Deal economics
Dataset gross margin rework: the cycles a reported margin leaves out
By SourceX Editorial · Reviewed by Noah Loul ·
10 sections
Short answer
Dataset gross margin rework is the preparation repeated after a dataset fails a check, and it usually hides in overhead. Log every cycle with its trigger, owner and hours, separate buyer-requested changes from preventable defects, then recharge rework and unsuccessful candidates to the accepted cohort. In this article's illustration, a 58 percent reported margin falls to 36, and unwritten acceptance criteria caused nearly half the rework hours.
Value drivers
Dataset gross margin rework
Key takeaways
- Rework is any preparation cycle after the first full pass that exists because output failed a check; log each one with trigger, owner, hours and metered re-runs.
- Buyer changes to a signed requirement are recoverable; extraction, de-identification and documentation defects are not, and unwritten criteria count as the seller's defect.
- In the illustration, recharging rework and unsuccessful candidates cuts gross margin from 58 to 36 percent of net revenue.
- At a constant pass rate p, a dataset needs (1 - p)/p repeat cycles on average: one at 50 percent first-pass acceptance, a third at 75 percent.
- Automation is billed on every pass, including failed ones, so a pre-check that lifts first-pass acceptance only from 50 to 60 percent can save nothing.
- SourceX has not published rework or margin metrics; it shares company and commercial metrics with investors directly.
Where does rework hide in dataset gross margin?
Rework hides in three places: repeat preparation cycles booked to overhead instead of to the dataset, hours spent on candidates that were never accepted, and buyer-requested changes done without a change order. Each leaves an accepted dataset's reported gross margin looking better than the process that produced it.
Rework, here, means any preparation cycle after the first full pass that exists because the output failed a check, whether the check was an internal review, a privacy scan or the buyer's acceptance test. It differs from remediation of the source records, which is a decision to change scope, and from a second license, which reuses work that already passed.
The question for an investor is not whether a data company has rework; any process that prepares real operating records will have some. It is whether rework is recorded per dataset, attributed to a cause and charged back before margin is reported. A company that can show that has a margin. A company that cannot has an estimate.
Even heavily studied public datasets carry defects that take repeated human review to fix. The authors of Pervasive Label Errors in Test Sets released corrections for the test sets of ten widely used benchmarks, from ImageNet to Amazon Reviews, built from Mechanical Turk reviews with an option to change a label only when all five reviewers agreed, and two of the sets had too many label errors to correct.
What belongs in a dataset rework ledger?
A rework ledger records one row per preparation cycle with eight groups of fields, enough to tie every repeat hour to a dataset, a cause and a decision about who pays. Without the trigger and detection fields, rework cannot be split into recoverable and preventable work, and margin cannot be corrected.
Two rules keep the ledger honest. A cycle is logged when it opens, not when it is billed, so abandoned fixes still count. And the first full pass is never a cycle, so the ledger measures repetition rather than the base preparation already inside the deal margin.
| Field | Definition | Example entry (illustrative) | Why an investor needs it |
|---|---|---|---|
| Cycle number and dates | Sequence after the first full pass, with open and close dates | Cycle 3, opened day 88, closed day 95 | Shows how long fixes hold up acceptance |
| Trigger | The failed check that opened the cycle | Buyer sample review: outcome field ambiguous | Separates internal catches from buyer catches |
| Detected by | Internal review, privacy scan or buyer acceptance test | Buyer acceptance test | Buyer-detected defects cost trust as well as hours |
| Stage reopened | Extraction, de-identification, linking, outcome labels or documentation | Outcome labels | Locates the process that failed |
| Owner and hours | Role responsible and person-hours spent | Data engineer 14 hours; reviewer 6 hours | Converts the cycle into cost |
| Metered re-runs | Pages, gigabytes or jobs processed again | 4,200 pages re-extracted | Automation is billed again on every pass |
| Recoverable flag | Whether the buyer changed a signed requirement | Not recoverable: the criterion was never written | Decides whether the cost can be billed |
| Root cause and fix | Coded cause and the change that prevents a repeat | Spec gap: no definition of resolved; definition added | Shows whether the same defect will return |
Buyer-requested change or preventable defect: who pays for the cycle?
A cycle is recoverable only when the buyer changed an agreed requirement after sign-off; every other cycle is a preventable cost of the seller's process, including cycles caused by criteria nobody wrote down. Treating an unwritten requirement as the buyer's change is an easy way to relabel rework as scope.
