Skip to content

Latest commit

 

History

535 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

ClaimBound Evidence

Where is the evidence?
Evidence cards for narrow public AI, ML and data claims: frozen protocol, hashed sources, honest result status.

PyPI Python versions conda-forge OpenSSF Best Practices OpenSSF Baseline PyPI downloads per month Total PyPI downloads tests License: Apache-2.0 Latest release Evidence cards

Terminal recording: claimbound validate-card and inspect card on a real evidence card

pipx install claimbound-evidence        # or: uv tool install claimbound-evidence
                                        # or: conda install -c conda-forge claimbound-evidence
claimbound validate-card path/to/card.json

ClaimBound turns a narrow public AI, ML, data or software-development claim into a small evidence card: protocol, source boundary, hashes, exact result status, claim boundary and reproduction level.

It is not a model leaderboard, hosted scoring service or certification authority. It is an open-source toolkit for asking one plain question:

Where is the evidence?

If there is no evidence card, the statement is still only a claim. Green means one narrow claim passed under a frozen protocol — not "trust everything". PASSED_UNDER_PROTOCOL is not the same as independently reproduced: most public cards are still SINGLE_OPERATOR / not independently reproduced until a separate rerun records otherwise. Negative, blocked and drift outcomes are first-class evidence too.

Read the one-screen walkthrough in ClaimBound in 30 seconds, the reviewer path, the non-developer start guide or the documentation index for reviewer, operator and advanced paths.

ClaimBound workflow

What It Is And Is Not

ClaimBound is ClaimBound is not
A card per one narrow claim: protocol, source boundary, hashes, result status, claim boundary, reproduction level A model leaderboard, benchmark ranking or quality score
A validator (validate-card, validate-all) that anyone can rerun locally A hosted scoring service or certification authority
Honest about negative, blocked and drifted results (they are first-class evidence) A way to declare a system safe, compliant or correct
Open registry of sanitized metadata and hashes An archive of raw payloads or private data

Nearby tools answer different questions: model and data cards describe a system as written by its authors, leaderboards rank systems, supply-chain provenance (SLSA, in-toto) covers how software was built, and fact-check markup records a publisher's verdict. ClaimBound records what one frozen, reproducible check of a public claim found, and how far anyone has independently reproduced it.

Public Workflows

Workflow Question it helps answer Start here
Public AI transparency Did a public system-card or model-card source boundary pass a frozen audit? Current evidence tracks
European and public open data Did an official public-data source pass a source-boundary or narrow empirical gate? Evidence card examples
Software development evidence Did a narrow build, validator or regression claim pass under fixed commands? Software development workflow

Blocked cards, artifact-only records and optional advanced protocol layers are documented separately. They are not hidden failures. See future applications, artifacts catalog and advanced optional docs.

Example Cards

Each SVG uses separate chips. Green and red on the result-status chip mean a narrow gate pass or honest non-pass. Yellow appears on the reproduction chip when reproduction_level records source-byte drift — not on the gate outcome itself. None of this implies general model quality.

Green ClaimBound card for Anthropic system-card source audit ClaimBound card with green gate pass and yellow reproduction chip for NASA POWER source-byte drift Red ClaimBound card for NOAA negative result

More examples, including EU source audits and blocked-source cards: evidence cards · registry index

For Reviewers And External Operators

Independent from EviBound; see scope split. Not affiliated with commercial or similarly named ClaimBound / Claimbound projects or domains (including on GitHub or GitLab); see similar names. Canonical home: GitHub ClaimBound/claimbound-evidence; the GitLab repo is a public mirror only.

Install

Works on Windows, macOS and Linux. See platform support.

From PyPI — card-level CLI (validate-card, inspect card, hash, run-root):

pipx install claimbound-evidence      # or: uv tool install claimbound-evidence / pip install claimbound-evidence
claimbound --version
claimbound validate-card path/to/card.json

validate-all, demo, rerun, drift and verify need the cards, registry and scripts of a repository clone (see below); from a pip install they print how to get one. The full guide, including a lighter sparse clone, is the documentation site.

From source (dev extras + full checkout):

git clone https://github.com/ClaimBound/claimbound-evidence.git
cd claimbound-evidence
uv sync --extra dev
uv run claimbound doctor

Reproduce In Three Commands

uv run claimbound validate-all
uv run --extra dev python -m pytest -q
uv run claimbound demo eea-source-audit

If validate-all passes, the committed cards and registry match the current validators. A scaffold or demo output is not evidence until a protocol is frozen, the run completes, the card validates and the registry is updated.

Quick Start

uv run claimbound new
uv run claimbound demo grok-source-audit
uv run claimbound run-root \
  --protocol-id EXAMPLE_D001 \
  --source-url https://example.org/source \
  --operator your-name-or-handle

Non-interactive scaffold, EU runners, card protocols and operator runbooks: getting started.

Portable AI control rules for evidence-bound work: 12 AI Life Rules.

Boundary

This repository is an independently usable open evidence foreground. It does not include or require private background technology.

Evidence cards are reusable examples and validation records, not a hosted review service, legal advice, or on-demand third-party check queue. The registry stores sanitized metadata and hashes, not raw payloads.

See governance, maintainers, architecture, security assessment, DCO, maintainer boundary and release process.

Community

License

Apache-2.0. See LICENSE.

Releases

Packages

Used by

Contributors

Languages