# Read, explore, or reproduce

> Choose the level of work you need. Reading and browser exercises require no Python or model service.

[Canonical HTML page](https://trustcyber.ca/research/allow-hold-deny/reproduction/)

- Author: Junior Williams
- Type: Research companion addition

## In the browser

Every chapter is a native TrustCyber page. The eight teaching tools use small, local JavaScript modules and the selected page’s data. They make no model calls. Policy inputs and consequence weights are hypothetical; the accuracy, confusion, case, and timing tools inspect reported or clearly attributed derived observations. The Brier calculator works through rounded percentage examples.

Each chapter loads independently. The original PDF and figure files are fetched when you open them; the entire standalone reader is not downloaded in the background.

## Learn through browser activities

The [guided walkthrough](/research/allow-hold-deny/lessons/01-ai-and-agents/) explains the paper in order, beginning with AI and statistics basics. Worked examples and interactive activities teach the policy, validity, accuracy, uncertainty, timing, probability quality and comparisons. Teaching inputs remain separate from historical observations.

## Original evidence and analysis

The [public evidence deposit](https://doi.org/10.5281/zenodo.22998941) contains the frozen reviewed cases, original response bytes, manifests, dependency details, and original analysis code. The v1.0.0 archive is approximately 16.2 MB. It is separate from the companion teaching activities.

For this reader, a read-only independent audit checked all 9,385 manifest entries and all 1,200 scored response records. The local archive’s MD5 matched the public deposit listing. Accuracy, coverage, balanced accuracy, Wilson intervals, Brier scores, skill, AUROC, reliability counts, timing quantiles, planned paired intervals, disagreement rows, and changed-reference decisions matched the manuscript’s reported precision. This verifies arithmetic and transcription; it does not independently validate the references or repeat inference.

For full original analysis reproduction, first verify the archive and its manifest, then inspect the original schema and analysis source. Keep the frozen key, invalid outcomes, case alignment, original decimal spellings, and declared probability tolerance. Recompute from recorded observations and compare resulting tables and figures. Running recorded-data analysis requires no paid inference. Repeating model calls is a distinct experiment, and hosted outputs are not guaranteed to repeat.

## Original paper

[Open the unchanged 11-page PDF](/research/allow-hold-deny/original.pdf) · [Manuscript DOI](https://doi.org/10.5281/zenodo.23005044) · [Figures and tables](/research/allow-hold-deny/figures/)

Paper SHA-256: `c2cd849482ec027c7079929b9ef85a57f3b7b75a4d46970f96f87f23da66524d`.

Evidence archive SHA-256: `9a8a4af8e98ed737351b99b0ae84939c5d5d2367202054cc7903ca7765c80b46`.

## Companion navigation

- [Overview](https://trustcyber.ca/research/allow-hold-deny/index.md)
- [Both authorization studies](https://trustcyber.ca/research/authorization/index.md)