Everbench evaluates evolving predictors against continuously arriving observations. Autonomous model research extends that domain with controlled proposals and evidence-based replacement of a serving predictor.
Champion: The predictor currently responsible for serving predictions and learning from observations. Avoid: Current model, production model
Candidate: A proposed successor to a champion, compared from a fresh state on an archive week. Avoid: Experiment model, new model
Candidate program: The immutable source and hyperparameters that construct a fresh, untrained candidate. Learned state and observations are not part of the candidate program. Avoid: Search configuration, bounded model space
Archive week: One immutable Parquet file containing every observation made available during one UTC week. It is the common offline corpus used to compare fresh champion and candidate instances. Avoid: Archive shard, research history
Serving state: Bounded learned parameters and sufficient statistics used by a champion while serving and learning. It excludes observations and raw payload identifiers. Avoid: Model data, training history
Model snapshot: A durable checkpoint of serving state used for restart recovery. It is not a source of offline research observations. Avoid: Model artifact, training archive
Weekly comparison: Measurements produced by replaying one archive week causally through fresh champion and candidate instances. The research agent may iterate on this same corpus; River's delayed progressive validation defines the replay semantics. It is not a separate holdout. Avoid: Sealed cohort, live score
Research evaluation: Repeatable causal measurements produced by replaying the same archive week through fresh champion and candidate instances while the agent edits a candidate program. Avoid: Production score, live evaluation
Objective: The owner-defined, agent-immutable criteria that determine whether a candidate may replace a champion. Avoid: Reward, agent goal
Reflection: One weekly research session that uses an archive week and owner context to evaluate a configured budget of candidate programs, then selects at most one for promotion. Avoid: Self-edit, retraining