How evals are computed
pass@k, paired comparison, bootstrap intervals, McNemar's test, the ship-or-hold gate and judge calibration, each with its formula.
Every number on this page is computed on your machine from the stored runs. Seeded resamples give the same interval on every rerun. Terms used throughout:
| Term | Meaning |
|---|
| Case | One input, with an expected answer when there is one. A repo task is a case too. |
| Trial | One run of a case. -k sets the trials per case; the default is 3. |
| n, c | A case's trials and how many of them passed. |
| Majority outcome | A case passes when most of its trials pass. Paired comparisons and the gate use it. |
| Paired | Both runs answered the same cases, so each case is compared with itself. |
What eval run and exp show print for one run.
| Number | Formula | Notes |
|---|
| pass@1 | Σc / Σn over every trial | 95% Wilson score interval, which stays inside [0, 1] at 0/n and n/n. |
| pass@k | 1 − C(n − c, k) / C(n, k), averaged over cases | The chance that at least one of k trials drawn from a case's trials passes. Unbiased for k ≤ n. |
| pass^k | C(c, k) / C(n, k), averaged over cases | The chance that all k pass. Use it when the agent must work every time. |
| k | The fewest trials any case has | pass@k and pass^k appear only when k is 2 or more. |
| Flaky | 0 < c < n | The case passed some trials and failed others. pass^k counts it as a failure. |
| Saturated | 90% or more of the cases pass every trial | Capability suites only. The suite no longer separates versions; add harder cases. |
| Small suite | Fewer than 20 cases | The report says so: only large changes show. |
A regression suite exits 1 when any case fails most of its trials. A capability suite always exits 0; its score is expected to start low.
With exp run --seq K, a failed trial is tried again with feedback, up to K attempts.
| Number | Meaning |
|---|
| Seq@k | The share of cases solved by attempt k. |
| Seq@k against pass@k | pass@k counts k independent tries; Seq@k counts k tries where each sees why the last one failed. The two rank agents differently, so the report shows the whole curve. |
What exp compare A B and coding.sh compare print. Each case gives one paired difference dᵢ = Bᵢ − Aᵢ: its pass rate, or score, in B minus the same in A.
| Method | Formula or rule | What it answers |
|---|
| B − A | mean of dᵢ | How much B moved the pass rate. |
| Paired bootstrap interval | Resample the cases with replacement 10,000 times (seeded) and take the 2.5th and 97.5th percentiles of the mean dᵢ | The 95% range for B − A. |
| Sign-flip test | Under "no difference", each dᵢ is as likely to be +dᵢ as −dᵢ. p is the share of the 2ⁿ sign assignments whose mean is at least as extreme as the one observed | Exact up to 20 cases, a seeded Monte Carlo estimate above. It holds its level on small suites, where a bootstrap does not. |
| McNemar's test | b cases only A passed, c only B passed. p = min(1, 2 · P(X ≤ min(b, c))) with X ~ Bin(b + c, 1/2) | Exact test on the cases that changed. Cases both or neither passed say nothing about the difference. |
| Verdict | better or worse when the interval excludes 0 and the sign-flip p is below 0.05; else inconclusive | p cannot go below 0.05 with fewer than 6 changed cases, so a smaller change is always inconclusive. |
| Check | Rule |
|---|
Clusters (--cluster repo) | The bootstrap and the sign-flip test resample whole clusters, such as every case from one repo, instead of single cases. Use it when cases in a cluster are not independent. |
| Bayesian check | Below 100 paired cases: a Dirichlet posterior with a uniform prior over {only A passed, only B passed, tie}. P(B > A) is the chance that the only-B share of changed cases is above 1/2, from Beta(c + 1, b + 1). |
| Minimum detectable effect (MDE) | The smallest true gap the comparison finds 80% of the time, from the exact power of McNemar's test at this many cases. A small suite may have none: no 80% power at this n. |
Non-inferiority (--margin D) | B is at most D worse than A when the interval's lower bound is above −D and a one-sided sign-flip test of "B − A ≤ −D" has p below 0.025. Exits 1 otherwise. Use it for a cheaper or faster B. |
| Regression to the mean | When B ran only on the cases A failed, B looks better even if it changes nothing. The verdict is forced to inconclusive. |
| Interleaving | Runs that did not alternate A and B may differ because the model or the network changed between them. compare warns. |
exp compare A B --gate and compare --gate decide whether to ship B. Cases where both runs agree are dropped. Each remaining case is a win (B solves a case A failed) or a loss (B breaks a case A solved). solved is the number of cases A solved.
