Short answer: because “what we think people feel” and “what people actually do physiologically and behaviorally” are not the same thing. Hand-authored labels give you tidy opinions; biometrics give you falsifiable, person-specific / sub-population-specific / population-specific, time-resolved evidence. We use both: defined Scenarios as priors, and multimodal biometrics to validate, quantify, and continually refine those priors into probabilistic models that actually predict outcomes and minimize harm.

Long answer:

• Ground truth vs. guesses. A coder can label a Scenario “likely frustrating,” but only measurement tells you how often, how intensely, for whom, and for how long. Without signals (sEMG/ECG/EEG/IMU/PPG/EDA/voice/facial/inputs), you’re guessing on all of that nuance – and baking designer bias into the database.
• Individual differences. People diverge—culture, neurotype, fatigue, meds, sleep, personality. Biometrics let the same Scenario produce different Affective State Vectors (ASVs) across users and moments, rather than one static tag. That variability is the data, not noise.
• Intensity and dynamics (time matters). Emotions aren’t on/off labels. They ramp, stack, blend, and decay. Signals let us learn trajectories and Pathways (transition probabilities), not just categories. That’s essential for “what should the AI do next?” safety decisions.
• Multimodal dissociations. People can say they’re fine while EDA/EMG spike; facial affect can mislead; voice can steady while heart rate varies with respiration. Multiple channels reduce single-modality failure modes and deception (intentional or not).
• Uncertainty you can calibrate. Probable Emotional States (PESs) need confidence intervals. Those shrink only when predictions repeatedly match measured responses at scale. Hand labels can’t self-calibrate or tell you when they’re likely wrong.
• Non-obvious effects and edge cases. Some Scenarios produce counterintuitive reactions (humor under stress, relief after a “loss,” paradoxical calm). You don’t discover these from first principles—you observe them.
• Causal testing and ethics. With measured responses, you can run A/B/C Scenarios, quantify risk, and reject interventions that increase distress—before deployment. That empirical safety loop is how EmPath enforces the normative rights kernel in practice.
• Generalization to wearables-free mode. Training with biometrics lets us learn a common ASV space that behavioral signals (voice, face, text/inputs) can map into at runtime. That’s why EmPath can run without wearables in production while still benefiting from the rigor of sensor-rich training.
• Longitudinal learning. People change (health, mood baselines, seasons). Ongoing measurements keep the database current and help distinguish trait from state—something static labels can’t do.

Why not “just hand-define” the database?

You can—and we do use Scenario design as structure—but without measurements you get: (a) untested assumptions, (b) no intensity/duration distributions, (c) no calibrated uncertainty, (d) poor personalization, and (e) no reliable Pathways. In other words, you get a taxonomy, not a safety-relevant model.

Bottom line: expert-written Scenarios give us hypotheses; multimodal biometrics turn those hypotheses into tested, calibrated, bias-checked models that can safely guide AI behavior—especially in edge cases where it matters most.