At a glance

The sampling worksheet: measuring completeness where it can differ

Sample size follows the precision each stratum needs. With about 96 cases in a stratum, a completeness share near 50 percent carries a 95 percent margin of roughly 10 percentage points, and quadrupling the sample halves the margin. Small strata that matter to the claim, such as abandoned cases, should be sampled to that size even if it means oversampling and reweighting.

Record each line of the worksheet for every sampled case:

  • Sampling unit and frame: case ID, the export it came from and the frozen list it was drawn from.
  • Stratum: system, record type, period, migration era and outcome status (resolved, open, abandoned or unknown).
  • Each timestamp field the claim needs: present, absent, import date or reconstructed, with the source of any reconstruction.
  • Resolution and offset: date only, full instant, offset known, or UTC with the local offset unknown.
  • Ties: whether the case's events share a timestamp, and how many do.
  • History depth: the earliest status change on record for the case, to detect truncated histories.
  • Rung: the highest rung the case reliably holds for the claim under review.
  • Result: completeness by stratum with intervals, and the difference in completeness between outcome classes.

Worked example (illustrative): a helpdesk export that is mostly dated

Illustrative: a team samples 400 cases from a helpdesk export, 100 from each of four strata, to decide whether it supports a time-to-resolution claim and a post-cutoff evaluation set. Overall, 341 of the 400 sampled cases carry an original creation date, about 85 percent, which sounds sufficient.

The strata tell a different story. Records from before a system migration lost their dates where the import overwrote them, and abandoned or merged cases rarely received a close time. A duration computed from the 177 sampled cases with both times would describe mostly resolved work after the migration and almost none of the abandoned cases, which are the cases a support model most needs to handle.

The decision follows from the strata, not from the 85 percent. The team can claim a post-cutoff evaluation set drawn from the post-migration stratum, durations for resolved cases with open cases reported as censored, and nothing time-based about abandoned cases. At 100 cases per stratum each share carries a margin of about plus or minus 10 points or less, so the gap between 98 and 71 is real while the gap between 98 and 96 is not. The figures are invented to show the method and describe no real company or export; SourceX has not published measured results from this worksheet.

Worked example (illustrative): a helpdesk export that is mostly dated
StratumSampled casesOriginal creation dateBoth open and close timesWhat the stratum supports
Resolved after migration1009897Durations and a post-cutoff test set
Resolved before migration1007160Period counts; durations only for the 60, flagged
Open at export100960Arrival counts; durations only as censored cases
Abandoned or merged1007620Content claims; no duration or outcome-time claim

Large record counts do not rescue a dated subset

The strongest objection is volume: with millions of records, even a large undated share leaves plenty to work with. That holds only if the dated records are a fair sample of the whole, and the reasons records lose their time usually make them unfair.

Three mechanisms hide behind a big count. Repeated cases, such as reopened tickets or duplicated threads, inflate the number of dated records without adding independent cases. Selection, such as a migration that kept dates only for resolved cases, makes the dated subset systematically different. Unobserved failures, cases that ended without any recorded outcome, are the least likely to carry the timestamps that would show how long they took.

Volume does help in one way: it lets you set aside an undated stratum and still keep enough records for a narrower claim. The claim has to shrink with the data, though. An illustrative export of two million records whose undated share sits in its first three years supports claims about its later years, stated as such, and nothing about the whole span.

Why post-cutoff and contamination claims need trustworthy creation dates

Because the claim is about dates. A buyer who wants evaluation records the model under test could not have seen needs to know that each record was created after the model's training cutoff, and an undated or import-dated record cannot show that.

Public benchmarks treat the date as the control. LiveCodeBench tags each problem with its release date so that newer models can be scored only on problems released after their cutoff, and its authors report that some models did noticeably worse on problems released after their cutoff dates, which they read as a sign of contamination on older problems. GSM1k reaches the same concern without dates: on fresh problems written to match GSM8K, the current version of the paper reports accuracy drops of up to 8 percent for some model families.

Replayed business tasks add a further requirement. Anthropic's engineering guide to agent evaluations treats each task as one test whose inputs and pass criteria are fixed in advance. Replaying a historical case faithfully means the defined inputs are what the record showed at that moment, which again requires knowing when each field was written, and a record that cannot be placed in time cannot be replayed as it was.

Writing the limitations statement, and asking for chronology in a data request

State limitations as claims with denominators: which claims the dataset supports, for which strata, at which rung, and which it does not. In the illustrative example, 'durations are reported for 177 of 400 sampled cases, and abandoned cases are excluded from all time-based claims' is useful; 'some timestamps are missing' is not.

When buying records, put the rung in the request. SourceX keeps no shelf of ready datasets; it looks for records at U.S. businesses once a buyer asks, and the request-data form is where the buyer sets out the scope, time span, format and licensing terms it needs. A request that names the claims the buyer intends to make, the rung each needs and an acceptance test, such as original creation dates on a stated share of sampled cases in every stratum, gives a supplier a concrete target.

SourceX's rule of thumb is that buyers favor several years of steady history from a single system, usually at least two, with each deal judged on its own facts. Steady is the word that matters: two years of records whose dated share collapses before a migration support fewer claims than their record count suggests.

Frequently asked questions

Can a dataset with missing timestamps still be used for AI training?

Yes, for purposes that do not depend on time. Undated records can still show what kinds of cases occur and how they are handled. What they cannot support is any claim about order, duration, recency or change, so a training set may include them while an evaluation set that claims a temporal split should not.

What share of missing timestamps is acceptable?

There is no universal threshold, because the answer depends on the claim and on where the gaps fall. A small undated share spread at random across strata may be harmless, while the same share concentrated in one outcome class can bias every duration. Set the threshold per claim and per stratum, and write it into the acceptance test.

Should missing timestamps be imputed?

Only from a documented source, in a separate field, and only for claims that tolerate the added uncertainty. A close time bracketed by a linked invoice may support a day-level duration. A value filled in from averages supports nothing, because it manufactures the pattern a time-based claim is supposed to test.

How do missing close dates affect time-to-resolution figures?

They usually make resolution look faster. Open and abandoned cases lack close dates, and dropping them leaves the cases that finished, which tend to be the quicker ones. Report closed-case durations with open cases counted as censored rather than dropped.

Does a UTC timestamp tell you the local time of an event?

Not necessarily. RFC 3339 used the offset -00:00 for a UTC time whose local offset is unknown, and RFC 9557 now says Z should be used instead, so a Z stamp gives the instant without promising the local clock time. Claims about local working hours need the offset or a documented system time zone.

Sources

  • With Field Audit Trail turned off, Salesforce keeps field history for up to 18 months, and up to 24 months through the API. Source
  • RFC 3339 used the offset -00:00 for a time whose UTC value is known but whose local offset is unknown, distinct from Z or +00:00. Source
  • RFC 9557 (2024) updates RFC 3339 so that Z should be used for a UTC time with an unknown local offset. Source
  • LiveCodeBench tags problems with release dates so newer models can be evaluated only on problems released after their cutoff; some models did worse on post-cutoff problems. Source
  • On GSM1k, fresh problems matched to GSM8K, the current version reports accuracy drops of up to 8 percent for some model families. Source
  • Rubin (Biometrika, 1976): ignoring the process that causes missing data is justified only under stated conditions such as data missing at random. Source
  • Anthropic's engineering guide to agent evals describes a task as a single test with defined inputs and success criteria. Source

Tell us what data you need

Share scope, volume, format, timing and licensing needs. Datasets are sourced on request.

See if your company qualifies