--- title: "Quality of hire: check the cohort before making a score" canonical: https://metix.ai/blog/quality-of-hire published: 2026-11-04 updated: 2026-11-04 author: "Metix AI" --- # Quality of hire: check the cohort before making a score Canonical: https://metix.ai/blog/quality-of-hire Missing reviews and different observation windows can distort a quality-of-hire score. Read the cohort, record the gaps and compare the same work at the same age. Metix AI editorial team The short version Quality of hire needs defined post-hire evidence. Before combining inputs, show which starts reached the chosen checkpoint and which records are known. The independent kit reports separate counts and missingness, without assigning employees a quality score. Decision guide ## Whose work could actually be observed? 1. ### Choose the work and window Agree on a role-specific goal, review checkpoint and reporting dates. 2. ### Check the cohort Retain pending starts, early exits and unknown employment history visibly. 3. ### Show each denominator Report work and manager record coverage before combining any inputs. A retained employee and an observed work goal answer different questions. The 90-day example is not a recommended review period. For employer recruiting and HR teams reviewing post-hire observations. Sources checked November 4, 2026. Every employment episode and result below is invented. Metix publishes this guide and sells recruiting software. ## Quality of hire starts with observed work after joining Quality of hire describes how new employees contribute against the outcomes an employer expects after they start. It is assessed from post-hire evidence, not from an interview score or a candidate profile. Performance, retention and manager feedback can inform that review. They measure different things, though. There is no universal quality-of-hire formula, and a percentage says little without its work definition, observation window and data coverage. Even a correctly calculated score can rest on incomplete observations. Some hires have reached the review point; others joined last week. One manager has submitted a review, another has not. Before combining those inputs, find out whose work could have been observed and which records actually exist. This guide provides a local worksheet for that first check. It reports maturity, retention through a chosen checkpoint and coverage of work-goal and manager records. The output shows the separate counts; its combined quality score is null. You can use the files without a Metix account. ## Six starts, four mature records and one unknown Our invented support-team cohort contains six continuous employment spells. Four people started in July; two started in October. We inspect the snapshot on November 4, using a 90-day checkpoint solely to make the arithmetic visible. Ninety days is not a recommended review period. Choose a window that gives the actual work enough time to become observable. **Invented records at November 4, 2026; start date is day zero** | Episode | Start / checkpoint | Known employment history | Full-window records | | - | - | - | - | | A | July 1 / September 29 | Confirmed employed through November 4: retained through checkpoint | Goal met; manager record present | | B | July 2 / September 30 | Left October 15: retained through checkpoint | Goal not met; manager record missing | | C | July 3 / October 1 | Left September 1: early exit | No full-window records | | D | July 4 / October 2 | Continuous employment confirmed only through September 30: unknown | No full-window records | | E, F | October 1 and 5 / December 30 and January 3 | Checkpoint not reached: pending | Not yet due | B left before the report date but after the checkpoint. Treating B as an early exit would change the question from retention through 90 days to employment today. D's missing confirmation cannot establish either retention or an exit. E and F have not had the same follow-up time as the July starts. Among four mature spells, two are retained through the checkpoint, one ended early and one has unknown history. Retention status is known for 3 of 4, giving 75% coverage. Retained among known is 2 ÷ 3 = 66.67%. Across all four mature spells, the retained share lies between 2 ÷ 4 = 50% and 3 ÷ 4 = 75%, depending on D's missing record. These bounds describe the missing record. They are neither confidence intervals nor quality scores. Early exit C remains visible in the mature cohort even though the person did not complete the work-observation window. The two confirmed full-window spells both have a goal observation: 2 of 2, or 100% coverage. One met the goal and one did not. Only A has a manager record, so manager coverage is 1 of 2, or 50%. Report that denominator of two beside the other mature records. These work and manager counts describe confirmed full-window spells only; they cannot represent the entire start cohort. Keeping early exits outside these component denominators creates a selected group, which the separate exit count makes visible. ## Download the cohort review kit and check the result [Download the complete quality-of-hire input review kit (ZIP)](https://metix.ai/resources/quality-of-hire/quality-of-hire-kit.zip). It contains the [six synthetic episodes](https://metix.ai/resources/quality-of-hire/start-cohort.csv), an [empty