Practical guide

AI market-sizing model verification checklist

Use this checklist after a market-sizing model has been built and before anyone relies on it for research priority, investment, capacity, or commercial planning. It does not teach the BOUND construction method. It tests the finished model as a controlled decision artifact: whether its population can be reconstructed, its data can support the intended use, its transformations preserve meaning,.

By Two Prune

Use this checklist after a market-sizing model has been built and before anyone relies on it for research priority, investment, capacity, or commercial planning. It does not teach the BOUND construction method. It tests the finished model as a controlled decision artifact: whether its population can be reconstructed, its data can support the intended use, its transformations preserve meaning, its calculations behave under challenge, and its conclusion stays inside the evidence.

What this checklist can verify

The U.S. Government Accountability Office frames data reliability as fitness for an intended use and asks reviewers to consider source, controls, testing, limitations, and the effect of problems on the conclusion. Census County Business Patterns and the Bureau of Labor Statistics Quarterly Census of Employment and Wages publish different measures, populations, and exclusions. Those differences are exactly why a reviewer must test the unit and boundary rather than merely confirm that a government URL was used. These sources support a review method; they do not validate a particular estimate. GAO: Assessing Data Reliability Census: County Business Patterns methodology BLS: QCEW overview

Sources: [1] [2] [3]

Run the PROBE review

Mark every item pass, fail, not applicable, or not evidenced. A pass needs an evidence ID, reproducible calculation, or retained source extract. Record critical failures separately; do not average them into a reassuring score.

### P: Pin the decision and population

Stop condition: hold the model when the decision or counted unit is ambiguous, or when the conclusion silently changes from establishments to companies, jobs to people, or interest to demand.

### R: Reconstruct source populations

Stop condition: do not issue a sourced population count when the extraction cannot be repeated, the population is materially different from the model boundary, or a suppressed value has been silently invented.

### O: Observe assumptions and transformations

Stop condition: return the model when a material output depends on a percentage whose denominator is missing or incompatible, or when generated estimates are presented as observed data.

### B: Break the arithmetic and boundaries

Stop condition: hold the dependent conclusion when a sampled calculation is irreproducible, an edge case changes unrelated outputs, or an entity can enter the model twice.

### E: Expose uncertainty and decision limits

Stop condition: do not publish a precise headline when plausible uncertainty can reverse the action and the model offers neither a bounded range nor a contingency.

  • [ ] The decision, accountable user, decision date, horizon, currency, and acceptable uncertainty are explicit.
  • [ ] The counted unit is stable: enterprise, establishment, team, seat, transaction, or another defined unit.
  • [ ] Geography, industry classification and version, size bands, operating characteristics, buyer, use case, and exclusions are stated.
  • [ ] The model distinguishes addressable population from reachable population, demand, conversion, and revenue.
  • [ ] Every headline number identifies whether it is observed, derived, assumed, or not evidenced.
  • [ ] Each source record includes publisher, dataset, table or endpoint, release date, reference period, extraction date, unit, population, filters, and revision status.
  • [ ] The reviewer can reproduce the sourced total from a retained extract or documented query.
  • [ ] Coverage exclusions, suppressed cells, disclosure treatment, imputation, sampling, and classification changes are recorded.
  • [ ] Cross-source joins explain how units, geographies, periods, and classification systems were aligned.
  • [ ] A source's reported coverage is not misread as coverage of the proposed customer population.
  • [ ] Eligibility, reachability, price, frequency, adoption, and conversion assumptions each have an owner, rationale, range, and replacement evidence.
  • [ ] Survey interest, tool use, and website traffic are not treated as purchase probability without a defensible bridge.
  • [ ] Percentages taken from another study preserve its denominator, population, period, and question wording.
  • [ ] Crosswalks and normalization rules are visible and reversible.
  • [ ] AI-suggested classifications or estimates remain proposals until an accountable reviewer verifies them.
  • [ ] A second reviewer independently recalculates sampled totals, rates, joins, and scenario outputs.
  • [ ] Unit, period, sign, currency, duplicate, missing-value, range, and impossible-value checks pass.
  • [ ] Removing one segment changes only the intended totals; changing one assumption changes only its dependants.
  • [ ] Reordered rows, blank values, zero values, classification edge cases, and rounding do not create hidden changes.
  • [ ] Scenario labels describe coherent states and do not imply probabilities without a probability method.
  • [ ] Double counting across locations, subsidiaries, categories, or channels has been tested.
  • [ ] The model shows a range or scenario set and identifies the assumptions that drive the decision.
  • [ ] A sensitivity ranking separates high-impact uncertainty from cosmetic precision.
  • [ ] Missing evidence is visible and connected to a next research action.
  • [ ] The decision is robust across plausible cases, or a contingency and trigger are stated.
  • [ ] The release record names the model version, preparer, reviewer, open exceptions, review date, and refresh triggers.
  • [ ] The conclusion does not claim demonstrated demand, revenue, willingness to pay, or competitive win rate from an addressable-population calculation.

Verification record

| Review area | Result | Evidence IDs | Failure or uncertainty | Owner | Decision consequence | | --- | --- | --- | --- | --- | --- | | decision and population | | | | | | | source reconstruction | | | | | | | assumptions and transformations | | | | | | | arithmetic and boundary tests | | | | | | | uncertainty and release | | | | | |

Choose one disposition: verified for the stated use, verified with visible limitations, return for repair, or hold the dependent decision. The disposition applies only to the reviewed version, sources, and intended use.

Worked verification: a synthetic operations market model

A fictional team submits a market model for prioritizing research into a U.S. operations product. Its base case starts with 18,400 establishments, applies 22% workflow eligibility, assumes 65% reachability, and multiplies 2,631 rounded reachable establishments by $7,200 annual value to produce about $18.9 million.

PROBE first confirms that 18,400 is a reproducible establishment count, not a company count. The source extract and classification filters are retained. The reviewer then finds two boundary defects. A size-band crosswalk includes 310 establishments from an adjacent category, and the team's eligibility review sampled companies while the denominator counts establishments. The arithmetic is correct, but the evidence cannot support a 22% establishment-level eligibility rate.

The break tests also show that rounding each industry before aggregation adds 14 reachable units. That difference is not decision-changing, but it is corrected. A sensitivity table shows that eligibility is still the dominant uncertainty: at 12%, 22%, and 30%, with other base assumptions held constant, the illustrative opportunity is approximately $10.3 million, $18.9 million, and $25.8 million before the crosswalk correction.

The disposition is return for repair. The team removes the adjacent category, relabels the company review as exploratory rather than representative, and replaces the point eligibility assumption with a provisional interval. Its next evidence task is a blinded establishment-level review using a written eligibility rubric. The model may prioritize that research; it may not support a revenue commitment. The checklist has verified the model's limits, not the commercial opportunity.

Two Prune boundary

Two Prune can assess whether a participant reconstructs a supplied market model, checks its evidence and calculations, challenges unsupported assumptions, and communicates a bounded disposition. It reports observed performance in that assessment. It does not validate a customer's market estimate, provide investment advice, predict commercial outcomes or workplace performance, or make employment decisions.

Sources

U.S. GAO: Assessing Data Reliability

  • U.S. Census Bureau: County Business Patterns methodology
  • U.S. Bureau of Labor Statistics: QCEW overview

Sources: [1] [2] [3]

Sources

  1. 1.GAO: Assessing Data Reliability · U.S. Government Accountability Office
  2. 2.Census: County Business Patterns methodology · U.S. Census Bureau
  3. 3.BLS: QCEW overview · U.S. Bureau of Labor Statistics