opp-py 0.1.0

ODIN Provenance Passport — Python instrumentation + CLI

OPP Python SDK (opp_py)

Instrumentation + CLI for ODIN Provenance Passport (OPP).

Features

• @stamp(step_type, attrs=..., inputs=..., outputs=...) decorator emits start/end signed receipts.
• Automatic input/output content IDs (CIDs) via optional callables.
• Fallback lightweight OPE client (no external odin-sdk required) using env Ed25519 seed.
• Graph + Passport generation (opp graph, opp passport).
• Policy evidence aggregation & breach summary (opp policy).
• Optional C2PA bridge (opp.c2pa.embed_bundle_cid) to embed bundle CID into images (extra deps).

Install (editable dev)

pip install -e packages/opp_py

Required env for stamping (generate a random 32‑byte seed once):

export OPP_GATEWAY_URL=http://127.0.0.1:8080
export OPP_SENDER_PRIV_B64=<base64url_32byte_seed>
export OPP_SENDER_KID=opp-sender

Generate a seed (PowerShell):

$bytes = New-Object byte[] 32; [Security.Cryptography.RandomNumberGenerator]::Create().GetBytes($bytes); \
$seed = [Convert]::ToBase64String($bytes).TrimEnd('=') -replace '\+','-' -replace '/','_'; $seed

Decorator Usage

from opp.decorators import stamp

@stamp(
		"train.v1",
		attrs={"model":"resnet"},
		inputs=lambda args, kwargs: {"hyperparams": kwargs.get("hp")},
		outputs=lambda ret: {"metrics": ret.get("metrics")}
)
def train(data, *, hp):
		# ... training ...
		return {"metrics": {"accuracy": 0.93}}

train([], hp={"lr":1e-3})

Emitted start receipt includes inputs_cid; end receipt includes outputs_cid and status.

CLI

opp graph --trace TRACE --gateway $OPP_GATEWAY_URL         # build minimal graph
opp validate --trace TRACE --gateway $OPP_GATEWAY_URL      # continuity + basic validation
opp passport --trace TRACE --gateway $OPP_GATEWAY_URL      # rich model/data passport JSON
opp policy --trace TRACE --gateway $OPP_GATEWAY_URL        # summarized policy evidence + breaches

Passport Fields (excerpt)

{
	"trace_id": "...",
	"receipts": 12,
	"steps": ["ingest.v1","train.v1"],
	"dataset_roots": ["sha256:..."],
	"model_id": "model-123",
	"metrics": {"accuracy": 0.93},
	"safety_flags": {"pii_redacted": true},
	"policy_engines": ["opa"],
	"policy_breaches": [{"rule":"no_public_data","outcome":"deny"}]
}

Policy Evidence

If receipts carry a normalized policy object with engine + decisions[], breaches are auto‑derived when an outcome is not in {allow, pass, ok}.

Optional C2PA Bridge

pip install c2pa Pillow
python examples/c2pa_embed.py image.png

Adds a custom assertion with the bundle_cid to the image manifest.

Airflow & Dagster Examples

See examples/airflow_dag.py and examples/dagster_job.py for task/op instrumentation and trace continuity.

Testing

pytest -q packages/opp_py/tests

Backward Compatibility

New fields are additive; existing receipt payload fields preserved. Hashing uses canonical JSON (stable order, no whitespace).

License

Apache 2.0