detectkit-test

detectkit-test ships deterministic synthetic fixtures so engineers can prove their CSAM detection plumbing works end-to-end in CI, with zero real CSAM. It is a Python package: a versioned corpus, fingerprint generators, and a pytest plugin plus CLI that feed fixtures through your own pipeline adapter.

Install

pip install detectkit-test

What it does

• Ships a versioned corpus of procedurally generated, provably non-CSAM synthetic images and short clips, each with a manifest of expected fingerprints (MD5, SHA-1, PDQ, TMK+PDQF, vPDQ).
• Regenerates those fingerprints deterministically, so fixtures can be rebuilt and audited.
• Engineers near-duplicate pairs at exact target Hamming distances (0, 10, 31, 32, 90) to exercise match-threshold boundaries.
• Runs fixtures through a ~12-line MatcherAdapter you implement, asserting hit/miss/score and failing CI on plumbing regressions like truncation or wrong byte order.
• Emits a false-positive characterization table (recall, precision, FP-rate) as JSON plus JUnit XML, with a prebuilt GitHub Action.

Quickstart

from detectkit_test import MatcherAdapter, run_suite

class MyAdapter(MatcherAdapter):
    def match(self, fixture):
        # feed the fixture through your detection pipeline
        # and return whether it matched, plus the score
        result = my_pipeline.scan(fixture.bytes)
        return result.is_match, result.score

report = run_suite(MyAdapter())
print(report.fp_rate, report.recall)
report.to_junit("detectkit-results.xml")

You can also run the bundled suite from the command line:

detectkit run --adapter myproject.adapters:MyAdapter

Status

PRE-RELEASE: the first PyPI publish is still pending, and APIs may change before v0.1. Every fixture is synthetic and non-CSAM by construction (procedurally generated noise, gradients, and shapes), and the project never ingests real CSAM or real hash lists.

Source

packages/detectkit-test