# Noetryx public-data feasibility benchmark v0.1

Protocol fixed before viewing benchmark scores. Date: 2026-10-11.

## Question and scope
Can session-level physiological summaries estimate the post-session, self-reported mental-load label in the public Smerdov et al. League of Legends dataset better than a training-set median?

This is a retrospective label-estimation experiment. It is not a forecast, a brain-state measurement, a validated stress detector, or evidence that feedback improves esports performance. The data are third-party research data, not Noetryx's proprietary cohort. No scientific, commercial, or investment valuation is inferred from the results.

## Data and provenance
Source: https://github.com/smerdov/eSports_Sensors_Dataset
Citation: Smerdov, A., Zhou, B., Lukowicz, P. and Somov, A. (2020). Collection and Validation of Psychophysiological Data from Professional and Amateur Players: a Multimodal eSports Dataset. https://arxiv.org/abs/2011.00958
License: Creative Commons Attribution 4.0 International, as declared by the source repository. Preserve source attribution and the license. No exclusive ownership of source data is claimed. Sensor-specific vendor rights and suitability for future commercial products require separate review.
Pin the source commit. Download match metadata, per-player reports, and available heart-rate, GSR, EMG and facial-temperature CSVs. Record SHA-256 hashes. Do not download camera recordings or introduce identities beyond anonymous participant keys.

## Representation
One row per player-match. Participant key is team plus player ID, because player IDs repeat between teams. Label is only the self-reported `mental_load` value. Use finite numeric sensor values, excluding time columns. Calculate each sensor's median, interquartile range and standard deviation. No labels, match outcome, role, player rank, team name, or player ID are input features. Missing modalities are imputed using training-fold medians; preprocessing never sees the held-out rows.

## Fixed models and validation
Primary: leave-one-participant-out evaluation. Secondary: leave-one-match-out evaluation. Both protocols are reported; neither is a substitute for an independent external cohort. In the primary test, held-out players may share match conditions with training players; in the secondary test, players can recur. This limitation is disclosed.
Model: median imputation, standard scaling and ridge regression with alpha=10, fixed without tuning. Baseline: training-fold median label. Primary metric: participant-macro mean absolute error. Also report pooled MAE, per-group errors, sample counts, missingness and target distribution. No random window split. No oversampling or synthetic data.
Compute an exploratory paired group-bootstrap 95% interval for model-minus-baseline MAE using 2,000 resamples and seed 20261011. This interval reflects the available groups and does not establish population-wide generalization.
Internal advancement gate: at least 10% lower macro MAE and an entirely negative paired bootstrap interval. This is an internal screening rule, not an accepted clinical standard. A failed gate is retained and published without renaming or tuning this benchmark.

## Next validation stages (not yet completed)
1. Domain-expert review of signals, label meaning and confounds.
2. Permissioned Noetryx cohort with explicit collection/reuse terms, synchronized gameplay events and prespecified targets.
3. Frozen evaluation on different players, teams, dates and devices, with calibration and abstention criteria.
4. Prospectively registered controlled trial of coaching feedback versus a relevant baseline, powered after a pilot estimates variance and effect size.
5. Independent replication and paid product pilots. Public research accuracy alone is insufficient evidence of commercial demand.

No EEG sensing, neurofeedback or neuromodulation is implemented by this benchmark.