{"id":50349821,"url":"https://github.com/hinanohart/poolcheck","last_synced_at":"2026-05-29T21:00:31.808Z","repository":{"id":361073327,"uuid":"1252985622","full_name":"hinanohart/poolcheck","owner":"hinanohart","description":"Combinatorial group testing for verifier-cost-limited reasoning: localize defective reasoning steps from ~k*log(n/k) pooled verifier queries under an explicit asymmetric noise model (simulated headline; pre-alpha).","archived":false,"fork":false,"pushed_at":"2026-05-29T04:08:30.000Z","size":226,"stargazers_count":0,"open_issues_count":0,"forks_count":0,"subscribers_count":0,"default_branch":"main","last_synced_at":"2026-05-29T06:07:11.854Z","etag":null,"topics":["chain-of-thought","group-testing","llm","process-reward","reasoning","test-time-compute","verifier"],"latest_commit_sha":null,"homepage":null,"language":"Python","has_issues":true,"has_wiki":null,"has_pages":null,"mirror_url":null,"source_name":null,"license":"apache-2.0","status":null,"scm":"git","pull_requests_enabled":true,"icon_url":"https://github.com/hinanohart.png","metadata":{"files":{"readme":"README.md","changelog":null,"contributing":null,"funding":null,"license":"LICENSE","code_of_conduct":null,"threat_model":null,"audit":null,"citation":null,"codeowners":null,"security":null,"support":null,"governance":null,"roadmap":null,"authors":null,"dei":null,"publiccode":null,"codemeta":null,"zenodo":null,"notice":"NOTICE","maintainers":null,"copyright":null,"agents":null,"dco":null,"cla":null}},"created_at":"2026-05-29T03:58:34.000Z","updated_at":"2026-05-29T04:01:38.000Z","dependencies_parsed_at":null,"dependency_job_id":null,"html_url":"https://github.com/hinanohart/poolcheck","commit_stats":null,"previous_names":["hinanohart/poolcheck"],"tags_count":1,"template":false,"template_full_name":null,"purl":"pkg:github/hinanohart/poolcheck","repository_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/hinanohart%2Fpoolcheck","tags_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/hinanohart%2Fpoolcheck/tags","releases_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/hinanohart%2Fpoolcheck/releases","manifests_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/hinanohart%2Fpoolcheck/manifests","owner_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/owners/hinanohart","download_url":"https://codeload.github.com/hinanohart/poolcheck/tar.gz/refs/heads/main","sbom_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/hinanohart%2Fpoolcheck/sbom","scorecard":null,"host":{"name":"GitHub","url":"https://github.com","kind":"github","repositories_count":286080680,"owners_count":33670211,"icon_url":"https://github.com/github.png","version":null,"created_at":"2022-05-30T11:31:42.601Z","updated_at":"2026-05-26T15:22:16.424Z","status":"online","status_checked_at":"2026-05-29T02:00:06.066Z","response_time":107,"last_error":null,"robots_txt_status":"success","robots_txt_updated_at":"2025-07-24T06:49:26.215Z","robots_txt_url":"https://github.com/robots.txt","online":true,"can_crawl_api":true,"host_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub","repositories_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories","repository_names_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repository_names","owners_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/owners"}},"keywords":["chain-of-thought","group-testing","llm","process-reward","reasoning","test-time-compute","verifier"],"created_at":"2026-05-29T21:00:18.873Z","updated_at":"2026-05-29T21:00:31.795Z","avatar_url":"https://github.com/hinanohart.png","language":"Python","funding_links":[],"categories":[],"sub_categories":[],"readme":"# poolcheck\n\n**Combinatorial group testing for verifier-cost-limited reasoning.