Evaluator Guild

Reviewer handbook

Everything a reviewer needs to know about working with us, in one place, written before we had a roster so that the first person to read it gets the same terms as the hundredth.

Last updated 13 September 2026 · Protocol v9.3

1. Getting in

  1. Apply on the guild page. We reply within a week.
  2. The battery. Fifty items, four hours, proctored, unpaid. Drawn from the eight validated failure families with matched clean controls. Passing standard: at most one Sev-1 miss, ≥85% Sev-2/3 recall, ≤5% false positives, Brier ≤ 0.15 with positive skill.
  3. Calibration. Paid at your rate from the first hour: a dense overlapping set of ≥50 challenge queries rated alongside other reviewers, which is what makes the latent-class model identifiable and your competency vector meaningful.
  4. The roster. Identity verification, confidentiality agreement, tax and payment setup. Then work, as it exists.

What we promise about timing

A reply to every application within a week; a battery result within five business days of sitting it; payment weekly for approved hours. If we miss one of these, tell us and we will say why.

2. Pay

RoleRateGate
Qualified reviewer (Tier 3)$65 per hour, loaded≥98% Sev-1 recall on the 50-item battery; ≤5% false-positive rate on matched controls
Senior assurance engineer$120 per hourAuthors Type B/C fixtures; dbt and Snowflake execution-plan fluency
Lead adjudicator (Tier 4)$150 per hourResolves Sev-1 disputes; Brier score ≤0.15 and BSS > 0 over the trailing 90 days

Calibration hours are paid at your rate. The only unpaid step is the 50-item qualification battery itself. Rates are published per role and change with 30 days notice, never retroactively. Competency is a vector (gold accuracy, sample size, recency, dialect, domain, overturn rate, Brier), not a years-of-experience proxy. Reviewers never pay us. Not for the qualification battery, not for calibration, not to stay on the roster. We make money when you find real defects, not when you sit an exam.

3. How items reach you

Deterministic AST checks and the Tier-2 semantic judge clear most items. You see Tier-3 items — edge cases, invariant exceptions, high-consequence vectors — and, if you are a lead adjudicator, Tier-4 disputes. Items arrive with the invariant, the generated SQL or its pseudonymised AST, the execution plan, and the counterfactual result. You rate the outcome state, name the defect family, and give a probability that the anomaly is a true Sev-1/2 defect. The probability is scored: it is your Brier score.

4. Quality

5. Confidentiality and evidence modes

In Mode 1 you never see a customer’s raw query or data; you adjudicate on sanitised evidence. In Mode 2 you see pseudonymised AST snippets with a 14-day TTL. In Mode 3 you work inside the customer’s VDI with clipboard, downloads and capture disabled; that session is regulated processing and your processing location at the time is recorded. Never copy, photograph, transcribe or discuss customer material anywhere outside the platform. Your confidentiality agreement is indefinite.

6. If something goes wrong

7. Difficult content

Warehouse SQL rarely contains distressing material, but prompts can. You may decline any item without penalty. Tell us and we re-route it.

8. What we ask of you

9. What this is not

Apply → How review is routed