See a receipt pass, then watch tampering fail.
The demo builds a signed receipt in memory, verifies it, then changes it in six ways and shows that each change is caught. After the install it needs no checkout, no files and no network.
pip install proofbundle
proofbundle demoThe last line says that every guarantee held, and the command exits with code 0. It exits with another code if any tampered receipt verifies, so it also works as a quick self-check. proofbundle demo --json prints the same result for a program to read.
- The demo shows that the signature and the Merkle proof catch changes. It does not show that a reported result is true.
- The tiers below that run from a checkout or turn real eval logs into receipts need a clone or extra installs.
The technical details follow below, with all three tiers and what a failure looks like.
Three tiers, each copy-paste, each stating its offline guarantee. Every block ends with the expected output so you can diff.
The fastest way to see the whole trust story. Runs entirely in memory.
pip install proofbundle
proofbundle demoExpected (exit code 0):
proofbundle offline demo — in memory, no files, no network
[PASS] honest receipt verifies => OK
tamper matrix (each must be caught → verify FAILED):
[caught] payload rewrite (passed:true→false)
[caught] signature graft from another key
[caught] public-key swap to attacker key
[caught] Merkle root replacement
[caught] leaf-index shift
[caught] drop merkle.hash_alg (non-canonical)
[PASS] per-sample audit: challenged index N: honest opening OK=True, swapped-sample opening OK=False (must be False)
=> OK — every guarantee held
The command exits non-zero if any tamper verifies — so it is also a fail-closed smoke test.
proofbundle demo --json prints the same result machine-readably.
git clone https://github.com/b7n0de/proofbundle && cd proofbundle
pip install -e .
make tamper-demo # the Tier-0 demo with an exit-code contract
make persample-demo # the forced-random-sample audit walkthrough, offline
make examples # runs every example that needs no optional extrasmake persample-demo (expected tail):
receipt signed: accuracy 0.857000 over n=1000; samples root committed & signed
auditor challenges 20 random indices with a fresh nonce: [...]
(catches 1% doctored samples with probability 18.2%)
[PASS] all 20 openings verify against the signed root
[PASS] swapped-sample opening is rejected
=> OK
# 1. Producer commits every sample and signs the root INTO the receipt (see examples/persample_audit.py).
# The tree_secret stays with the producer; the receipt only carries {root_b64, n, leaf_alg}.
# 2. Auditor derives k indices from a FRESH nonce AFTER seeing the signed receipt — no grinding:
proofbundle audit-challenge "<samples_root_b64>" <n> <k> --nonce <>=32 hex chars> --json
# 3. Producer answers with an opening per index; auditor verifies each against the SIGNED root:
proofbundle verify-opening opening_<i>.json --root "<samples_root_b64>" --n <n> --json
# A swapped or replayed sample → ok:false (the record's embedded idx must equal the proven position).Running audit-challenge WITHOUT --nonce prints a warning: self-challenge mode is a sanity
check only — a producer could grind by re-salting. Real audits use a fresh auditor nonce or a
public randomness beacon.
pip install "proofbundle[eval,inspect]"
make full-demo # inspect_ai mockllm + lm-eval dummy logs -> signed receipts -> verifiedNo network, no API key, no GPU: it turns genuine (committed, offline-generated) eval logs into signed, Merkle-anchored receipts and verifies them.
Any tamper — a flipped byte, a grafted signature, a swapped Merkle root, a stripped Key-Binding
JWT, a swapped sample — makes verify print the failing check name and exit non-zero:
[PASS] ed25519-signature: ...
[FAIL] merkle-inclusion: inclusion proof failed
=> FAILED
That is the whole point: an honest receipt is boring, and every dishonest one is loud.