**\n\nWhen you verify a chain of reasoning, the expensive part is the *verifier* (an\nLLM judge, a process reward model, a test suite). poolcheck spends fewer verifier\ncalls by **pooling** items into a single query — \"does *any* step in this pool\ncontain a mistake?\" — and **decoding** the answers to localize the defective\nitems, instead of checking each of the `n` items one at a time.\n\nFor `k` defectives among `n` items, the number of pooled queries can be\n`~k·log(n/k)` instead of `n` — *under an explicit noise model* that this library\nmakes you state.\n\n\u003e **Status: `v0.1.0a1` (pre-alpha).** The decoders are correct and tested; the\n\u003e headline numbers come from a **simulated** noise model, not a deployed judge.\n\u003e See [Claims](#claims-and-non-claims) and [`s0_report.json`](s0_report.json)\n\u003e before relying on anything here.\n\n---\n\n## Install\n\n```bash\npip install poolcheck            # core (numpy + scipy only)\npip install 'poolcheck[hf]'      # + Hugging Face Inference verifier\n```\n\n## Quickstart\n\n```python\nimport numpy as np\nfrom poolcheck import ItemSet, NoiseChannel, SimulatedJudge, localize\n\n# 8 chain-of-thought steps; the faulty one is step 5 (unknown to the decoder).\nitems = ItemSet.from_cot([f\"step {i}\" for i in range(8)])\n\n# A judge that misses real errors 40% of the time and false-alarms 10%\n# (a lenient operating point grounded in the LLM-as-judge literature).\nnoise = NoiseChannel(alpha_fa=0.10, beta_md=0.40)\njudge = SimulatedJudge(truth={5}, noise=noise, n=8, rng=np.random.default_rng(0))\n\nresult = localize(items, judge, budget=12, noise=noise, k=1, rng=np.random.default_rng(0))\nprint(result.defectives)   # -\u003e localized faulty step(s)\n```\n\nMeasure the simulated budget → accuracy frontier from the CLI:\n\n```bash\npoolcheck frontier --n 32 --k 1 --alpha 0.1 --beta 0.4\n```\n\n## How it works\n\n1. **Design** a pooling (test) matrix — which items go in which pooled query.\n   Default is a near-constant-column-weight design (outperforms i.i.d. Bernoulli,\n   arXiv:1612.07122); a deterministic Reed-Solomon (Kautz-Singleton) `d`-disjunct\n   design is also provided.\n2. **Query** the verifier once per pool.\n3. **Decode** the defective set. Noiseless: COMP / DD / SCOMP. Noisy: a per-item\n   separate-decoding **log-likelihood-ratio** decoder tuned to the channel's\n   asymmetric false-alarm (`alpha_fa`) and missed-detect (`beta_md`) rates. A\n   threshold trades precision for recall.\n\nThe core (`design`, `decode`, `adaptive`, `noise`, `frontier`) never imports a\nverifier and never touches the network — it is fully deterministic given a seed.\n\n## Headline (simulated)\n\n\u003c!-- BEGIN GENERATED:headline --\u003e\n_Generated by `scripts/measure.py` (seed=0, n=32, k=1, 300 trials). Simulated noise model only — not a deployed-judge benchmark._\n\n**noiseless** (alpha_fa=0.0, beta_md=0.0)  \nPer-item baseline: **32 queries**, F1=1.000 (95% CI [1.000, 1.000])\n\n| pooled queries | F1 | 95% CI | vs per-item baseline |\n|---:|---:|:---:|:---:|\n| 5 | 0.027 | [0.010, 0.047] | worse |\n| 8 | 0.323 | [0.270, 0.377] | worse |\n| 12 | 0.923 | [0.893, 0.950] | worse |\n| 16 | 0.990 | [0.977, 1.000] | matches (fewer queries) |\n| 24 | 1.000 | [1.000, 1.000] | matches (fewer queries) |\n\n**lenient_judge** (alpha_fa=0.1, beta_md=0.4)  \nPer-item baseline: **32 queries**, F1=0.264 (95% CI [0.236, 