passwedge 0.0.1a2

Reliability-science metrics for repeated-attempt evaluation of long-horizon LLM agents: pass@k, pass^k, Bayesian posteriors, RDC/VAF/GDS/MOP.

passwedge

Reliability-science metrics for repeated-attempt evaluation of long-horizon LLM agents.

pass@1 tells you whether a model can do a task once. It says nothing about whether an agent does so consistently across repeated attempts — the property that actually matters when a task takes hundreds of tool calls and one slip ends the run. passwedge measures that gap: capability vs. reliability.

• Pure Python (numpy + scipy only), CPU-only, no GPU, no network, no API keys.
• Every metric is implemented exactly as defined in its source paper, with the equation and citation in docs/DEFINITIONS.md. Where a paper gives a concept without one canonical estimator, passwedge documents its operationalization and never presents it as a verbatim reproduction.
• Ships three ways: a library, a pytest plugin (@pytest.mark.passk), and a GitHub Action that comments a reliability table on your PRs.

Alpha (0.0.1a1). The API may change before 0.1.0.

Install

pip install passwedge

Quickstart (library)

import passwedge as pw

# 12 repeated attempts at one task; True = the attempt passed.
trial = pw.coerce_trial([True, True, False, True, True, False, True, True, True, False, True, True])

print("pass@1 :", pw.pass_at_k(trial.n, trial.c, 1))   # >=1 of 1 succeeds
print("pass@5 :", pw.pass_at_k(trial.n, trial.c, 5))   # >=1 of 5 succeeds
print("pass^5 :", pw.pass_pow_k(trial.n, trial.c, 5))  # ALL 5 succeed (reliability)

# Bayesian posterior over the task's true success probability (Jeffreys prior).
post = pw.beta_posterior(trial.c, trial.n)
print("posterior mean :", round(post.mean(), 3))
print("95% credible   :", tuple(round(x, 3) for x in post.credible_interval(0.95)))
print("E[p^5]         :", round(post.expected_pow_k(5), 3))  # expected all-5-success

Quickstart (CI gate / GitHub Action)

Given a JSON file of task outcomes:

[
  {"task_id": "t1", "duration_bucket": "short", "outcomes": [true, true, true, false]},
  {"task_id": "t2", "duration_bucket": "long",  "outcomes": [true, false, false, false]}
]

passwedge ci --input trials.json --k 1,2 --metric pass_pow_k --fail-under 0.5

prints a Markdown report, emits $GITHUB_OUTPUT values, and exits non-zero if the chosen metric is below the threshold. In a workflow, use the bundled action:

- uses: hinanohart/passwedge@v0.0.1a2
  with:
    input: trials.json
    fail-under: "0.5"
    metric: pass_pow_k

Quickstart (pytest plugin)

import pytest

@pytest.mark.passk(attempts=20, k=5, min_pass_pow_k=0.9)
def test_agent_is_reliable():
    assert run_agent().solved   # executed 20×; passes iff pass^5 >= 0.9

Metrics

Metric	Meaning	Source
pass_at_k	probability ≥1 of k attempts succeeds	Chen et al. 2021 (arXiv:2107.03374)
pass_pow_k	probability all k attempts succeed	Beyond pass@1 (arXiv:2603.29231), Def. 2
reliability_decay_curve / _slope	how pass^k decays with task duration (RDC/RDS)	arXiv:2603.29231, Def. 3
variance_amplification_factor	VAF: variance ratio long-vs-short bucket	arXiv:2603.29231, Def. 4
graceful_degradation_score	GDS: weighted partial credit over subtasks	arXiv:2603.29231, Def. 5
meltdown_onset_point	MOP: tool-call entropy collapse detector	arXiv:2603.29231, Def. 6
beta_posterior / dirichlet_posterior	Bayesian posterior mean + credible interval, E[p^k]	Don't Pass@k (arXiv:2510.04265)

Honest-marketing note. arXiv:2510.04265 does not define a metric called "Bayes@k"; passwedge's Bayesian helpers are our operationalization of that paper's Dirichlet framework. The scorio package (the paper's reference implementation) is the numeric baseline our test suite reproduces under a uniform prior.

MOP thresholds are dataset-specific. meltdown_onset_point requires explicit theta_h, delta, w (no defaults). The paper's calibration is exposed as MOP_PAPER_DEFAULTS but applying it verbatim to other data produces false positives.

Where it fits

passwedge is the measurement layer: it consumes repeated-attempt outcomes (a bool list, a JSON file, or a trace export) and reports reliability. It deliberately does not run rollouts, score reward/fitness functions, audit reward gameability, or detect reward hacking — those are separate concerns handled by tools such as scorewright (fitness scoring + anti-gaming), rewardfuzz (reward gameability auditing), and mav-bench (multi-agent verification). passwedge sits one layer up from those: feed it the per-attempt pass/fail outcomes they (or any eval harness) produce, and it tells you how reliably the agent succeeds. For confidence / hallucination fragility see yuragi; for streaming inference verification see conformlock.

License

Apache-2.0.