Execution Evidence Contracts

Experimental reference implementation for fail-closed validation of execution evidence.

The project validates consistency between three independently supplied inputs:

1. an untrusted execution record,
2. expected identities and test policy fixed before execution,
3. a verifier_context representing independent observations.

It is intentionally small and uses only the Python standard library.

Important: valid=True means only that the supplied data satisfy this contract. It is not proof that tests actually ran, that the observer is authentic, that isolation succeeded, or that a process/VM really terminated. The result never grants release or qualification authority.

Quick Start

Use Python 3.12 or 3.13. There are no third-party dependencies, API keys or installation steps. In a new checkout:

git clone https://github.com/borek504/execution-evidence-contracts.git
cd execution-evidence-contracts
python -B demo_execution_contract.py

The demo prints these three results, followed by explanations and error codes:

[1] Consistent report: ACCEPTED (data only)
[2] Wrong attempt: REJECTED
[3] Missing required test: REJECTED

Every case prints qualification_authorized=False. The final line should be Demo checks: 3/3 matched the expected outcomes. Exit code 0 means the demo matched all three expected results, including both rejections; it does not mean three reports were accepted. An unexpected result exits with code 1.

See demo_execution_contract.py for complete, editable inputs and calls to validate_execution_record. The demo constructs example data in memory and prints to the terminal; it does not run the test workload described by those data or write evidence files.

What the rejection cases teach: the wrong-attempt report keeps the expected identity fixed, and the incomplete report keeps the required test inventory fixed. Only the synthetic observed-report digest is refreshed to simulate receipt of each changed report. Matching that digest is not enough to pass validation. This fixture is not an evidence collector: a real integration must establish expectations and observations independently, not copy them from a candidate's report.

Current status

This is an experimental reference profile. The current schema names retain their original jarvis-r14-* identifiers so the first public version does not silently change semantics.

The profile currently pins:

• python_version == "3.13.15" inside the validated evidence,
• CANDIDATE_QUALIFICATION -> LOCAL_LOOPBACK_ONLY_SEATBELT,
• FINAL_REGRESSION -> DENY_ALL.

The python_version value is a synthetic evidence-profile field, not a reading of your installed interpreter version.

Those are profile rules, not a claim that every user or every execution system should use those exact settings. Generalizing the profile is a separate design task.

What the validator checks

The validator rejects unknown schema versions and keys, mismatched identities, incomplete or duplicate test inventories, unauthorized skips, failed or incomplete execution states, missing isolation-check/log inventories, report substitution, and termination data that does not bind to the expected attempt/unit/supervisor.

ContractValidation.qualification_authorized is always False, and converting a result directly to bool raises TypeError. Callers must inspect .valid explicitly.

Run the tests

python -B -m unittest -v test_r14_execution_contract test_demo_execution_contract

The existing 33 contract tests remain unchanged. Seven additional demo tests check the expected outcomes, error explanations, lack of authority and the nonzero exit gate. CI also runs the exact Quick Start demo command separately. All test data are synthetic; no supervisor, VM, agent, network service or private runtime is started.

Design boundary

The validator is deliberately not an execution engine. A real deployment still needs a trusted component that establishes the origin of expectations and observations, safely collects artifacts, and proves whole-unit termination. Copying candidate-controlled values into the supposedly independent inputs defeats the intended trust boundary.

See PROFILE.md for the current reference profile and SECURITY.md for security reporting and non-goals.

License

Licensed under the Apache License 2.0. See LICENSE.