0.292])\n\n| pooled queries | F1 | 95% CI | vs per-item baseline |\n|---:|---:|:---:|:---:|\n| 5 | 0.100 | [0.067, 0.133] | worse |\n| 8 | 0.130 | [0.093, 0.170] | worse |\n| 12 | 0.270 | [0.220, 0.323] | matches (fewer queries) |\n| 16 | 0.387 | [0.330, 0.443] | **better** (fewer queries) |\n| 24 | 0.580 | [0.523, 0.637] | **better** (fewer queries) |\n\n_\"vs per-item baseline\" compares the pooled F1 bootstrap CI to the baseline CI (which uses n per-item queries): non-overlapping above = **better**, overlapping = statistically indistinguishable (**matches**), non-overlapping below = worse. Pooled always uses fewer queries than the baseline._\n\u003c!-- END GENERATED:headline --\u003e\n\n## Claims and non-claims\n\n**CLAIM.** poolcheck decodes a defective set (e.g. the first faulty CoT step)\nfrom `~k·log(n/k)` pooled verifier queries instead of `n` per-item queries,\n**under an explicit asymmetric false-alarm / missed-detect noise model**. The\nheadline budget → accuracy frontier is computed with a *simulated* judge oracle\n(deterministic, no API key, reproducible via `scripts/measure.py`); its noise\nparameters are grounded in published single-item LLM-judge error rates\n(see [`s0_report.json`](s0_report.json)).\n\n**NON-CLAIMS.**\n\n- **No speedup is claimed for any specific deployed model.** Every headline\n  number is from the simulated noise model above, not a live-judge benchmark.\n- **The \"1-good-among-N\" regime is explicitly out of scope.** Picking the single\n  good answer out of `N` candidates is the `k = N-1` regime, where group testing\n  provably **cannot** beat individual testing (`Θ(n)` tests when `k = ω(n/log n)`,\n  arXiv:2006.01325, arXiv:2106.06878). `ItemSet.from_candidates` is a\n  `k \u003c\u003c N` shortlist-narrowing convenience only — it is **not** the headline.\n- **poolcheck is not a process reward model.** It does not *score* steps; it\n  *localizes* defectives. `ItemSet.priors` is an unused experimental seam (off by\n  default); supplying priors lets a downstream PRM bias the decoder, but that path\n  is unbenchmarked here.\n- **The pooled-query premise is unverified in this release.** The simulated noise\n  uses *single-item* literature rates and assumes they also hold for *pooled*\n  queries. Whether pooling degrades a real judge's FA/MD (residual risk #1) is\n  **OPEN** — see below.\n\n## Did pooling break my judge? (`s0-gate`)\n\nThe one empirical question poolcheck cannot answer for you: does asking your judge\n\"are *any* of these N steps wrong?\" make it noticeably worse than asking about one\nstep at a time? Measure it on your own judge:\n\n```bash\npoolcheck s0-gate --cases your_cases.json --verifier hf:Qwen/Qwen2.5-7B-Instruct \\\n    --pool-sizes 4 8\n```\n\n`PASS` (pooled FA/MD ≤ 1.5× single, with bootstrap CIs) means the simulated\nadvantage should transfer. `FAIL` means it may not. This build ships the tool but\n**did not run it against a live judge** (no inference token was available in the\nbuild environment); see [`s0_report.json`](s0_report.json).\n\n## Public API\n\n`Verifier` · `ItemSet` · `localize` · `Strategy` · `frontier`\n(plus supporting types `NoiseChannel`, `SimulatedJudge`, `DeterministicJudge`,\n`LocalizeResult`).\n\n## License\n\n[Apache-2.0](LICENSE). © 2026 the poolcheck authors.\n","project_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2Fhinanohart%2Fpoolcheck","html_url":"https://awesome.ecosyste.ms/projects/github.com%2Fhinanohart%2Fpoolcheck","lists_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2Fhinanohart%2Fpoolcheck/lists"}