1 min lesson
Quantify uncertainty
Name the parts in "Quantify uncertainty" and give the practical job of each one.
Step 1 of 2
Quantify uncertaintyone run is not a result
RL training is high-variance: seed, data shuffle and rollout sampling can swing a SWE-bench-style number by more than the effect you're chasing. A single lucky run is a story, not a finding. Report across seeds with a spread and ask whether your claimed delta clears the noise floor.
- Multiple seeds
- Repeat the key comparison across several seeds and report mean and spread, not a single point.
- Noise floor first
- Estimate run-to-run variance under no change, so you know how big a delta has to be to mean anything.
- Eval variance
- Small or contaminated eval sets are noisy on their own. Account for benchmark variance, not just training seed variance.
Beware metrics that move for the wrong reason. A SWE-bench jump can come from train/eval contamination, a grader the model learned to game or a harness that silently caught more partial credit. Reward hacking looks exactly like progress on the dashboard. Before you celebrate, ask what else could produce this number and go check.
Learn more
Full explanation
Leakage and reward hacking
Leakage and reward hackingthe two ways a coding eval lies
- Failure
- Train/eval contamination
- What it looks like
- Suspiciously high eval; gains don't transfer to live sessions
- How you catch it
- Dedup by repo/problem; hold out truly unseen tasks; check overlap
- Failure
- Reward hacking
- What it looks like
- Reward climbs, real quality doesn't
- How you catch it
- Inspect winning trajectories by hand; audit the grader on adversarial cases
- Failure
- Wrong-reason metric
- What it looks like
- A number moves with no plausible mechanism
- How you catch it
- Trace the cause; reproduce the delta on a controlled slice
| Failure | What it looks like | How you catch it |
|---|---|---|
| Train/eval contamination | Suspiciously high eval; gains don't transfer to live sessions | Dedup by repo/problem; hold out truly unseen tasks; check overlap |
| Reward hacking | Reward climbs, real quality doesn't | Inspect winning trajectories by hand; audit the grader on adversarial cases |
| Wrong-reason metric | A number moves with no plausible mechanism | Trace the cause; reproduce the delta on a controlled slice |
Offline eval can rise while shipped impact flatlines - always close the loop against real Cursor usage.
Learn more
Optional practice
Practice: Quantify uncertainty
QYour new grader raises an offline coding benchmark by 2 points, but live edit-accept rate in real Cursor sessions doesn't budge. What are the leading explanations and what do you do?