At a glance
Buyer-requested change or preventable defect: who pays for the cycle
The spec-gap rule borrows from evaluation practice. Anthropic's guide to agent evals treats each task as one test whose inputs and success criteria are spelled out. A dataset's acceptance test should meet the same bar before the first hour of preparation; otherwise the first buyer review writes it, at the seller's expense.
- Recoverable: the buyer adds a field, changes a time window or tightens a threshold that the signed acceptance test did not contain. Price it as a change order or into the next license.
- Preventable, extraction: wrong joins, dropped attachments, truncated text or broken encodings produced by the export or the pipeline.
- Preventable, de-identification: identifiers missed, or context over-redacted. Automated detection cuts review hours but does not remove them: Microsoft's Presidio documentation warns that because it relies on automated detection mechanisms, it may not find all sensitive information, and that additional systems and protections should be used.
- Preventable, documentation: field meanings, units, time zones or outcome definitions missing from the data dictionary.
- Preventable, spec gap: the acceptance test never defined what a pass looks like, so the buyer's first review becomes the specification.
The process-cost heat map: which stage and trigger consume the hours
In the illustrative dataset, spec gaps consumed 40 of 84 rework hours, nearly half, and extraction was the costliest stage. The heat map crosses five preparation stages with four triggers so that the largest cell, not the loudest complaint, sets the fix.
Read the columns before the rows. Source defects, 14 hours, are partly a sourcing problem that intake checks can catch. Tool errors, 18 hours, are an engineering problem. Spec gaps, 40 hours spread across every stage, are a contracting problem: they start before preparation, and no pipeline improvement removes them. Buyer changes, 12 hours, are the only column that can be billed.
Hours are person-hours from the ledger. The illustration assumes one dataset that needed three repeat cycles after a 150-hour first pass, so rework added 56 percent to its preparation hours. A company's own ledger replaces every cell.
| Stage | Source defect | Tool error | Spec gap | Buyer change |
|---|---|---|---|---|
| Extraction | 5 | 9 | 7 | 2 |
| De-identification | 1 | 6 | 10 | 0 |
| Linking and joins | 6 | 2 | 7 | 1 |
| Outcome labels | 2 | 1 | 9 | 5 |
| Documentation | 0 | 0 | 7 | 4 |
| Total, 84 hours | 14 | 18 | 40 | 12 |
Gross margin before and after charging rework and failed candidates
Charged back, rework and unsuccessful candidates cut the illustrative dataset's gross margin from 58 to 36 percent of net revenue. The corrected waterfall uses the same 84 hours as the heat map, converted at 6 person-hours per point of net revenue, plus 48 hours per accepted dataset from candidates in the same cohort that were never accepted.
Net revenue here is the buyer's payment minus the supplier's payout. Under ASC 606 as amended by ASU 2016-08, a company acting as principal reports the gross buyer payment as revenue and an agent reports only its fee or commission, so percentage margins are comparable across companies only after both are restated to the same net base.
The 2 points of unbilled buyer changes look small, but they are the only lost points a contract clause recovers directly. The 12 points of preventable rework need process changes, and the 8 points from unsuccessful candidates need earlier screening, which is a sourcing decision rather than a preparation one.
| Step | Points of net revenue | Margin after the step | Source of the adjustment |
|---|---|---|---|
| Reported gross margin | 58 | 58 | Deal margin with rework booked to overhead |
| Preventable rework charged back, 72 hours | -12 | 46 | Ledger: source defects, tool errors and spec gaps |
| Recoverable buyer changes not billed, 12 hours | -2 | 44 | Ledger: recoverable flag with no change order |
| Unsuccessful candidates, 48 hours per accepted dataset | -8 | 36 | Cohort record of candidates never accepted |
| Corrected gross margin | 36 | 36 | Every line charged to the accepted dataset |
What is higher first-pass acceptance worth once automation is billed?
Higher first-pass acceptance is worth less than it looks when automated checks are billed on every pass, including the passes that fail. With a constant chance p that a pass is accepted, a dataset needs on average 1/p passes and (1 - p)/p repeat cycles: two passes and one repeat at 50 percent, about 1.33 passes and a third of a repeat at 75 percent.
