Skip to content
mentorship

Post-training research work sample

Post-training environment and grader lab

A scalar reward is an attack surface. Separate task success, policy compliance, and process evidence before training optimizes the wrong behavior.

60 minutes plus 20-minute research discussionBroken structured grader, adversarial episodes, environment-design readout

Do this only if the format matches

Confirm the round with recruiting first. This is an original practice exercise, not a claim that a specific company uses this exact prompt.

Protocol

  1. State the capability, prohibited behavior, state, action space, and terminal condition.
  2. Run tests and identify how outcome quality hides process violations.
  3. Implement hard disqualification gates and bounded component scores.
  4. Add adversarial episodes for reward hacking and grader blind spots.
  5. Separate deterministic checks, model-graded judgments, and human review.
  6. Explain grader versioning, drift, and online validation.

Deliverables

  • A repaired structured grader.
  • Two new adversarial episodes.
  • An environment and coverage specification.
  • A decision rule for whether the training intervention helped.

Graduation gates

01

Unsafe process cannot be offset by final-answer quality.

02

Evidence accompanies every gate and penalty.

03

The environment has explicit reset, timeout, and terminal semantics.

04

Grader and environment versions are part of every result.

Starter files

Answer key is a rubric, not a solution dump

The linked question explains what strong performance looks like, common failure modes, and senior follow-ups. Attempt the lab before reading it.

Read the post-training design rubric →

← All executable labs