Skip to content

Systems and records

Conditional approval dataset: four clocks that decide when a yes holds

By SourceX Editorial · Reviewed by Noah Loul ·

10 sections

Short answer

A conditional approval dataset must keep each yes attached to what it depends on: the conditions, the authority that granted it and the window in which it holds. Store conditions as typed clauses with their own evidence and deadlines, derive the approval's status for any date from dated events, and join only records that settle a condition. Flattened to approved, the data teaches a model that conditions never bind.

Provenance trail

Conditional approval dataset

This page covers conditional approval dataset: four clocks that decide when a yes holds.Illustrative

Key takeaways

  • An approval with conditions is a bundle of facts: a decision, the clauses it depends on, the authority behind it and a validity window.
  • Conditions run on four clocks: before the approval takes effect, by a deadline, for as long as the approved object stays unchanged, and within stated limits.
  • Common approval tools record Approved or Rejected and push conditions into free-text comments, so a dataset has to rebuild them as typed clauses.
  • Status for a given date is derived from events, which is how a model learns that an approved vendor can still be blocked on day 75.
  • Join a downstream record only when it changes a condition's state, the authority evidence or the validity window; the rest adds noise and linkage risk.

What is a conditional approval dataset, and what must each record prove?

A conditional approval dataset is a set of business decisions in which a yes was granted subject to stated terms, recorded so that a model can tell whether the approval actually permitted a later action. Each record has to prove four things: what was decided, which conditions the decision depended on, who had the authority to grant it, and the period in which it held.

The question a training or evaluation task asks is rarely whether something was approved. It asks whether, on a given date, a given action was allowed under that approval. A purchase order sent to a vendor approved subject to an insurance certificate is allowed only once the certificate is on file. A research protocol approved with required consent-form changes cannot enroll anyone yet. The decision record alone answers neither question.

The evidence behind a useful answer has three parts: the condition text as the approver wrote it, dated evidence that each condition was met, missed or waived, and the authority record that applied on the decision date. Where any part is absent, the dataset should say so record by record instead of filling the gap with the most common outcome.

Where conditions go when a workflow tool only knows Approved

Most approval tools store a terminal status and leave conditions to free text, so the conditions survive only as prose nobody parses. Salesforce's classic approval history tracks a record through the statuses Submitted, Pending, Approved, Rejected, Reassigned and Recalled. None of them means approved subject to terms, so an approver who wants to attach terms types them into a comment, which Salesforce caps at 4,000 characters.

Steps

Where conditions go when a workflow tool only knows Approved

Some platforms do enforce one kind of condition. GitHub's branch protection can dismiss an approving pull request review as stale when new commits change the diff the reviewer approved; GitHub records the state of the diff at approval, and the pull request cannot merge until someone approves again. The approval is valid only while the reviewed content stays the same, and the dismissal leaves a trace in the history.

Regulated workflows spell conditions out because the consequences are legal. OHRP's 2010 guidance on IRB approval of research with conditions lets a review board require specified changes or confirmations, has a designated reviewer check the investigator's response, and treats the initial approval as effective only once that reviewer accepts the revised materials. The federal grant rule at 2 CFR 200.208 lets an agency attach specific conditions to an award, requires a notice naming the condition, its reason, the action that removes it and the time allowed, and requires the condition to be removed promptly once satisfied.

Flattening any of these into a single approved flag produces three different errors, and each one teaches a model something false:

  • False permission: the approval was granted but had not taken effect, because a precondition was still open.
  • Stale permission: the approval took effect and later lapsed, expired or was invalidated by a change to what was approved.
  • Lost scope: the approval held, but only for less than was requested, such as a smaller amount, fewer sites or a narrower use.

Four clocks a condition can run on

Classify every condition by the event that makes it bite, because that event decides how the approval's status changes over time. Four clocks cover the conditions found in research, regulatory, grant, engineering and commercial approvals. Expiry of the whole approval is recorded as a validity window beside them rather than as a fifth kind of condition.

The European Medicines Agency's conditional marketing authorisation runs several clocks at once. EMA describes it as granted for one year and renewable annually, carrying specific obligations, typically studies to complete or data to collect, that are reviewed at each renewal, and expiring automatically if no renewal application is made. A dataset that stored only the authorisation date would miss both the obligation deadlines and the expiry.

