Metrics Reference

This is the exhaustive list of fields on the BARTMetrics object returned by score_bart(). Ranges are the engine’s documented output ranges; conceptual explanations live in The scoring engine.

Note

Color-name independence. Every metric is now config-agnostic. The EV-based metrics — risk_calibration_score, ev_ratio_score, ev_efficiency_uniformity, explosion_penalty, risk_adjustment_score, and the per-color breakdown — are computed from each color’s precomputed survival/EV curve, valid for any color names, counts, caps, and hazard families. The persona metrics — the learning-rate family (learning_rate, half_split_learning_rate, tercile_learning_rate), color_discrimination_index, color_discrimination_trajectory, patience_index, orange_avg_pumps, and the risk_style classifications — resolve behavior by risk role rather than by literal color name: the study’s colors are ranked by EV-optimal stop (safest = highest optimum, riskiest = lowest), and the two-color contrasts run between that safest and riskiest color, excluding the mid-risk ones. A renamed, re-counted, or re-ordered study therefore has all of these metrics computed coherently, and the session_warnings completeness/balance checks are judged against the study’s own colors and trial counts (issue 57). orange_avg_pumps keeps its legacy field name but reports the study’s highest-risk color. See ADR 0004 (issue 56).

Volume & outcome

Field	Range	Description
total_balloons	int	Number of balloons in the session.
total_pumps	int	Total pump events across all balloons.
total_explosions	int	Balloons that burst.
total_collections	int	Balloons successfully banked.
explosion_rate	0–1	Fraction of balloons that exploded (gross, uncensored).
average_pumps_adjusted	≥0	Mean pumps per non-exploded balloon (adjusted BART score).
avg_pumps_all_balloons	≥0	Mean pumps across all balloons, including exploded — not subject to censoring bias.
money_collected	≥0	Total earned: banked pumps × $0.25.

Calibration (sequential EV)

Field	Range	Description
ev_ratio_score	0–100	EV(participant) / EV(optimal) × 100, EV-weighted across colors. Primary calibration measure.
risk_calibration_score	0–100	Identical to ev_ratio_score; the explosion penalty is reported separately to avoid double-penalizing.
explosion_penalty	0–1	Mean excess burst rate beyond the rate expected at EV-optimal play.
rng_normalized_pumps	≥0	Mean collected stop ÷ EV-optimal stop, averaged across colors. 1.0 = optimal, <1 conservative, >1 over-pumping.
ev_efficiency_uniformity	0–1 / None	1 − CV of per-color EV efficiencies (consistency across hazard levels). None if <2 colors have usable data.
money_efficiency	0–2	money_collected ÷ the study’s expected EV-optimal earnings (Σ trials × EV-optimal per color, config-derived; ≈27.03 for the default study).
patience_index_normalized	0–1	Lowest-risk color’s EV efficiency: distinguishes patience from reckless over-pumping.
ev_optimal_stops	dict	EV-optimal stop per color, plus per-color efficiency entries (_purple_efficiency, etc.).

Learning & adaptation

Field	Range	Description
learning_rate	−1…1	\(R^2\)-weighted regression slope of pumps on trial number, sign-adjusted by risk role (mid-risk colors excluded).
half_split_learning_rate	−1…1	First-half vs. second-half mean-pump change per color. Robust at low trial counts.
tercile_learning_rate	−1…1	First-third vs. last-third change, dropping the middle third.
color_discrimination_trajectory	≈−1…1 / None	Change in safest-minus-riskiest separation from first to last third, normalized by the study’s EV-optimal spread (low_opt − high_opt; 9 for the default study).
post_explosion_sensitivity	≈−2…2 / None	Mean pump reduction on the next same-color balloon after a burst, normalized by \(s^*\). Positive = adaptive.
risk_adjustment_score	0–100	Linear distance-to-optimum alignment, referenced to the EV-optimal stops (11 / 5 / 2). Diagnostic only.
risk_sensitivity	−1…1	Pearson correlation between balloon capacity and pumps.
color_discrimination_index	0–1 / None	Deprecated (Cohen’s d, safest vs. riskiest color). Kept for backward compatibility; use ev_efficiency_uniformity.

Behavioral indices

Field	Range	Description
impulsivity_index	0–1	\(1 - \mathrm{clip}(\text{mean latency}/800\text{ ms})\). Higher = faster, more reflexive pumping.
patience_index	≥0	Mean pumps on the lowest-risk color (raw, behavioral).
within_balloon_consistency	≥0	Mean CV of inter-pump latencies within balloons.
between_balloon_consistency	≥0	CV of pump counts across balloons (high = erratic strategy).
response_consistency	≥0	CV of all inter-pump latencies (lower = more consistent).
mean_latency_between_pumps	ms	Mean inter-pump interval (intervals ≥ 2000 ms are dropped as off-task).

Composite, profile & validity

Note

Analysis-ready primitives vs. exploratory composites. adaptive_strategy_score and behavioral_profile (risk_style) are unnormed heuristics, not validated psychometric constructs: the score is an arbitrarily fixed-weighted composite and the risk style is a hand-tuned decision tree — neither has a norming sample or a reliability estimate. Use the analysis-ready primitives as dependent variables — average_pumps_adjusted, explosion_rate, ev_ratio_score, and the per-color {color}_ev_efficiency — and treat these two composites as descriptive summaries only, validating independently before reporting them. color_discrimination_index is deprecated (see below); prefer ev_efficiency_uniformity.

Field	Range	Description
adaptive_strategy_score	0–100	Exploratory (unnormed composite). Weighted blend: calibration 35%, learning 25%, uniformity 25%, money 15%.
flat_strategy_detected	bool	True if pumping is undifferentiated across colors (active learners exempted).
behavioral_profile	dict	Exploratory (unnormed heuristic). risk_style, description, dominant_traits — see Behavioral profiles.
session_valid	bool	True if the session passes all validity checks.
session_warnings	list[str]	Human-readable validation warnings (empty if fully valid).
color_metrics	list	Per-color breakdown; see below.

Per-color metrics

Each entry of color_metrics is a ColorMetrics:

Field	Range	Description
color	str	The study’s color name (default purple / teal / orange).
risk_profile	str	low / medium / high, assigned by EV-optimal risk rank (safest = low).
average_pumps	≥0	Mean pumps over all balloons of this color.
behavioral_avg_pumps	≥0	Mean pumps over collected balloons (falls back to all if <2 collected).
explosion_rate	0–1	Burst rate for this color.
total_balloons	int	Balloons of this color.
collected_count	int	Non-exploded balloons of this color.
used_fallback	bool	True if behavioral_avg_pumps fell back to all balloons.
ev_efficiency	0–1 / None	EV(behavioral mean) / EV(optimal) for this color.
ev_optimal_stop	int / None	EV-optimal pump count for this color (11 / 5 / 2).
excess_explosion_rate	float / None	Observed minus expected-at-optimal burst rate. Positive = over-pumping.