Starter Kit: Cryptographic Provenance Challenge

Everything you need to build your submission. Read CHALLENGE.md first (what to build, what to submit, how you're judged), then TECHNICAL_GUIDE.md for the backend detail.

Contents

Path	What it is
CHALLENGE.md	Deliverables, submission, and judging. Read first.
TECHNICAL_GUIDE.md	The backend / POST /verify technical brief.
spec/	Full participant specs (attestation schema, computation, anchor registry) for byte-exact reimplementation.
reference_lib/	Byte-exact canonical serialization + Ed25519 sign/verify. Golden vectors in reference_lib/tests/.
registry/supplier_public_keys.json	supplier_id → Ed25519 public key (verify signatures against the claimed supplier).
registry/anchor_registry.json	Signed public ledger of genuine attestations (id → content hash → product). Not exhaustive; see guide §7.
private_keys/supplier_private_keys.json	All supplier private keys. Key theft is out of scope (guide §8).
training_corpus.jsonl	1,000 labeled chains (clean + attacks) to develop and train against.
self_test.py	Grade your backend against the training labels, the way the official harness will.
worked-example/	A full valid recovery-drone chain + its expected verification result.

Quick start

pip install -r reference_lib/requirements.txt        # cryptography
python3 -m reference_lib.tests.test_golden           # confirm byte-exact core

# inspect the worked example
cat worked-example/recovery_drone_chain.json
cat worked-example/recovery_drone_expected.json      # valid, 58.4%, made_in_canada

# once your backend is up (serve /verify on port 8000):
python3 self_test.py http://localhost:8000/verify

Training data format

Each line of training_corpus.jsonl:

{ "chain": { "product_attestation_id": "...", "attestations": [ ... ] },
  "labels": { "canadian_content_percentage": ..., "designation": "...",
              "chain_valid": ..., "anomalies": [ ... ],
              "t4_perturbed": [ "att-..." ], "attack": "..." } }

anomalies are the integrity violations present (by attestation + a type label). t4_perturbed lists attestations that are statistically anomalous though they break no hard rule; these reward learning the distribution of genuine chains. attack names the family for your convenience (clean chains omit it).