PQID-Bench User Manual¶
This manual guides a user from installation through scientific reproduction,
live model collection, candidate comparison, isolated replay, numerical
reporting, interactive exploration, troubleshooting, and citation. It
documents pqid-bench 1.2.1 as implemented against the frozen PQID-Bench
v1.0.0 evidence release.
Manual Chapters¶
- Installation
- Core Concepts
- Capability Guide
- Live Model Testing
- CLI Reference
- Reporting And Numerical Exports
- Interactive Explorer
- Python API
- Data Dictionary
- Metrics And Invariants
- End-To-End Workflows
- Reproducibility Contract
- Security And Governance
- Troubleshooting
- Glossary
Reading Paths¶
First-time user
Read Installation, Core Concepts, and Capability Guide, then complete Workflow 1 in the workflow chapter.
Paper reviewer
Use the Reviewer Quickstart, followed by the Reproducibility Contract when auditing claims and denominators.
Researcher with new model outputs
Read Live Model Testing and the evaluation and run-manifest portions of the Data Dictionary, then the candidate-summary and comparison workflows. Use Reporting And Numerical Exports when moving results into R, pandas, or a manuscript.
Package integrator
Read the Reporting And Numerical Exports, Python API, Data Dictionary, Metrics And Invariants, and Security And Governance chapters.
Maintainer or software-paper author
Read the complete manual and the Package Paper Blueprint.
Conventions¶
RELEASE_DIRmeans the root of the extracted frozen evidence bundle. It is the directory containingARTIFACT_MANIFEST.tsv.EVALUATIONS.jsonlmeans newline-delimited canonical evaluation objects.RESPONSES.jsonlmeans archived canonical model-response objects accepted by the replay harness.- Shell examples use
\for line continuation. In PowerShell, place the command on one line or replace\with the PowerShell backtick. - Counts and rates in this manual refer to the frozen version 1.0.0 release unless a candidate or matched subset is explicitly named.
Scope¶
The manual covers offline verification, reproduction, live OpenAI-compatible collection, summarization, comparison, and archived or newly collected code replay. Live calls always create a new replication rather than reproducing archived model text. Native non-compatible provider protocols and automatic cost estimation remain outside package version 1.2.1.