Contract-talent screening is the assessment a buyer runs on a freelancer, platform worker or agency contractor before awarding paid work. It exists because platform ratings are top-censored: on one large online labour market the share of contracts awarded five stars rose from 33% to 85% in six years, so a 4.9-star average no longer separates one candidate from another.
What follows is an illustrative walkthrough — not a client engagement — of the six decisions behind one contract-talent screening process. The situation is a composite: a services firm needs 40 contractors for a six-month data-remediation project, receives roughly 600 applications through two freelance platforms and one agency, and finds that almost every profile sits between 4.8 and 5.0.
Decision 1: Settle what the rating can and cannot carry
The honest answer is that it carries recency and completion, not capability.
The best evidence on this comes from Filippas, Horton and Golden's study of a large online labour market, where public feedback scores inflated steadily over a decade:
- What happened to the scale
- Average public feedback rose from 3.74 in early 2007 to 4.85 by May 2016.
- The share of contracts receiving the maximum rating went from 33% to 85%.
- Comparing numerical scores against scores predicted from the written review text, the authors attribute 67.7% of the increase to inflation rather than improvement.
- What happened to candour
- Among employers who privately said they would "definitely not" rehire a worker, 28.4% nonetheless gave four or more stars publicly.
- The same words drifted: reviews containing "terrible" averaged about 1.4 stars in 2008 and about 2.4 stars in 2015.
A scale on which 85% of observations sit in the top bin cannot rank 600 people. That is a ceiling problem, not a cynicism problem — and it is why the team treated the rating as an eligibility filter and nothing more.
Decision 2: Name the criterion before choosing the instrument
The project's failure mode was not slow work. It was silent error — contractors mis-coding records without flagging ambiguity. So the criterion was error rate on ambiguous records, plus escalation behaviour.
That choice determines the instrument, and the current validity evidence narrows the field. Sackett and colleagues' re-analysis, which corrected the range-restriction adjustments behind the older figures, reports operational validities of roughly:
- Structured interviews: .42
- Job knowledge tests: .40
- Empirically keyed biodata: .38
- Work sample tests: .33
- Cognitive ability: .31
Two things follow for contract work. First, a work sample is not automatically the strongest predictor, despite the intuition that contract work should be judged by contract output. Second, a job knowledge test does most of the work here at a fraction of the administration cost — which matters when the pool is 600 and the engagement is six months.
The team's design: a 20-minute job knowledge test on the specific data standard, then a short AI-graded scenario set on ambiguous records for everyone who cleared it, then a structured interview for the final 60.
Decision 3: Price the candidate's time, not just yours
A contractor pool is not a captive applicant pool. Someone with billable alternatives will abandon a two-hour unpaid assessment, and the people most likely to abandon are the ones with the most work already — exactly the wrong selection gradient.
The practical constraints the team set:
- Total unpaid candidate time under 35 minutes before any human contact.
- Assessment sequenced so the cheapest, highest-validity instrument runs first.
- Score reporting back to the candidate, so the time spent produces something they keep.
- Per-candidate cost, not per-seat licensing — AssessAll's pay-as-you-go credits at ₹30 (about US$0.50) each make a 600-candidate screen a line item rather than a procurement exercise.
Point 3 is not courtesy. In a pool where everyone's platform rating is uninformative, a result the candidate can show elsewhere is the only compensation on offer for unpaid assessment time.
Decision 4: Decide the integrity question explicitly
Remote, unsupervised, unpaid, and assessed on a tested skill the candidate will later be paid for: that is the highest-incentive cheating configuration in assessment. The team's answer was proportionate rather than maximal — identity confirmation at the structured interview stage, integrity bands from AI proctoring on the scenario set to triage who gets a closer look, and no automatic rejection on a flag alone.
The reason for the last point is now partly legal. The EU's Platform Work Directive, Directive (EU) 2024/2831, must be transposed by member states by 2 December 2026. It requires human oversight of automated monitoring and decision systems by trained staff able to override them, bars processing of data on a worker's emotional or psychological state or data collected while the person is not working, and requires platforms to review the effects of those systems at least every two years. A screening process that auto-rejects on a proctoring signal, with no human in the loop, is the pattern the directive is written against.
Decision 5: Write down what the process will be asked to justify
In June 2026 the International Labour Conference adopted the Decent Work in the Platform Economy Convention, 2026 (No. 193) — the first international labour standard for the platform economy. It applies whether or not a platform worker is in an employment relationship, and it includes principles on the transparent use of automated and algorithmic decision systems and on access to review mechanisms where those systems affect workers.
Ratification is a slow, country-by-country process and the Convention does not reclassify anyone's employment status. But the direction is settled enough to design for. The records the team chose to keep were the ordinary ones: the job analysis linking the test content to the task, the cut score and how it was set, pass rates by group, the override log for every proctoring flag, and a dated note of who reviewed each automated decision.
Decision 6: Decide who owns the result afterwards
A contract pool refreshes constantly. Screening 600 people for a six-month engagement and then discarding the results means screening them again next year.
The scale argument is real. India's gig and platform workforce was estimated at 7.7 million in 2020-21, projected to reach 23.5 million by 2029-30 by NITI Aayog. Meanwhile the ILO's February 2026 research brief counted 653 active digital labour platforms as of October 2025 — fewer than the 777 reported for comparable sectors in 2021, which the brief attributes to closures and consolidation. Reputation built inside a platform is reputation that can disappear with the platform.
A candidate-held, verifiable record — a Skill Passport rather than a profile badge — survives both the engagement and the platform. The team's rule was that every assessed candidate keeps their result regardless of outcome, and re-verification is required only when the underlying standard changes.
When the rating is enough
For a one-off task under a few hours, where the cost of a bad hire is the cost of redoing the task, screening is not worth building. The rating plus a scan of the written reviews is a reasonable filter at that stake, and the inflation evidence is about the numerical scale, not the review text — the text kept carrying signal even as the stars stopped. Screening earns its cost when the engagement is long, the pool is large, or an undetected error is expensive.
The takeaway: a platform rating is a completion record on a ceiling-bound scale, so a contract pool needs the same thing a permanent one does — a named criterion, an instrument chosen on validity evidence, and a written record of how the decision was made.