Recursim Glossary

Core Simulation Terms

Scenario Pack: a versioned contract that defines the scenario, seeds, scripted steps or fixtures, oracle assertions, and scoring weights.
Pack Class: readiness scope tier for a pack.
- class-1: can become gating after readiness evidence.
- class-2+: advisory until explicitly promoted.
Seed Set: deterministic input set used to make runs reproducible and comparable.
Run Scope: isolated runtime namespace for one run (state/toolstore/tmp paths).
Run Scope ID: unique identifier bound to a run scope and persisted in run evidence.
Oracle: machine-checkable expectation attached to one step.
Oracle Verdict: pass/fail result for a single oracle check.
Score Vector: weighted metric summary (correctness, contract_adherence, repair_efficiency, reuse_effectiveness, overall).

Execution Lane: the behavior channel used to execute a pack.
Deterministic Lane: fixture/replay execution where reproducibility is the first-class objective.
Live-Shadow Lane: runtime execution lane that calls real agents/tools in isolated scope and scores observed outcomes.
Fixture Mode: generate baseline artifacts/evidence for later replay.
Replay Mode: rerun against fixture contracts to test stability and score consistency.
Live Mode: execute live-shadow behavior directly and record lane evidence.
Replay Stability: reproducibility measure for comparable replay runs.

G0: scenario contract validity.
G1: replayability.
G2: score consistency.
G3: trace schema integrity.
G4: baseline diff completeness.
G5: operationalization context (local, ci, nightly).
Run Ledger: append-only JSONL evidence stream for simulation runs.
Baseline Diff: structured comparison between current run and prior comparable run.
Advisory Status: non-gating signal summary for packs not promoted to release gates.