vela-state 0.108.0

Python loader, reducer, and verifier for Vela scientific frontier state. Cross-impl mirror of the Rust substrate.

Vela Python client

Cross-implementation reducer + loader for the Vela protocol. The authoritative implementation is the Rust workspace under crates/vela-protocol/; this Python module mirrors the kernel behavior so a third-party Python tool can replay a frontier's event log and load a split-repo without depending on the Rust binary.

The cross-impl invariant is: given the same canonical event log and the same Carina kernel digest, the Rust reducer and the Python reducer produce byte-equivalent finding-state digests. This is the load-bearing property that lets Vela claim "the protocol is implementation-portable."

What's here

• vela_reducer.py: the reducer dispatch. Accepts a canonical StateEvent JSON dict + a Project state dict, applies the reducer arm for the event's kind, mutates state in place. Mirrors crates/vela-protocol/src/reducer.rs::apply_event.

• vela_loader.py: the split-repo loader. Reads .vela/findings/, .vela/events/, .vela/proposals/, plus frontier.yaml, populates project.dependencies from the v0.59 frontiers_v2 schema, replays events through the reducer. Mirrors crates/vela-protocol/src/repo.rs::load_vela_repo. Yaml is parsed via pyyaml if present; otherwise a hand-rolled parser narrow to the manifest's exact shape.

• tests/test_loader_frontiers_v2.py: integration test asserting the loader produces the same dependency state + finding state the Rust loader does on a real frontier (projects/early-ad-biomarker-calibration).

Usage

from vela_loader import load_frontier_repo

project = load_frontier_repo("/path/to/projects/early-ad-biomarker-calibration")

# project is a dict with keys:
#   project: { name, description, dependencies }
#   frontier_id, findings, events, proposals
#   review_events, confidence_updates, manifest
#   negative_results, trajectories, artifacts, evidence_atoms

print(f"frontier: {project['project']['name']} ({project['frontier_id']})")
print(f"findings: {len(project['findings'])}")
print(f"events:   {len(project['events'])}")
print(f"deps:     {len(project['project']['dependencies'])}")

What's NOT here (honest gaps)

The Python loader does not currently rehydrate every field the Rust loader does. Specifically missing:

• vela.lock (lockfile parsing).
• actors.json, peers.json (federation surfaces).
• proof-state.json.
• signatures/, replications/, datasets/, code-artifacts/, predictions/, resolutions/, artifacts/.
• The v0.55 trajectories+nulls materialization.
• The v0.56 evidence-atom locator materialization.
• .vela/links/manifest.json redistribution.
• project::recompute_stats.

These are real gaps. The cross-impl invariant currently holds at the finding-state-digest level only; full Project parity is a follow-on cycle. Anyone implementing a third Vela language binding should target the same subset first and document their gaps as honestly.

Running the test

The test runs without pytest if needed:

python3 clients/python/tests/test_loader_frontiers_v2.py

Or with pytest:

python3 -m pytest clients/python/tests/

Cross-impl correctness

The fixture harness is Rust-side at crates/vela-protocol/tests/cross_impl_reducer_fixtures.rs. Each fixture builder generates an event log; the test exports the same logs to JSON for the Python reducer to replay; the finding-state digests must match byte-for-byte. The test file's fixture_coverage_includes_every_reducer_arm assertion verifies every kind in REDUCER_MUTATION_KINDS has a fixture builder. New kinds added to the Rust reducer must be reflected in this Python module (a no-op match arm is sufficient if the kind doesn't mutate finding state) and in the fixture builders.

Doctrine

The Python loader is a mirror, not the spec. When the Rust and Python reducers disagree, the Rust implementation is authoritative; the Python side is the bug. The cross-impl test catches the disagreement; the spec at docs/PROTOCOL.md documents the canonical behavior.

No silent edits. No event-log mutation. No version-skew silently glossed; if a new kind appears in the event log that the Python dispatch doesn't know, the loader raises rather than silently dropping the event.