Dataset Viewer
Auto-converted to Parquet Duplicate
file
stringclasses
9 values
expected_result
stringclasses
2 values
signature_status
stringclasses
3 values
structure_status
stringclasses
2 values
case_type
stringclasses
2 values
purpose
stringclasses
9 values
receipts/valid_signed_receipt.json
VALID
VALID
VALID
positive
valid signed receipt
receipts/unsigned_evaluation_receipt.json
INVALID
UNSIGNED
VALID
negative
well-formed unsigned evaluation receipt
receipts/contradiction_unsigned_receipt.json
INVALID
UNSIGNED
VALID
negative
failing-verdict receipt shape, unsigned
receipts/invalid_signature_receipt.json
INVALID
INVALID
VALID
negative
bad signature
receipts/invalid_schema_receipt.json
INVALID
UNSIGNED
INVALID
negative
schema count invariant failure
receipts/tampered_payload_receipt.json
INVALID
INVALID
VALID
negative
post-signing tamper detection
receipts/soc2_receipt.json
VALID
VALID
VALID
positive
SOC 2 control coverage receipt (domain)
receipts/rag_receipt.json
VALID
VALID
VALID
positive
RAG faithfulness receipt (domain)
receipts/legal_receipt.json
VALID
VALID
VALID
positive
legal citation audit receipt (domain)

SVR Receipt Examples

Invariant builds deterministic verification infrastructure for AI systems that mutate state. SIGMA computes structural admissibility, SATYA wraps runtime outputs and workflow transitions, and SVR records the result as a signed, independently checkable JSON receipt. This dataset is the conformance corpus for that receipt layer.

Every receipt here has been run through the published svr-verify package, and its stated outcome is the outcome the verifier actually returns, not a label that merely looks right. SVR is a vendor-neutral signed JSON receipt for AI verification results. Registered media type: application/vnd.svr.receipt+json. File extension: .svr.json. The JSON Schema is in schemas/.

The dataset viewer shows fixtures_index.jsonl: one row per fixture, with its expected verifier result, signature status, structure status, and purpose. That index is the conformance matrix; the receipts themselves are in receipts/.

Paper

This corpus is part of the public evidence stack for:

Incremental Sheaf Cohomology on Cellular Complexes: O(1)-in-n Lazy Edit Processing under Bounded Local Geometry. arXiv:2606.04227 (https://arxiv.org/abs/2606.04227).

Fixture provenance

These are synthetic conformance fixtures, not benchmark replay. Reason text inside a receipt (for example a reference to a 5M-vertex benchmark graph, or a cells_total of 25,473) is illustrative fixture content carried in the signed payload. It is not an independently reproduced benchmark run. A receipt here proves only that the artifact verifies, that is, signature plus structure. It proves nothing about the computation the receipt describes. Benchmark numbers and their provenance are tracked separately in the Invariant benchmark ledger and in arXiv:2606.04227.

What is here

fixtures_index.jsonl                 conformance matrix (viewer rows)
receipts/
  valid_signed_receipt.json          VALID
  unsigned_evaluation_receipt.json   INVALID (unsigned)
  contradiction_unsigned_receipt.json INVALID (unsigned)
  invalid_signature_receipt.json     INVALID (bad signature)
  invalid_schema_receipt.json        INVALID (structure error)
  tampered_payload_receipt.json      INVALID (signature mismatch)
  soc2_receipt.json                  VALID (domain: SOC 2)
  rag_receipt.json                   VALID (domain: RAG)
  legal_receipt.json                 VALID (domain: legal)
schemas/
  svr_receipt_schema.json            SVR v1.0 JSON Schema

Receipt outcomes

Core verification is signature AND structure together, so a receipt is VALID only when its Ed25519 signature verifies over the canonical payload and its structure passes. The outcomes below were produced by svr-verify 1.0.3.

  • valid_signed_receipt.json: signature verifies against the embedded key, structure passes. Result VALID. This is the only fully valid receipt in the set.
  • unsigned_evaluation_receipt.json: a well formed unsigned evaluation receipt. Structure passes, signature status is UNSIGNED, so the overall result is INVALID. Unsigned does not mean valid.
  • contradiction_unsigned_receipt.json: a well formed unsigned receipt whose verdict is unsafe_to_submit with H^1 dimension 2. Structure passes, signature is UNSIGNED, overall result INVALID. Included to show a failing-verdict receipt shape, distinct from a failing verifier outcome.
  • invalid_signature_receipt.json: structure passes but the signature does not verify under the stated public key. Result INVALID.
  • invalid_schema_receipt.json: the count invariant is broken (items_checked is 4 but items_passed + items_failed + items_excluded is 3). Structure fails. Result INVALID.
  • tampered_payload_receipt.json: a copy of the valid signed receipt with one checked_items reason edited after signing. The structure still passes, but the signature no longer matches the canonical payload. Result INVALID. This is the tamper-detection case.

Domain receipts

Three signed, fully conformant domain receipts demonstrate that the same receipt format and the same verifier cover different verticals. All three verify VALID under svr-verify 1.0.3.

  • soc2_receipt.json: SOC 2 control-coverage audit (receipt_type soc2), four Trust Services Criteria controls, all evidenced.
  • rag_receipt.json: RAG faithfulness check (receipt_type rag), three answer claims grounded in retrieved sources.
  • legal_receipt.json: legal citation audit (receipt_type legal), three cited authorities verified against the propositions.

How to verify

pip install svr-verify
svr-verify receipts/valid_signed_receipt.json
svr-verify receipts/tampered_payload_receipt.json

Or in Python:

from svr_verify import verify
result = verify("receipts/valid_signed_receipt.json")
print(result["valid"], result["signature_valid"], result["structure_errors"])

The interactive verifier Space accepts these receipts directly: paste or upload any file from receipts/.

Intended use

A conformance and teaching corpus. The valid receipt confirms a verifier accepts a correct artifact; the five negatives isolate the distinct failure modes (unsigned, bad signature, structure error, post-signing tamper). Suitable for CI parity tests across independent SVR verifier implementations.

Limitations and scope

  • "Valid" means the receipt artifact verifies (signature plus structure). It does not mean the underlying SIGMA computation was independently re-run. See Fixture provenance above.
  • Domain receipts for SOC 2, RAG, and legal are included and verify VALID (see Domain receipts above). Receipts for clinical, financial, and defense use the same format and the same receipt_type enum, and can be added on the same basis.
  • Protein and cofold receipts are not yet included here. Those verticals are real (the SATYA routing demo signs a conformant protein receipt), but protein and cofold are not in the v1 receipt_type enum, so a schema-clean public example waits on a v1.1 enum extension.
  • All receipts in this dataset are signed with labeled, deterministic demo issuer keys. "Valid" means the signature matches the embedded key and the structure passes; it is not an issuer-identity claim.
  • Transparency-bearing receipts are not included. The VALID_WITH_TRANSPARENCY outcome requires a Locus transparency bundle and a pinned operator key, which are out of scope for this fixture set.

Related

Links

Downloads last month
-

Space using jasonlvolk/svr-receipts-examples 1

Paper for jasonlvolk/svr-receipts-examples