Telemetry fields

Engine batch telemetry is a plain dictionary scoped to one collection call. The keys are uniform across policies; Applies to identifies which policies populate a mechanism rather than leaving its counters at zero. This distinction prevents, for example, interpreting zero ensemble disagreement from EpsilonGreedyQ as measured agreement.

Field	Type / unit	Applies to	Meaning
episodes	list[tuple[list[float], int, bool]]	All engine policies	Completed (returns, length, seeded) tuples. Returns contain one scalar per player; length is in ticks.
decisions	int decisions	All engine policies	Policy decisions completed during collection.
max_depth	int tree depth	SelectiveExpectimax, Minimax, Mcts, AlphaZero	Maximum search depth observed.
mean_leaves	float leaves / decision	SelectiveExpectimax, Minimax, Mcts, AlphaZero	Mean leaf count over completed searches.
mean_rounds	float rounds / decision	SelectiveExpectimax, Minimax, Mcts, AlphaZero	Mean pooled inference/search rounds.
mean_expansions	float nodes / decision	SelectiveExpectimax, Minimax, Mcts, AlphaZero	Mean expanded-node count.
mean_sigma	float value units	SelectiveExpectimax, n_heads >= 2	Mean epistemic uncertainty across searched leaves; reported as 0.0 with one head.
mean_disagreement	float value units	SelectiveExpectimax, n_heads >= 2	Mean root action-value disagreement across ensemble heads; reported as 0.0 with one head.
infer_seconds	float seconds	All engine policies	Wall time observed inside inference callbacks.
infer_calls	int calls	All engine policies	Pooled callback invocations.
infer_rows	int rows	All engine policies	Real observation rows evaluated, excluding pad padding. Mean physical call size is (infer_rows + padded_rows) / infer_calls.
padded_rows	int rows	All engine policies	Zero pad rows forwarded by pad (0 when disabled).
cache_lookups	int rows	Any policy with infer_cache enabled	Persistent evaluation-cache lookups.
cache_hits	int rows	Any policy with infer_cache enabled	Successful persistent evaluation-cache lookups.
cache_demotions	int rows	Any policy with infer_cache enabled	Gated hits invalidated by a weights_updated() boundary at release and re-inferred instead of delivered.
stall_closes	int events	Any policy	Stalled routes whose partial batch the liveness ladder closed and fired early. Nonzero means some routes fill much slower than others; expect zero in balanced workloads.
stall_releases	int events	Any policy with infer_cache enabled	Stalled hit-only routes whose cached results the ladder released without an inference call. Same balance signal as stall_closes.
terminal_sims	int simulations	SelectiveExpectimax, Mcts, AlphaZero	Simulations ending at a terminal state.
depthcap_sims	int simulations	SelectiveExpectimax, Mcts, AlphaZero	Simulations ending at the configured depth cap.
requested_rows	int rows	All engine policies	Rows entering the inference queue (decision evaluations and tail bootstraps), pre-dedup and pre-cache, counted once at the queue seam; dedup and cache savings are the gap to infer_rows. Replaces 0.2.x's fresh_rows/hit_rows/shared_rows split — row provenance is an Evaluator fact.
tail_rows	int rows	All engine policies	The tail-bootstrap share of requested_rows; requested_rows - tail_rows is the decision-evaluation term of the simulation identity.
extra_eval_rows	int rows	Mcts, AlphaZero	Rows beyond one per simulation, from multi-perspective leaf evaluation or an ExpandAll chance fan.

The preceding Engine cache counters apply only when the engine was constructed with infer_cache; they remain zero when it is disabled. See the cache lifecycle guide. The seeded episode flag is relevant only when reached-state starts are enabled.

Deep CFR

rf.solvers.DeepCfr.collect returns a separate telemetry dictionary:

Field	Type / unit	Applies to	Meaning
player	int player id	Every Deep CFR collect	Traversing player for this collection.
traversals	int traversals	Every Deep CFR collect	External-sampling traversals completed.
advantage_samples	int records	Every Deep CFR collect	Advantage records produced.
strategy_samples	int records	Every Deep CFR collect	Average-strategy records produced.
infer_calls	int calls	Every Deep CFR collect	Pooled advantage-network callback invocations.
infer_rows	int rows	Every Deep CFR collect	Information-state rows sent to callbacks.
infer_seconds	float seconds	Every Deep CFR collect	Wall time observed inside inference callbacks.
collect_seconds	float seconds	Every Deep CFR collect	Total collection wall time.
cache_lookups	int rows	Every Deep CFR collect	Advantage-cache lookups within this collection.
cache_hits	int rows	Every Deep CFR collect	Successful advantage-cache lookups within this collection.

Deep CFR's per-player caches are unconditional implementation machinery rather than a constructor option. They reuse rows within one collect call and are force-cleared at the start of the next call, allowing networks to be retrained between calls without explicit invalidation.

Treat the dictionary as extensible. Read known keys and preserve unknown keys when forwarding structured logs. See telemetry and TensorBoard for aggregation.