Updated 2026-11-02: version 0.2.0 adds logical precision and recall bounds on the finite input set. The original 0.1.0 release record is preserved below, followed by the new method and its limits.
0.1.0: original release record
I released Rejected Check Audit, an offline tool for auditing denominators in multi-stage filters. It retains events rejected by earlier gates, keeps unknown labels separate from negative labels, and traces known misses to the stage where each event stopped.
Reviewing only the final releases can reveal errors among those releases. It does not reveal how many relevant events were rejected. Treating unreviewed events as negatives, or inventing downstream decisions for stages that never ran, changes the meaning of the reported metrics.
The repository contains twelve author-constructed events in three labeling views: complete labels, released-only labels, and a purpose-selected review subset. These are the same twelve identities, not thirty-six independent observations. When only releases are labeled, the auditor retains the original denominator and returns null for overall recall with an explanation. It does not extrapolate a selected review subset to the whole population.
The tool runs locally with the Python standard library and exports JSON plus stage, path and metric CSV files. A separate verifier, a frozen hand-counted contract and a CPU notebook are included. The final publication package passed fresh-directory execution and additional independent state-combination checks. Published HF file contents and the three searchable dataset views were also checked.
These results validate accounting behavior. They are not model-quality measurements or evidence that a filter improved. There are no production traces, model responses, inference calls or training runs. Real use still requires a defensible event identity, labeling rule, sampling design and stage semantics. Recall uses a conservative complete-label rule; sampling weights, population confidence intervals and partial-identification bounds are outside scope.
Dataset and code: https://huggingface.co/datasets/WideChaos/rejected-check-audit
Code and synthetic fixtures are licensed under Apache-2.0. By Renchao Wu.
2026-11-02: logical bounds in 0.2.0
I added exact logical bounds for precision and recall on the finite input set. Known labels and decisions stay fixed; unknown labels and unresolved predictions may take either binary value. Each metric is minimized and maximized over completions with a positive denominator, with explicit flags for possibly or always undefined metrics. Each endpoint is individually attainable; the precision and recall endpoints are not guaranteed to be jointly attainable in one completion.
The same twelve synthetic events and three labeling views remain unchanged. Recall bounds are [1/3,1/3] with complete labels, [1/5,1] with released-only labels, and [2/9,2/3] after reviewing two additional events. Precision is [1/2,1/2] in all three views. Existing point-value and null eligibility rules remain unchanged; the original recall field stays null with incomplete labels.
These are logical constraints on the input set under the stated completion model, not population estimates, sampling confidence intervals, or model performance results. The model does not guarantee that every completion is physically possible in a real pipeline. The original examples and limitations above remain part of the version history; updated code and proof are in the same HF repository at the fixed revision.
The HF release check verified the files and data rows in all three configurations; search was still unavailable at that check.
2026-11-06: browser entry
The same work now runs as a static page: the original Space uses sdk static, and the standalone entry is Open browser tool. Choose a synthetic example or edit events, labels and stages in the JSON, press run, and the page recomputes; you can download that run's input, full report and four CSV tables.
The Python tool, mathematical contract and twelve synthetic examples are unchanged. browser-core.js on the page implements the same contract in JavaScript and runs in the browser. The Static adapter is version 0.2.0.
Input limits are 128 KiB UTF-8, 64 events and 1 to 8 stages, with 512 Unicode code points per text field. Initialization and example selection request only fixed same-origin files. In the reviewed code and current deployment, audit input and computation stay local; hosting infrastructure and browser extensions have their own behavior. Editing input or changing examples clears old results and download links, and refreshing clears edited work. Downloads contain your identifiers, reasons and review text, so check them before sharing.
JSON exports preserve parsed values, while layout and number spelling may differ from Python exports. CSV output matches Python when compared as parsed fields; the difference sits in serialization and quoting, where some Python runtimes may leave a bare CR unquoted under an LF terminator while the JavaScript implementation quotes it safely. The bytes can differ; the parsed field value does not. CSV text starting with a formula marker gets an apostrophe, and JSON keeps the original string.
Originally published on widechaos.cn.
Top comments (0)