Each clock maps to a public example, the fields a record needs and the status that follows when the condition is not met:

Four clocks a condition can run on
ClockWhen it bitesPublic exampleFields to storeStatus if unmet
Before effectMust be met before the approval permits anythingOHRP approval with conditions: research may not start until a designated reviewer accepts the changesCondition text, evidence received, reviewer, acceptance dateApproved, not yet in force
By a deadlineThe approval runs now but must be earned by a dateEMA specific obligations, reviewed at each annual renewalDue date, evidence, consequence the terms nameLapsed, or in force with an open breach
While unchangedThe approval dies if the approved object changesGitHub dismissing a stale approving review after the diff changesVersion or fingerprint of the approved object, change eventsInvalidated until reviewed again
Within limitsThe approval holds only inside stated boundsODRL constraints, such as a date limit on a permissionOperand, operator, bound and unitIn force inside the bound, refused beyond it
Validity windowThe whole approval ends unless renewedEMA authorisation valid for one year unless renewedValid from, valid until, renewal eventsExpired

One status model for every workflow: derive it, never type it

Status should be computed from dated events for the moment a question is asked, not kept in a field someone edits. One small state model fits a research approval, a grant award, a vendor onboarding and a code review, and that shared model is what makes a cross-workflow dataset possible.

Steps

One status model for every workflow: derive it, never type it

The model has an entry state and six others. A request becomes approved with open terms; it moves into force once every before-effect condition has evidence; from there it can lapse when a deadline passes without the required evidence, which some terms let a late filing cure; be invalidated when the approved object changes; expire at the end of its window; or be revoked by someone with authority. Waivers and amendments are events as well, and they need the same authority evidence as the original decision.

The W3C's ODRL information model is a useful reference for the representation, though it says nothing about business status. It separates a permission from its constraints and duties: a permission may be exercised only if all its constraints are satisfied and all its duties fulfilled, and a duty can carry consequences that must also be fulfilled when its action has not been exercised. Mapping each business condition to a constraint, a duty or a consequence forces the annotator to decide which clock it runs on.

Every approval needs these fields, because each is an input to the status on a given date:

  • approval_id, plus the object approved and its version or fingerprint at decision time.
  • decision and decision_date, with the decision text verbatim.
  • approver and authority_ref: the delegation, policy or charter that empowered that person on that date.
  • conditions: one row each, with clock type, verbatim text, normalized bound or due date, and an evidence reference.
  • condition_events: met, missed, waived or amended, each dated and attributed to a person.
  • valid_from, valid_until and any renewal events.
  • derived_status for a query date, computed from the rows above and never typed by hand.
A state diagram for one approval: Requested leads to Approved with terms open, then In force, which can end as Lapsed, Invalidated, Expired or Revoked, with a dashed path showing that a late filing can return Lapsed to In force; a bracket notes that a flat status export shows Approved for all six post-decision states.
Approval status derived from dated events. Lapsed is highlighted because it is the state the illustrative vendor packet reaches on day 75.

Validation cases: five approvals the representation must get right

A representation counts as cross-workflow only if it derives the right status in every workflow it claims to cover. Before trusting one, run it against cases whose correct answer follows from the source rules, and compare its output with what a flat export would say.

The five cases below each name the clock, a query date and the status the representation must produce:

Validation cases: five approvals the representation must get right
WorkflowCondition and clockQuery dateExpected statusWhat a flat export says
Research ethics review (OHRP guidance)Consent wording change, before effectAfter the board's vote, before the reviewer accepts the changesApproved, not in forceApproved
Medicine authorisation (EMA)Annual renewal, validity windowAfter the one-year window with no renewal applicationExpiredApproved
Federal grant award (2 CFR 200.208)Specific condition with a removal action and time allowedAfter the agency confirms the action and removes the conditionIn force, condition removedApproved with conditions, never updated
Code review (GitHub branch protection)Approval tied to the reviewed diff, while unchangedAfter a new commit changes the diffInvalidatedApproved
Vendor onboarding (illustrative)Security questionnaire due in 60 days, by a deadlineDay 75 with no questionnaire on fileLapsedApproved

An annotated minimal task packet: a vendor approval with three open terms

A minimal task packet holds only what a model needs to answer one question about one approval, and annotates each element with the record that settles it. The case is illustrative: it describes no real company, and its dates are counted in days from the approval.

