Systems and records
Expert decision time pressure dataset: capture what the deadline ruled out
By SourceX Editorial · Updated
10 sections
Short answer
Record the clock alongside the choice. For each expert decision, capture when the request arrived, which deadline applied and where it came from, the working time actually available, every option considered, the options the deadline ruled out and why, and what the expert would have chosen with more time. Without those fields a rushed but sensible call looks like a preference, and a model learns the wrong lesson.
System export
Expert decision time
Key takeaways
- A decision record without its deadline makes triage look like preference, and a model trained on it learns the wrong rule.
- Clock time, working time and felt pressure are separate fields; SLA clocks pause and calendars skip hours.
- Record options the deadline eliminated with a reason code, and keep impossible options apart from risky ones.
- Ask the expert, within the hour, what they would have chosen with no deadline and what fact would have changed the call.
- Only four fields need the expert's hand; the rest come from system logs and the outcome review.
How do you capture the effect of time pressure in an expert decision time pressure dataset?
Treat the deadline as part of the decision, not as background. A dispatcher who sends the nearest technician because the customer's window closes in two hours is making a different decision from one who would send the specialist with a day to spare, even if both are experts and both are right for their circumstances. A dataset that keeps only the choice teaches a model that the nearest technician is simply preferred.
The research on deadlines points the same way. A systematic review of time pressure in software engineering by Kuutila and colleagues reports that most high-quality studies found productivity rising and quality falling under time pressure. Output under a deadline is a mix of judgment and triage, and a training record has to say how much of each it contains.
The fix is a set of fields captured at decision time: the clock, the source of the deadline, the options that were feasible inside it, and the expert's own view of what they would have done without it. The specification later on this page lists them.
Where does a deadline come from in ordinary business records?
Usually from a system clock that measures something narrower than the time the expert actually had. Help desks, service management tools, contracts and customer emails each record deadlines differently, and each leaves something out. Know which one you are reading before you call a decision rushed.
- Zendesk documents that requester wait time counts time in New, Open and On-hold and pauses while a ticket is Pending, while total resolution time does not pause for Pending.
- Jira Service Management SLAs start counting when a start condition is met, can pause on optional conditions such as waiting for the customer, and stop as soon as any stop condition is met.
| Deadline source | Where it is recorded | What the record leaves out |
|---|---|---|
| Help desk SLA target | SLA policy and per-ticket target status, for example in Zendesk | Paused time: Zendesk's requester wait time stops while a ticket is Pending, and targets can run on business hours |
| Service management SLA goal | Start, pause and stop conditions on an SLA, for example in Jira Service Management | Which pause condition was active when the expert decided |
| Contract or order due date | Promise date on a sales order or project milestone in the ERP or CRM | Whether the date was firm, negotiable or already missed |
| Customer-stated deadline | Free text in an email, chat or call note | Whether the expert saw it before deciding |
| Shift or dispatch window | Scheduling system appointment window or shift roster | Workload from other jobs competing for the same window |
| Self-imposed target | Rarely recorded at all | Everything; ask the expert directly at capture time |
Are clock time, working time and felt pressure the same variable?
No, and the specification keeps them apart. Clock time is the gap between the request and the deadline. Working time is the part of that gap the expert could actually spend on this decision after calendars, pauses and other work. Felt pressure is the expert's own rating of how rushed the decision was.
Felt pressure has a well-tested wording. The NASA Task Load Index includes a temporal demand item that asks how much time pressure the person felt due to the rate or pace at which the tasks occurred. Asking a version of that single item right after the decision costs seconds and gives you a per-decision rating that clock fields cannot provide.
Two common shortcuts measure something else. O*NET's Time Pressure work context reports how often an occupation requires meeting strict deadlines, on a scale from never to every day; it describes the job, not a decision. Annotation tools report time on task, but Label Studio documents that its lead time includes idle time and is cumulative across sessions, so it overstates how long the expert actually deliberated.
