falsify-giskard

PRML pre-registration for Giskard scenario results.

PRML v0.1 License: MIT

Commit a Giskard eval claim (a metric and threshold) to a SHA-256 before the run, then verify the realised ScenarioResult against it.

Why

Giskard runs a scenario of checks and reports pass/fail plus per-check metrics. But the report records what happened, not what was promised before the run. Pre-registering the claim means quietly relaxing a threshold or swapping a model after seeing results breaks the hash, so a passing scenario becomes tamper-evident.

This is the Giskard counterpart to falsify-inspect and uses the same PRML v0.1 manifest format.

Install

pip install falsify-giskard

Quickstart

from falsify_giskard import preregister, verify_scenario_result

# 1. Before the run — commit the claim (real PRML v0.1 manifest)
h, manifest = preregister(
    metric="pass_rate",            # or the name of a Giskard Metric (e.g. "groundedness")
    threshold=0.9,
    threshold_direction=">=",      # PRML `comparator`
    dataset="support-qa-v1",       # PRML `dataset.id`
    dataset_hash="e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855",  # PRML `dataset.hash`, 64 lowercase hex
    seed=42,
    producer_id="acme/eval-bot",   # PRML `producer.id`
    giskard_scenario="grounded-answers",
    output_path="grounded.prml.yaml",
)
print(h)  # 64-char canonical PRML SHA-256

# 2. Run your Giskard scenario as usual
result = await scenario.run()

# 3. After the run — verify (pass the lock-time hash for tamper-evidence)
verdict = verify_scenario_result(result, "grounded.prml.yaml", expected_hash=h)
assert verdict["status"] == "PASS"   # PASS / FAIL / TAMPERED

Metrics

  • metric="pass_rate" verifies the fraction of (non-skipped) checks that passed.
  • metric="<name>" verifies the value of a Giskard Metric with that name (for example a semantic_similarity or groundedness score).

Verdicts

  • PASS — the manifest hash matches and the observed metric satisfies the committed threshold.
  • FAIL — the hash matches but the observed metric misses the threshold.
  • TAMPERED — the manifest no longer matches the lock-time hash passed as expected_hash (it was altered after commit). Without expected_hash, the hash is taken from the file itself, so only PASS/FAIL is evaluated — pass expected_hash for real tamper-evidence.

For durable tamper-evidence, commit the returned hash to git or to the public registry at registry.falsify.dev right after pre-registration, so the claim is anchored somewhere immutable.

License

MIT. The PRML specification is CC BY 4.0.

Project links

Data verified by PyPI on Jul 30, 2026
Data provided by the project maintainers, verified at the time the release was uploaded to PyPI.

Key dates

PyPI data
Data sourced directly from PyPI's database.
  • Released:
Latest release

1 maintainer

PyPI data
Data sourced directly from PyPI's database.
Avatar for falsify from gravatar.com falsify

Credits

Author: Cüneyt Öztürk

GitHub Statistics

Data verified by PyPI on Jul 30, 2026
The GitHub source repository was provided by the project maintainers and verified by PyPI at the time of upload. Stars, forks, and open issues/PRs are derived from that repository and have not been independently verified.

License

MIT License (MIT)

Requires

Python >=3.12

Provides Extra

dev giskard

Tags

ai-safety evaluation giskard llm ml pre-registration prml reproducibility

Classifiers

Report project as malware

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

falsify_giskard-0.3.0.tar.gz (13.0 kB view details)

Uploaded Source

Built Distribution

If you're not sure about the file name format, learn more about wheel file names.

falsify_giskard-0.3.0-py3-none-any.whl (9.4 kB view details)

Uploaded Python 3

File details

Details for the file falsify_giskard-0.3.0.tar.gz.

File metadata

  • Download URL: falsify_giskard-0.3.0.tar.gz
  • Upload date:
  • Size: 13.0 kB
  • Tags: Source
  • Uploaded using Trusted Publishing? Yes
  • Uploaded via: twine/7.0.0 CPython/3.13.14

File hashes

Hashes for falsify_giskard-0.3.0.tar.gz
Algorithm Hash digest
SHA256 a9e4c86621f433a304110009f03d1ee1552ddc932d3dcffd642c2f6b53c7c336
MD5 d257605d8d4f94c78671c4d637376d67
BLAKE2b-256 415681c8f3b09483780efc19dbb10c90dd2d6787846d209009759c6ef78db358

See more details on using hashes here.

Provenance

The following attestation bundles were made for falsify_giskard-0.3.0.tar.gz:

Publisher: publish.yml on studio-11-co/falsify-giskard

Attestations: Values shown here reflect the state when the release was signed and may no longer be current.

File details

Details for the file falsify_giskard-0.3.0-py3-none-any.whl.

File metadata

File hashes

Hashes for falsify_giskard-0.3.0-py3-none-any.whl
Algorithm Hash digest
SHA256 ce793f5b41e36ba3a2bd663bddd73d4243b89f1320256e7c0b3945539c2a1463
MD5 dd508fdedbb6bd964e67fff17cbec98b
BLAKE2b-256 ece84c57495aa37d7e433cb4ce1549a0f5cf2f3ea56e37cb54e6341377a084c1

See more details on using hashes here.

Provenance

The following attestation bundles were made for falsify_giskard-0.3.0-py3-none-any.whl:

Publisher: publish.yml on studio-11-co/falsify-giskard

Attestations: Values shown here reflect the state when the release was signed and may no longer be current.

Supported by

AWS Cloud computing and Security Sponsor Datadog Monitoring Depot Continuous Integration Fastly CDN Google Download Analytics Sentry Error logging StatusPage Status page