The table holds the cost of a repeat cycle at 5 points of net revenue, charges a metered pre-check at 1 point per pass and starts from 50 points of margin after unsuccessful candidates. A pre-check that lifts first-pass acceptance only from 50 to 60 percent saves 1.7 points of rework and spends 1.7 points on meters, so it pays nothing.
The meters are real and published. On October 26, 2026, Google Cloud listed Document AI's Form Parser at $30 per 1,000 pages for the first million pages a month, and its Sensitive Data Protection pricing example charges $3.00 per GB inspected plus $2.00 per GB transformed. Every re-run of the same pages or gigabytes is billed again.
Meters are rarely the expensive part of a pass, though. At that Form Parser rate, a machine pass over 60 pages costs $1.80, while one person-minute of checking per page over the same 60 pages is a full hour of labor. Automation earns its place when it removes human minutes per page or prevents a repeat cycle, and it should be credited only with what the ledger shows it removed.
| Case | First-pass acceptance | Expected passes per dataset | Points lost to rework and meters | Corrected margin |
|---|---|---|---|---|
| Manual review only | 50 percent | 2.0 | 5.0, rework only | 45.0 |
| Pre-check with a modest lift | 60 percent | 1.67 | 3.3 rework plus 1.7 meters: 5.0 | 45.0 |
| Pre-check with a strong lift | 75 percent | 1.33 | 1.7 rework plus 1.3 meters: 3.0 | 47.0 |
| Pre-check plus a written acceptance test | 90 percent | 1.11 | 0.6 rework plus 1.1 meters: 1.7 | 48.3 |
When does a profitable-looking dataset stop being worth preparing?
A dataset becomes unattractive when its expected repeat cycles exceed the margin left above the investor's hurdle divided by the cost of one cycle. With 50 points after unsuccessful candidates, a 30-point hurdle and 5 points per cycle, the limit is 4 repeat cycles, which a constant pass rate reaches at 20 percent first-pass acceptance.
Averages hide the tail, so the operating thresholds below apply per dataset and per buyer, each with the action it should trigger. They are illustrative rules to adapt, not industry standards.
The hurdle is the investor's choice and should be stated. What matters is that the threshold is computed from ledger data before a dataset enters preparation, so a decision to reprice or decline is made while it is still cheap.
| Signal | Illustrative threshold | What it indicates | Action |
|---|---|---|---|
| Expected repeat cycles per dataset | More than 4 at 5 points per cycle | Margin falls below a 30-point hurdle | Reprice before preparation, or decline |
| Spec-gap share of rework hours | More than a third | Buyer reviews are writing the acceptance criteria | Require a written acceptance test before preparation |
| Unbilled buyer-change hours | Any recurring pattern | Recoverable work is being given away | Add a change-order clause to the license |
| First-pass acceptance | Below 50 percent | One or more repeat cycles per dataset on average | Fix the largest heat map cell before adding volume |
| Same root cause in consecutive cycles | Twice in a row | The fix is not holding | Stop, re-scope and record the decision |
Rework evidence that can mislead a diligence team
Rework data is easy to make look better without changing the work, so test how it was recorded before trusting the margin it supports.
None of these requires bad faith. They follow when a young company's time tracking was built for billing rather than cost attribution, which is why a ledger defined in advance is worth more than a margin reported after the fact.
- Relabeling. Cycles caused by unwritten criteria recorded as buyer changes make preventable rework look recoverable. Ask for the signed acceptance test beside each change claimed as the buyer's.
- Overhead absorption. Repeat hours charged to a shared data operations cost center vanish from dataset margin. Reconcile ledger hours with total preparation payroll hours for the same period.
- Gross presentation. A company that reports buyer payments gross shows a lower percentage margin than one that reports net for the same deal. Restate both to net revenue before comparing.
- Automation credit. A pipeline that cuts hours per pass but adds passes can raise total cost. Ask for hours and metered units per accepted dataset, before and after the change.
- Defects found after acceptance. Errors a buyer finds after paying, such as residual identifiers, create remediation and trust costs outside any margin. Presidio's documentation advises additional protections because automated detection may miss sensitive information, so ask what share of each delivery a person reviews.
Diligence questions on rework, SourceX included
Ask for the ledger itself, then for the corrected margin it implies, and put the same questions to SourceX.
For businesses licensing operational data to AI developers, SourceX handles rights review, the contract, preparation of the data, its delivery and payment, so its preparation cycles, and any rework in them, sit inside one operator. Its standing offer terms leave de-identification to SourceX itself or to its processing partner, so any rework cycle runs inside that one preparation step, as a schedule risk as well as a cost.