CSV](https://metix.ai/resources/quality-of-hire/blank-cohort.csv), [a definition card](https://metix.ai/resources/quality-of-hire/definition.json), [the local calculator](https://metix.ai/resources/quality-of-hire/review_cohort.py), [the expected result](https://metix.ai/resources/quality-of-hire/expected-nov4.json), [eight boundary tests](https://metix.ai/resources/quality-of-hire/test_review_cohort.py), [field and method instructions](https://metix.ai/resources/quality-of-hire/METHOD.md) and a [manual investigation log](https://metix.ai/resources/quality-of-hire/review-log.csv). ``` python3 review_cohort.py start-cohort.csv --as-of 2026-11-04 --checkpoint-days 90 ``` Extract the ZIP, open its folder and run the command. The script uses Python 3's standard library and prints JSON locally. It makes no network requests or file writes. Compare the result with the worked example before replacing any inputs. The method file includes the test command. Keep the synthetic file unchanged and put your own episodes in a separate CSV using the blank header. Edit the definition card before reviewing real records. Use one opaque ID per continuous employment spell; a rehire needs a separate episode. Keep identifying details and source evidence in the employer's record system. Start is day zero, and the checkpoint is start plus the chosen number of calendar days. Record local dates in one reporting timezone. The CSV records either continuous employment confirmed through a date or a known exit date. A confirmation means an uninterrupted spell from the start, not simply a recent file export. Here, exit date means the first date the person was no longer employed. An exit on the checkpoint does not count as retained through it. Choose this boundary before collecting data. All supplied dates must be known by the snapshot cutoff. The calculator rejects future evidence, duplicated IDs, partial observation records and mismatched windows. To recreate an earlier report, use a separately retained snapshot containing only what was known then; today's status cannot reconstruct what an earlier export omitted. This file is not an event-history system. An empty denominator returns `null`. Missing feedback is not a zero rating. Different IDs for the same spell still need private manual deduplication; no script can repair that identity problem from opaque IDs alone. ## Define a work outcome before asking for a rating Agree on the role's expected contribution and the evidence that would demonstrate it. A work milestone needs an owner, a relevant observation period and a clear meaning for met or not met. The goal belongs to the actual job. A generic impression of potential does not supply that definition. For the kit, the employer supplies a goal observation and its review date. The recorded observation window must end at the exact checkpoint. A review can be written later, but it must describe that agreed window. A manager record has its own window and date. The calculator checks their presence; it cannot verify the underlying work or the fairness of a judgement. For an invented support role, an employer might choose completion of a defined set of supervised cases, with documented checks against a case rubric. Record what counted as completed, who reviewed the cases and which work dates the observation covers, including any change in duties during that window. An unfinished assignment and an unsubmitted review are different states. The example goal needs adapting to the actual job; it is not a recommended performance standard. Keep the observation policy in the definition card, including role family and version. The script does not read that card or decide whether records are comparable. Run one agreed role family and goal definition per file. Retain the input, command and definition version with the output. ## Use a formula only after agreeing what its inputs mean [Lever's calculation guide](https://www.lever.co/blog/how-to-calculate-quality-of-hire) describes simple averages, weighted composites and threshold approaches. Those choices require defined inputs and consistent review windows. Its sample weights are illustrative. They do not establish a standard for another employer. A weighted composite has the form: sum of each comparable input score multiplied by its agreed weight. The weights must total 100%. A simple average assumes equal importance. Before either calculation, explain why the scales represent comparable evidence and document how each conversion works. Combining a retention flag, a work milestone and a manager opinion does not make their meanings interchangeable. If an input is missing, silently reallocating its weight changes the definition. Keep missingness visible and set the reporting policy before looking at results. Our kit provides the underlying counts and no default composite. [Workable's quality-of-hire guide](https://resources.workable.com/tutorial/quality-of-hire) also warns that turnover can reflect onboarding or management conditions. Retention alone cannot tell you whether someone met the job's goals, or whether a recruiting decision caused the outcome. ## Compare the same work at the same age Compare like roles after a suitable common observation period, using the same goal definition and coverage policy. A six-month engineering cohort and a 90-day support cohort answer different questions. A