The deadline constrained decision collection specification
Capture these fields for every decision you include. Fields marked at decision time should be written while the case is open; reconstructing them days later invites hindsight. Values in the third column are illustrative.
| Field | Definition | Illustrative value | When captured |
|---|---|---|---|
| deadline_at and deadline_source | The binding deadline and which record set it, from the table above | Friday 14:00, sales order promise date | At decision time, from the system |
| time_available | Working minutes between receipt and the latest point a decision could still meet the deadline, under the calendar that applied | 380 working minutes | Computed afterward from logs and calendars |
| time_used | Minutes from receipt to the recorded decision | 95 minutes | Computed afterward from logs |
| felt_pressure | Expert's rating on a single temporal demand item, with the scale stated | 4 on a 1 to 5 scale | Within the hour, by the expert |
| options_considered | Every option the expert weighed, in their own words | Four options, listed | At decision time, by the expert |
| options_ruled_out_by_time | Options rejected only because they could not finish before the deadline, each with a reason code | Two, lead_time and approval_latency | At decision time, by the expert |
| evidence_at_decision | Documents and data the expert had in hand, with retrieval times | Stock report, customer email | At decision time, from the system |
| evidence_after | Information that arrived after the decision and would have mattered | Supplier confirmation at 16:10 | At outcome review |
| unpressured_choice | What the expert says they would have chosen with no deadline, with confidence | Drop-ship, medium confidence | Within the hour, by the expert |
| outcome | What happened, checked against a record rather than memory | Delivered on time, one return | At outcome review |
How should an expert describe an option the deadline eliminated?
With a reason code and a number, so a later reader can tell whether the option was impossible or merely risky. Hard infeasibility and a judgment call about risk look identical in prose and lead to different training signals: the first is a constraint a model should learn to check, the second is a tradeoff it should learn to weigh.
At a glance
How should an expert describe an option the deadline eliminated
- lead_time: the option's own duration exceeds the time available, with both durations stated.
- approval_latency: a required sign-off could not arrive in time, with the usual turnaround stated.
- information_gap: data needed to choose the option could not be obtained before the deadline.
- resource_window: the person, part or vehicle the option needs was not free inside the window.
- risk_tradeoff: the option could have fit, but the expert judged the chance of missing the deadline too high; record that judgment as a tradeoff, not as infeasibility.
Which fact would have flipped the call, or made a second option just as good?
Ask for it explicitly, because experts rarely volunteer it. The unpressured_choice field is the first half: it names the path the expert prefers when time is not binding. The second half is the trigger, the specific fact that would have flipped the decision, such as a supplier confirming stock or a customer accepting a later date.
Three further checks keep the record honest. Disagreement: where possible, have a second expert review a sample of cases with the same clock fields and record whether they would have chosen differently under the same deadline. Calibration: compare the confidence the expert gave with what actually happened, by deadline band. Incentives: note whether missing the deadline carried a visible penalty for the expert or team, since a breach counter on a dashboard changes behavior as surely as a contract does.
When two options remain equally valid after review, keep both and mark the case as tied. A forced single label on a tied case is noise dressed up as ground truth.
How do you collect the expert's fields without adding to the pressure?
Split capture into a ten-second step during the decision and a two-minute step after it. Asking an expert under a deadline to fill in a form changes the decision you are trying to record, and asking a week later produces a tidy story instead of what happened.
At a glance
How do you collect the expert's fields without adding to the pressure
During the decision, collect only what the expert would write anyway: the options on the table and the one chosen. Many teams already do this in a ticket note or an order comment; a short template with one line per option turns that habit into structured data. Within the hour, a follow-up prompt collects the felt-pressure rating, the reason codes for ruled-out options and the unpressured choice. Everything else is computed from logs.
Pilot the routine on a small batch before scaling it. Watch for two failure signs: experts skipping the follow-up on exactly the most rushed cases, which biases the set toward calm decisions, and unpressured choices that always match the actual choice, which usually means the question is being answered by habit.
- During the decision: options considered, choice made.
- Within the hour: felt pressure, ruled-out options with reason codes, unpressured choice and confidence.
- From system logs: deadline, source, time available, time used, evidence on hand.
- At outcome review: what happened and what evidence arrived too late.
Worked example (illustrative): an expedite request at an industrial distributor
This case is invented to show how the fields work together; it does not describe a real company. A customer emails a distributor at 09:40 on Thursday needing a replacement pump seal by Friday at 14:00, or a production line stays down. The last premium freight pickup that can still arrive in time leaves at 16:00 on Thursday, so an inside sales specialist has 380 working minutes to decide how to fill the order.