SourceX has not published rework, first-pass acceptance or margin metrics; its company and commercial metrics are discussed with investors directly. The company's Rule 506(c) raise is open to accredited investors, and the investor page is where materials can be requested.
- How many repeat cycles did accepted datasets need over the last four quarters, and what share was detected by buyers rather than internal checks?
- What share of rework hours traces to acceptance criteria that were unwritten or changed?
- Which preparation stage and trigger hold the largest share of rework hours, and what changed after you found it?
- How are recoverable buyer changes priced, and how many were billed?
- What is gross margin before and after charging rework and unsuccessful candidates, on a net revenue basis?
- What share of each delivery does a person review after automated de-identification?
Methods note: what the ledger model assumes and what it cannot show
The rework ledger, heat map, corrected margin and first-pass model are editorial tools built for this article on 2026-10-26. All values are illustrative: one hypothetical dataset with a 150-hour first pass and 84 hours of rework across three repeat cycles, a conversion of 6 person-hours per point of net revenue for the corrected margin, and separate assumptions of 5 points per repeat cycle and 1 point per metered pass for the first-pass model. None of it describes SourceX or any other company.
Public inputs are FASB's ASU 2016-08, Google Cloud's Document AI and Sensitive Data Protection price lists as listed on 2026-10-26, Microsoft's Presidio documentation, the label-errors repository behind Pervasive Label Errors in Test Sets, and Anthropic's guide to agent evals. Not measured: real rework hours, first-pass acceptance, trigger mix or corrected margins at any company. The constant pass rate is a simplification; in practice later cycles can get cheaper as fixes accumulate, or harder as easy defects run out.
Frequently asked questions
What counts as rework in data preparation?
Any preparation cycle after the first full pass that exists because the output failed a check: an internal review, a privacy scan or the buyer's acceptance test. Base preparation is not rework, and neither is reusing approved work for a second license. Log each cycle with its trigger, owner and hours.
Should buyer-requested changes count as rework?
Count them, but flag them as recoverable when the buyer changed a requirement after signing the acceptance test, because those hours can be billed or priced into the next license. Changes that fill a gap in criteria nobody wrote down are the seller's preventable cost.
How does first-pass acceptance affect dataset gross margin?
With a constant pass rate p, a dataset needs on average (1 - p)/p repeat cycles, so moving from 50 to 75 percent first-pass acceptance cuts expected repeats from one to a third. Multiply by the cost of one cycle to see the margin effect, then subtract any metered automation that runs on every pass.
Does automation remove rework costs?
It moves some cost from hours to meters and can prevent cycles, but it is billed on every pass and still needs human checks. Presidio's documentation, for example, advises additional protections because automated detection may miss sensitive information. Credit automation only with the hours and cycles the ledger shows it removed.
Does SourceX publish rework or margin metrics?
No. SourceX has not published rework, first-pass acceptance or margin metrics, and it shares company and commercial metrics with investors directly. To see how rework and first-pass acceptance are defined, request investor materials via sourcex.si/investors; only verified accredited investors can take part in any offering.
Sources
- The label-errors repository behind Pervasive Label Errors in Test Sets provides corrections for the test sets of ten common benchmarks, including ImageNet, MNIST and Amazon Reviews, based on Mechanical Turk reviews that can require all five reviewers to agree, and says Amazon Reviews and QuickDraw had too many label errors to correct. Source
- Microsoft's Presidio documentation says that because it uses automated detection mechanisms it may not find all sensitive information, and that additional systems and protections should be employed. Source
- Anthropic's guide to agent evals (January 9, 2026) defines a task as a single test with defined inputs and success criteria. Source
- Google Cloud's Document AI price list showed Form Parser at $30 per 1,000 pages for up to 1,000,000 pages a month, as listed on October 26, 2026, with pricing examples computed per page sent. Source
- Google Cloud's Sensitive Data Protection pricing example charges $3.00 per GB inspected and $2.00 per GB transformed. Source
- FASB's ASU 2016-08 says a principal, which controls the specified good or service before transfer, recognizes the gross amount, while an agent recognizes only its fee or commission. Source
Request SourceX investor materials
SourceX shares company and commercial metrics with investors directly. Any offering is made only to verified accredited investors.