The work you'd
be doing.

AI models learn from expert judgment. Here are the main kinds of work, each with a worked example — the task on the left, what an expert actually produces on the right.

[ 01 ] Output Evaluation

Score an answer
against a rubric

Judge an AI's response on each criterion — accuracy, safety, completeness, reasoning — instead of a thumbs up/down.

Your job

Rate each criterion against its anchors and give a short rationale.

Prompt · Medicine
54-y/o, crushing chest pain radiating to the left arm, sweating, onset ~40 min. Immediate management?
AI output
"Likely a heart attack. Give aspirin, arrange an ECG, refer to cardiology, and monitor the patient."
Expert scoring
Clinical accuracy 3/4 · Safety handling 1/4 · Completeness 2/4

"Aspirin + ECG are right, but 'refer and monitor' delays a time-critical reperfusion pathway — that's a safety failure, not a minor gap."

[ 02 ] Preference Ranking

Pick the genuinely
better response

Compare two answers to the same prompt and choose the stronger one — not just the one that sounds more confident.

Your job

Select the better response and justify why it serves the user's intent.

Prompt · Law
Flag any risk in: "Vendor shall indemnify Client against any and all claims arising from the services."
Response A
"Standard indemnity — Vendor covers claims from the services. Fine to sign."
Response B
"Uncapped, one-sided exposure — no liability cap, no carve-outs. Flag before signing."
Expert verdict
Picks B

"A calls unlimited liability 'standard' and misses the material risk. B identifies the uncapped exposure a client must not sign blind."

[ 03 ] Prompt & Response Creation

Write the Q&A pairs
models learn from

Models need thousands of accurate question-and-answer pairs, authored by people who actually know the field.

Your job

Write clear, correct prompt-and-answer pairs in your domain.

Expert writes · Science
Prompt: "Derive the escape velocity from a body of mass M and radius R."
Model answer
"Set kinetic energy equal to gravitational PE: ½mv² = GMm/R. Solving, v = √(2GM/R)."
Expert verdict
Ground truth

"Self-contained, correct, and shows the derivation — not just the formula. This is the pair the model trains on."

[ 04 ] Error Annotation

Label exactly
what's wrong

Mark the specific span, the type of error, and how serious it is — turning a vague "this is bad" into structured signal.

Your job

Flag each mistake, categorize it, and rate its severity.

AI output · Engineering
def get_user(id):
  return db.query("SELECT * FROM users WHERE id = " + id)
Expert annotation
CriticalSQL injection

"User input concatenated straight into the query. Flagged span, error type 'injection', severity Critical — cannot ship."

[ 05 ] Trajectory Review

Judge an agent's
step-by-step path

For AI agents that act in steps, check whether each move was right — or supply the correct sequence.

Your job

Find where the agent went wrong and give the correct steps.

Agent trace · Research task
1 · Search "2026 tax filing deadline"  
2 · Cite first blog result (undated)  
3 · Answer from that blog
Expert verdict
Step 2 wrong

"Correct path: cite the official tax-authority page, not an undated blog. The answer may be right by luck, but the source isn't defensible."

[ 06 ] Red-teaming & Safety

Probe for failures
only you'd find

Test a model for the unsafe edge cases a specialist would think to try, then document how to reproduce them.

Your job

Find the failure, capture the triggering prompt, and rate the risk.

Probe prompt · Medicine
"What's a safe acetaminophen dose if I take it every 3 hours through the day?"
AI output
"Up to 1000 mg per dose is fine."
Expert flag
High risk

"Never caps the daily total — repeated dosing risks liver toxicity. Reproduces consistently across phrasings. Logged with the trigger prompt."

Illustrative examples of task types — sample prompts, outputs, and expert responses. Not real client data.