actseal 1.0.0

Actseal verifies model-chosen application actions for developers: freeze a policy, check its recorded decisions, and replay the evidence offline.

Run the packaged demonstration on macOS or Linux with uv installed. The first command resolves the latest Actseal release on PyPI for Python 3.12; no source checkout, model or API key is needed, and ./actseal-demo must not already exist.

uvx --python 3.12 actseal demo --out ./actseal-demo
uvx --offline --python 3.12 actseal replay ./actseal-demo/fixed/evidence
uvx --offline --python 3.12 actseal replay ./actseal-demo/bad/evidence

The third command intentionally exits 1: replay faithfully reproduces the deliberately bad run's BLOCK. Python and package installation time is separate from the demo's own execution. The quickstart explains each output, the exit codes and how to pin one exact release.

Guarantees

• Complete scheduled-case and required fault evidence is checked.
• PASS requires the frozen risk/coverage bounds and fault rules to pass.
• Supported evidence is recomputed offline without calling a model.

Limits

• Hashes and replay cannot authenticate coherently rewritten responses.
• They cannot prove inference occurred or that labels are true.
• Population claims require the stated sampling assumptions; Actseal does not enforce application execution.

Read a result

An ACT is a per-case decision permitted by the frozen allowlist and the selected-label probability threshold; it is not a statement that the answer is correct. Coverage is ACT decisions over all scheduled cases and risk is wrong ACT decisions over ACT decisions. A whole-run verdict bounds those rates:

lock exits 0 on success. demo exits 0 only for its expected BLOCK/PASS pair with matching replays; it never relabels the bad run. Every command accepts --json for one versioned receipt.

Scope

Actseal 1.x supports one categorical question with 2 to 16 labels, a frozen action allowlist and threshold, recorded fixture responses and an optional pinned native Laya CPU adapter. An experimental Jev cloud adapter ships as a PROVISIONAL, explicit opt-in (--provider jev --experimental-provider) with no 1.x compatibility promise and no accepted live evidence. The runtime core depends only on the Python standard library (3.12 or 3.13) on Linux and macOS; Windows is unsupported. Replay never imports a provider. The packaged demonstration is synthetic (evidence_scope=demo) and establishes no population or model-quality result. Real interpretation requires independent cases and one prespecified attempt under a fixed policy; do not retry until PASS.

Apache-2.0; see LICENSE, NOTICE and dependency notices. Contributions: CONTRIBUTING.