On day 75, a buyer at a distributor asks whether a second purchase order to a newly approved packaging vendor may go out. The procurement director approved the vendor on day 0 subject to three terms: a certificate of insurance on file before the first order, spending held under a cap for the first six months, and a completed security questionnaire within 60 days. The approval is valid for 12 months.

The correct answer is no. The certificate arrived on day 4 and the first order went out on day 9, both within the terms, but the questionnaire deadline passed on day 60 with nothing on file, and the onboarding policy in this example suspends ordering until it arrives. A flat export shows an approved vendor with one past order and invites the opposite answer.

Each element is listed with the record that settles it and what a model can learn from it:

An annotated minimal task packet: a vendor approval with three open terms
Packet elementIllustrative valueSettled byWhat it teaches
DecisionApproved with three terms on day 0Approval record and the approver's verbatim commentA yes can carry obligations
AuthorityProcurement director, vendor approvals delegated in policy version 3Delegation register as it stood on day 0Who could say yes on that date
Term 1, before effectInsurance certificate before the first order; received day 4Vendor document store, dated uploadOrder 1 on day 9 was allowed
Term 2, within limitsFirst-six-month spending cap; order 1 below itOrder totals in the ERP up to day 75Scope is part of the approval
Term 3, by a deadlineSecurity questionnaire due day 60; none receivedVendor risk tool, no submission eventSilence after a deadline is evidence
ValidityValid to day 365, never revokedApproval recordExpiry is not the reason here
Derived status on day 75Lapsed until the questionnaire arrives; order 2 blockedComputed from the rows aboveThe label a flat export gets wrong

Context completeness map for the vendor packet

A completeness map records, element by element, whether the evidence was captured at the time, rebuilt later or is missing, and which uses each state supports. It keeps the packet honest by telling a buyer which labels can train a model and which can only illustrate one.

In the vendor packet, the decision and validity window were captured by the approval record. The condition clocks were rebuilt by an annotator from a comment. The authority came from a delegation register restored to its day-0 version. The questionnaire term rests on the absence of a submission event, which is only as reliable as the vendor risk tool's logging, and evidence of absence is the weakest kind a packet can hold.

Read across each row before deciding how the packet may be used:

Context completeness map for the vendor packet
ElementEvidence stateWeak pointSupports
Decision and decision dateCaptured at the timeNone notedTraining and evaluation labels
Condition textCaptured as free textWording can be ambiguousTraining, after annotation review
Clock type for each conditionRebuilt by an annotatorA judgment call that needs double annotationTraining, with agreement reported
Authority on day 0Rebuilt from register historyRegister edits may lack datesTraining when the version is dated
Condition met or missedCaptured for terms 1 and 2; inferred from absence for term 3A logging gap looks like a missEvaluation of term 3 only after the logging is checked
Validity windowCaptured at the timeRenewals may sit in another toolTraining and evaluation labels

Would joining every downstream record settle the conditions?

No. Past the records that settle a condition, more joins mostly add noise and linkage risk, and they can change a label for the wrong reason. The strongest version of the opposing view is that a model handed the vendor's full history, every email, order and document, can work the conditions out for itself.

Two problems undercut it. Language models are measurably distracted by irrelevant material: Shi and colleagues (2023) built GSM-IC by adding irrelevant information to grade-school math problems and found performance fell sharply against the original problems across the prompting methods they tested, with some recovery when the model was told to ignore what was irrelevant. A vendor history full of unrelated threads poses the same test with business records.

The second problem is linkage. Every join made on a vendor name or an email domain can attach a certificate, an order or a questionnaire that belongs to a sister company or an older vendor with a similar name, and one wrong join can flip a condition from missed to met. The packet should state which key brought in each joined record, and joins made on names alone deserve a human check.

The stopping rule is easy to state and to audit: a record enters the packet only if it changes a condition's state, the authority evidence or the validity window. Everything else stays in the source system, where a reviewer can still find it.