| Condition | Formula | Default |
|---|
| Net gain | wins − λ · losses > 0 | λ = 2 (--lambda): one break costs two fixes. |
| Significance | P(X ≥ wins) < α, with X ~ Bin(wins + losses, λ / (1 + λ)) | α = 0.05 (--gate-alpha). λ / (1 + λ) is the win share at which B only breaks even. |
| Breakage | losses / solved ≤ τ | τ = 0.10 (--tau). |
| Smallest clean win | The smallest w with (λ / (1 + λ))ʷ < α | 8 wins with no losses at the defaults. |
| Headline | When | Exit |
|---|
Ship it: B wins N cases, loses M | All three conditions hold. | 0 |
Hold: not enough evidence (…) | Any other failed condition, most often too few wins to pass the significance test. | 1 |
Don't ship: B breaks N cases A solved | B's losses beat chance: an exact one-sided binomial test of the losses among the cases that changed outcome, at p = 1/2, is under α. A group under --by or a guard set says it when B broke a case and the net gain or the breakage condition fails. | 1 |
With --lambda 1 --gate-alpha 1 --tau 1 the gate reduces to "ship when wins exceed losses".
judge calibrate <experiment> compares an LLM judge's verdicts with your labels on the same cases. Only cases in a held-out split count (test by default): a rubric written while reading the other splits agrees with you there by construction.
| Number | Formula | Notes |
|---|
| Agreement p₀ | Share of cases where judge and label agree | |
| Chance agreement pₑ | P(judge pass) · P(label pass) + P(judge fail) · P(label fail) | |
| Cohen's κ | (p₀ − pₑ) / (1 − pₑ) | 1 is perfect agreement, 0 is chance. 95% interval from 2,000 bootstrap resamples. |
| Sensitivity q₁ | P(judge says pass, given you labelled pass) | Stored with the judge. |
| Specificity q₀ | P(judge says fail, given you labelled fail) | Stored with the judge. |
| Status | Rule |
|---|
calibrated | κ ≥ 0.6 (--threshold) on at least 20 held-out cases (--min-n). |
insufficient-data | Fewer than 20 held-out cases with both a label and a verdict, or κ is undefined because you and the judge each gave every case the same label. |
uncalibrated | κ below the threshold. |
in-sample | The split you asked about is one the rubric was tuned on (tunedOn). |
| Method | Formula | Used by |
|---|
| Rogan-Gladen correction | θ̂ = (p̂ + q₀ − 1) / (q₀ + q₁ − 1), with p̂ the judge's pass rate | exp compare for a calibrated judge. Refused when q₀ + q₁ ≤ 1. The interval counts the uncertainty in q₀ and q₁ too. |
| Judge-corrected delta (PPI++) | θ = mean(y)ₙ + λ (mean(f)_N − mean(f)ₙ): the mean label on the n labelled cases, plus λ times the judge's mean on all N cases minus its mean on the labelled ones | exp compare --labels-a A.jsonl --labels-b B.jsonl, with at least 20 labels per run (--min-labels). λ is tuned so the interval is never wider than labels alone give. |
exp compare also warns when the judge's sensitivity or specificity differs between the two runs, because that biases B − A even when each run's judge looks fine.
| Term | Meaning |
|---|
| Signal validity | Whether a failure flag or tool errors predict failure beyond how hard the case is. compare --signals prints the AUROC within each task, beside difficulty and length references. Needs -k 2 or more. |
| Plausibly best set | With three or more configurations, the ones not significantly worse than the best. exp sweep prints it. |
| Item difficulty | A per-case difficulty from an item response model, P(configuration j passes case i) = σ(aᵢ (θⱼ − bᵢ)), fit on stored runs. compare --pick N --pick-by irt runs the cases that best separate the configurations first. |
Sources
- pass@k estimator: Chen et al., "Evaluating Large Language Models Trained on Code", arXiv 2107.03374.
- pass^k: Yao et al., "τ-bench", arXiv 2406.12045.
- Seq@k: Yen et al., arXiv 2610.37859.
- Small-sample intervals and the Bayesian check: Hariri et al., "Don't Pass@k", ICLR 2026.
- Sign-flip test: Fisher, *The Design of Experiments* (1935); p-values never zero: Phipson and Smyth, SAGMB 9(1), 2010.
- Non-inferiority: Piaggio et al., CONSORT extension, JAMA 308(24), 2012.
- Ship-or-hold gate: Wang et al., SAGE, arXiv 2609.36043.
- Judge correction: Lee et al. 2026, "How to Correctly Report LLM-as-a-Judge Evaluations" (Rogan-Gladen); Angelopoulos, Duchi and Zrnic, "PPI++", arXiv 2311.01453; Boyeau et al., "AutoEval Done Right", arXiv 2403.07008.