company-wide average can hide that difference even when every calculation is correct. Start with the gaps in the underlying records. Ask the data owner whether D's history can be confirmed, and whether B's manager review exists before trying to explain a difference between sources. Do not infer the missing answer from a profile, a recruiting grade or an accepted offer. Keep unknown values until the evidence arrives. Once records are usable, review a specific mismatch with the people who know the work. Did the agreed role change? Was the milestone realistic? Did the person have the tools and support required? Record the observation, alternative explanation and next check in the review log. A small cohort cannot isolate a sourcing-channel or product effect. If this cohort has only one complete pair of records, begin by repairing collection. A source ranking built from that pair would hide more than it explains. Use the [recruitment funnel guide](https://metix.ai/blog/recruitment-funnel-metrics) for movement beginning at first contact, the [offer acceptance guide](https://metix.ai/blog/offer-acceptance-rate) for formal-offer decisions and the [candidate experience survey](https://metix.ai/blog/candidate-experience-survey) for feedback about the process. Those are separate records from work after joining. ## Bring a verified role mismatch back to the brief If an employer review shows that the job's real work differed from the hiring brief, make that discrepancy specific before changing the next search. For example, ask which job requirement was missing and what evidence would have tested it. The aggregate result cannot tell you which new requirement would help. The [Metix AI recruiting workflow](https://metix.ai/features/ai-recruiter) structures the role, sources candidates, screens for interest and books interviews after you approve outreach. The employer supplies the bar and keeps interviews and hiring decisions. Any change to that bar needs an employer's judgement about the work. Metix is not a post-hire performance system or an HRIS. Work reviews, employment history and quality-of-hire definitions remain with the employer. This kit does not import into Metix, predict performance or demonstrate that Metix improves retention. If the next task is an approved role whose upstream search needs to change, create the role, review matching evidence and approve outreach. A missing manager review needs the employer's record owner first. ## Sources and limits We read Workable's complete quality-of-hire article, including its three formula images and historical survey chart, Lever's complete calculation guide and tables, and the current Metix product page. Sources were checked November 4, 2026. Workable's chart cites a 2016 survey; we do not use it as a current benchmark. The worksheet, episode dates, missing-record bounds and calculator are our independent example. They describe record coverage rather than a validated assessment model. Have the people responsible for the role choose its work measures and review period. Use the file to ask about the evidence; it supplies no decision about an employee. ### What is quality of hire? Quality of hire describes how new employees contribute against an employer's expected outcomes after they start. Work performance, retention and manager feedback can inform it, but the role definition, observation window and data coverage determine what a result means. ### What is the quality-of-hire formula? There is no universal formula. A simple average gives defined inputs equal weight; a weighted composite multiplies comparable input scores by agreed weights totaling 100%. Document scales, windows and missing-data policy first. This worksheet reports observation coverage and supplies no combined quality score. ### Should quality of hire be measured at 90 days? Choose a review point suited to the role and the time needed to observe its work. Ninety days is only the arithmetic example in this kit. A later checkpoint may be necessary; compare cohorts at a common appropriate age. ### How does the worksheet handle missing data? Younger employment spells remain pending. Mature spells with incomplete employment history remain unknown. Work-goal and manager records have separate coverage denominators among confirmed full-window spells; early exits stay visible in the mature cohort. A zero denominator produces null rather than zero. ### Does Metix measure or predict employee performance? Metix is not a post-hire performance system or an HRIS. Employers own work reviews and employment records. If a reviewed mismatch concerns the original hiring brief, Metix can support the upstream role-to-interview workflow after employer approval. The kit establishes no Metix effect on quality of hire. About the editorial team ## [Metix AI editorial team](https://metix.ai/about) Research & buying guides The Metix AI editorial team writes hiring guides and product comparisons. We publish source dates, distinguish vendor claims from observations, and keep worked examples separate from measured outcomes. Metix AI also sells recruiting software. [About Metix AI ↗](https://metix.ai/about) ## Next step See plans, Credit rules, and trial terms on the [Metix AI pricing page](https://metix.ai/pricing). A 14-day trial includes a one-time 1 AI Agent and 500 Credits, with no card required.