Customer releaseNorthstarComposite Operations World
A deterministic, stateful companion to the substantive customer release—100 tasks across five MCP servers, with executable verifiers, replayable traces, and a source-backed research bundle.
oracle tasks verified
The — reference-policy traces are deterministic fixtures used to validate failure analysis. They are not provider outputs or model rankings. Empirical model traces in this artifact: 0. The separately named Northstar regression archive includes a metered live runner with an optional spend ceiling for new provider evidence.
Replay a grounded task
Choose any of the 100 tasks, then step through its executable reference policy. Each tool call mutates a fresh logical episode and ends in exact SQL state checks.
This player replays the programmatic oracle—not a provider trace. Every run shown here is deterministic and uses the checked-in seed.
Diagnose every reference policy
The three rows are intentionally injected policy profiles, paired with deterministic signatures and exact verifier evidence. They are pipeline tests, not stand-ins for model scores.
— canonical oracles + — valid alternate plans passed; all — adversarial mutants were rejected.
Every call. Every check. All —.
Switch reference policies, search their records, isolate a failure signature, and replay the observable tool trajectory beside its verifier result.
A real DAG, with proof
All — functions are nodes in a —-edge directed acyclic graph. Kahn’s algorithm visits every node, and every edge points from a lower to a higher rank.
Ship the substantive customer world.
The primary archive contains — executable tools across — namespaces, — training-eligible tasks, content-bound seed assets, executable verifiers, oracle receipts, and a live single-model failure dossier.
Download customer releaseLoading size · tar.gz↓loading…python3.12 -m frontier_lab_world.seed
python3.12 servers/composite_server.py \
--db runtime/frontier-world.sqlitePython 3.12 · zero runtime dependencies · MCP stdio