The specialist weighs four options. Drop-shipping from the manufacturer takes five working days, so it is ruled out with reason lead_time. A compatible substitute is in local stock but needs engineering sign-off, which normally takes three working days, so it is ruled out with reason approval_latency. A transfer from a regional warehouse arrives Friday morning by premium freight. A partial shipment of the one seal on hand covers half the order.
The specialist chooses the regional transfer in 95 of the 380 available working minutes and rates felt pressure 4 on a 1 to 5 scale. Asked what they would do with a week, they answer drop-ship with medium confidence, because it avoids a warehouse imbalance. At outcome review, the record shows on-time delivery and one return of an unneeded seal.
Stored without the clock fields, this case teaches that regional transfers are preferred for seals. Stored with them, it teaches that regional transfers win when the deadline is under two working days and the substitute path needs approval, which is the lesson the specialist actually applied.
When should the deadline be left out of the training signal?
Leave it out, or flag the case, in three situations. When every record in the set faced the same deadline, the field has no variance and adds nothing a model can learn from. When the deadline was set by a system default rather than by any business need, such as a blanket SLA applied to every ticket, the pressure may be real but its cause is arbitrary. And when the rushed decision is exactly the behavior you do not want a model to copy, keep the case for evaluation and mark it constrained rather than using it as a target.
Know the limits of the specification as well. Self-reported pressure and unpressured choices are recollections, collected close to the event to reduce hindsight but never free of it. The specification captures what experts did and say they would do; it does not prove that the unpressured path would have turned out better.
Preparing deadline-stamped decision records for licensing
Clock fields are often the easiest part to export, because help desks, ERPs and scheduling tools already store them. The expert's own fields, felt pressure, ruled-out options and the unpressured choice, have to be collected deliberately and with the experts' knowledge, since they describe identifiable employees' judgment and workload. Plan the de-identification of those fields before the first case is collected.
A company with years of deadline-stamped decisions can learn whether it qualifies through SourceX's short company assessment, which asks for no files. Scope and terms are agreed and signed before anything moves, and the removal of identifying details, the experts' own included, is carried out by SourceX through approved processing routes as that agreement sets out.
Frequently asked questions
Is the SLA clock a good measure of the time pressure an expert felt?
Only partly. An SLA target measures a defined interval under defined pause rules and business-hour calendars, which may not match the time the expert could actually spend. Keep the SLA fields, but add working time and a short felt-pressure rating collected from the expert.
Why ask experts what they would choose without a deadline?
Because it separates their preference from the constraint. If the answer differs from what they chose, the case shows how the deadline changed the decision, which is the signal a model needs. If it matches, the deadline was not binding for this case.
Can annotation time stand in for decision time?
Not reliably. Annotation tools usually measure time in the labeling interface, and some include idle time and time across several sessions. Decision time should come from the operational record of when the request arrived and when the decision was logged.
How many fields does the expert have to fill in by hand?
Four in the specification above: felt pressure, options considered, options ruled out by time and the unpressured choice. The rest come from system logs or the outcome review. Keeping the manual part short is what makes capture at decision time realistic.
Should rushed decisions be removed from a training set?
Not automatically. A rushed decision that was correct for its deadline is useful when the deadline fields travel with it. Remove or reserve it for evaluation only when the rushed behavior is something you do not want a model to imitate.
Sources
- A systematic review of time pressure in software engineering reports that most high-quality studies found increased productivity and decreased quality under time pressure. Source
- Zendesk's requester wait time counts time in New, Open and On-hold and pauses while a ticket is Pending; total resolution time does not pause for Pending; SLA targets can run on business or calendar hours. Source
- Jira Service Management SLAs start on start conditions, can pause on optional conditions such as waiting for the customer, and stop as soon as any stop condition is met. Source
- The NASA Task Load Index temporal demand item asks how much time pressure the person felt due to the rate or pace at which the tasks occurred. Source
- O*NET's Time Pressure work context asks how often a job requires the worker to meet strict deadlines, rated from never to every day. Source
- Label Studio's annotation lead time includes idle time and is cumulative across sessions. Source
See if your company qualifies
A short company assessment. No data uploads are needed.