Actseal

Decision contracts you can replay.

If your application uses a model to choose an action, a confidence score alone does not tell you when that action is justified. Actseal freezes your policy, checks its errors and coverage on labelled cases, exercises provider failures, and saves evidence that you can replay without the model.

Try the packaged demo

With uv installed, run this on macOS or Linux. It selects Python 3.12 and installs the GitHub release wheel in uv's tool environment; no source checkout, model or API key is needed.

uvx --python 3.12 --from https://github.com/ajaysurya1221/actseal/releases/download/v0.1.0/actseal-0.1.0-py3-none-any.whl actseal demo --out ./actseal-demo

Choose an output directory that does not already exist. Its parent must exist; Actseal refuses to overwrite an existing destination. Use a different directory name for another demo run.

The demo evaluates the same policy against two authored sets of answers. The independently tested wheel produced bad: BLOCK, fixed: PASS, with both freshly replayed. The overall demo exits 0 only when those conditions hold. Here “fixed” names the fixture whose authored answers match the labels; no model was trained or repaired.

actseal-demo/
  inputs/          # copied contracts, labelled cases and recorded answers
  bad/lock.json
  bad/evidence/    # deliberately wrong authored answers
  fixed/lock.json
  fixed/evidence/  # authored answers that match the labels

The installed demo took 0.17 seconds on the reference Mac after environment preparation. Both bundles replayed correctly, and all eight Linux/macOS CI jobs produced identical evidence. The quickstart records exact bounds and measurement scope. First-time uv/Python/wheel downloads are separate; this is a synthetic fixture measurement, not an inference benchmark.

Replay an individual bundle in a separate command:

uvx --python 3.12 --from https://github.com/ajaysurya1221/actseal/releases/download/v0.1.0/actseal-0.1.0-py3-none-any.whl actseal replay ./actseal-demo/fixed/evidence

Replay reconstructs the policy decisions and verdict from recorded data; it does not call a provider. Replaying bad/evidence should return BLOCK and exit 1, even when replay correctly reproduces the result. See the detailed quickstart for JSON output, individual lock/verify commands and checking an externally obtained lock digest.

Read the result

An ACT is a decision permitted by the frozen allowlist and selected-label probability threshold. Coverage is ACT decisions divided by all scheduled cases. Risk is wrong ACT decisions divided by ACT decisions. Actseal reports bounds around those rates; a high score or zero observed errors alone is not PASS.

lock also returns 0 on success. demo returns 0 for its expected BLOCK/PASS pair plus matching replays; it does not turn the bad run into PASS. Append --json to a command for structured output. Failures and warnings are surfaced.

Use your own policy

v0.1.0 supports one categorical question with 2–16 labels, a frozen action allowlist and threshold, recorded fixtures, and an optional pinned native Laya CPU adapter. Start with the quickstart, statistical contract and provider setup and limitations. The core and replay require no model package or hosted service. Jev is deferred, not a v0.1.0 dependency.

The native Laya CLI check at candidate 434c682 also preserved an honest result: all 128 synthetic cases ABSTAINed at the fixed threshold, with no provider failures. Verification and replay both returned BLOCK for insufficient coverage; no accepted cases means risk remains unestimated, [0,1]. The policy was not tuned or retried to obtain PASS. Provider details record the environment, warnings and bounds; this is an integration check, not a population or model-quality benchmark.

The demonstration is synthetic, marked evidence_scope=demo, and establishes no population or model-quality result. For real data, the statistical interpretation requires independent cases and one prespecified attempt under a fixed policy and operating regime. Actseal does not enforce another application's behavior.

Hashes and replay check consistency. Even a trusted lock digest cannot authenticate rewritten responses under that lock, prove inference occurred or prove the labels true. Read the threat model before using evidence to authorize actions.

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