Label, narrow or exclude: deciding what each approval can teach

Decide per approval from the completeness map, and write the rule down before labeling starts so it is applied the same way to every record. Three outcomes cover the cases:

At a glance

Label, narrow or exclude: deciding what each approval can teach

  • Label: conditions are typed, each has dated evidence or a dated deadline, and the authority is known for the decision date. The derived status can train and evaluate a model.
  • Narrow: conditions are typed, but satisfaction evidence is missing or rests on absence. Keep the record for training condition extraction and clock typing, and keep its derived status out of evaluation labels.
  • Exclude: the record says the approval was conditional, but the condition text is lost or the decision cannot be dated. A model cannot learn the right behavior from it, and keeping it tilts the set toward unconditional yeses.

Unwritten terms, informal waivers and what to request from a supplier

Unwritten terms and informal waivers are gaps no packet can close, so they belong in the dataset card, and the data request should be written to shrink them. Only conditions someone wrote down reach the data; terms agreed on a call never do, and an informal waiver looks like a miss that was ignored. Annotators' clock-type judgments need an agreement rate from a double-annotated subset before anyone relies on them, and every condition whose status rests on absence needs a check of the logging behind it. Those two measurements are the next evidence to collect.

A buyer who needs approvals with their conditions intact should ask for them explicitly, because a supplier's default export is usually the flat status. Name the workflows, ask for verbatim condition text and condition events, require the authority source for each decision date, and state the acceptance test: for a sample of approvals, the status derived for stated dates must match the source records.

SourceX sources datasets against requests like this one rather than from stock, so the request decides what is collected; the request form asks for scope, volume, format, timing and licensing needs. The company that holds the records approves the scope and terms and signs before anything is delivered. Approver names and comments are personal data, and SourceX performs de-identification itself or through its processing partner, so the request should say whether roles can stand in for names.

Frequently asked questions

Is an approval with conditions the same as an approval with comments?

No. A comment explains or advises; a condition changes what the approval permits. A quick test is to ignore the comment and ask whether the approval would then allow something the approver did not allow. If it would, the comment holds a condition and needs its own clock type, evidence and status.

Should an expired or lapsed approval be labeled as rejected?

No. A rejected request was never allowed, while an expired or lapsed approval was granted and then stopped holding. Keep the original decision and add the derived status for each query date, so a model learns that the same approval can permit an action in March and block it in May.

How should a waived condition be recorded?

As a dated event with the person who waived it and the authority that let them do so. A waiver changes the approval as much as the original decision did, so it needs the same evidence. An undocumented waiver is indistinguishable from a missed condition that nobody enforced.

Can a language model extract conditions from approval comments?

It can draft the extraction, but the clock type, the bound and the due date should be checked by people on a sample before the labels are used. Report how often two annotators agree on the clock type, and keep the verbatim comment beside every extracted condition so errors can be traced.

Which business workflows produce conditional approvals worth modeling?

Workflows where a later action depends on the condition: vendor onboarding, discount and credit approvals, change approvals, research protocols, grants and code reviews. If nothing downstream checks the condition, the approval history teaches little beyond the decision itself.

Sources

  • OHRP's 2010 guidance describes IRB approval with conditions: the IRB requires specified changes or confirmations, a designated reviewer checks the response, and initial approval becomes effective only when that reviewer accepts the revised materials. Source
  • EMA describes the conditional marketing authorisation as granted for one year and renewable annually, with specific obligations reviewed at renewal, and expiring automatically without a renewal application. Source
  • 2 CFR 200.208 lets a federal agency impose specific conditions on an award, requires notice of the condition, its reason, the action needed to remove it and the time allowed, and requires prompt removal once the condition is satisfied. Source
  • GitHub's branch protection can dismiss stale pull request approvals when commits change the approved diff, after which the pull request cannot merge until approved again. Source
  • The W3C ODRL Information Model defines a permission as exercisable only if all constraints are satisfied and all duties fulfilled, with consequences that must be fulfilled when a duty's action is not exercised. Source
  • Salesforce's classic approval history uses the statuses Submitted, Pending, Approved, Rejected, Reassigned and Recalled. Source
  • Salesforce limits approval request comments to 4,000 characters. Source
  • Shi and colleagues (2023) built GSM-IC by adding irrelevant information to math word problems and found model performance fell sharply compared with the original problems. 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