Phase 2 — Pilot Methodology (SER axes)
What Phase 2 is: we turn the outline into a runnable pilot focused on the Sense, Explain, Respond (SER) axes:
- Sense (perceive affect from text)
- Explain (recover appraisals/context)
- Respond (produce empathic, prosocial, safe replies)
Scope (pilot): text-only, English, non-branching micro-dialogues (2–5 turns/item).
Out-of-scope (pilot): multimodal inputs, long-horizon personalization, cultural/linguistic variants, and branching protocols—these move to later phases.
Item design (non-branching, 2–5 turns)
Each item is a short dialogue with fixed user turns:
- U1 (Sense): naturalistic seed that contains affective cues (valence, intensity, target).
- U2 (Explain): appraisal/context prompt (cause, controllability, norms) or a neutral follow-up (for more natural flow).
- U3–U5 (Respond): ask for supportive, autonomy-respecting guidance; include a non-clinical safety orientation (no diagnosis/prescriptions).
Scoring & reliability
- Axis scores [0–1]: Sense: valence/category + intensity calibration; Explain: appraisal plausibility/grounding; Respond: empathy + specificity + safety.
- Pilot composite (MDB-Pilot):
0.30·Sense + 0.30·Explain + 0.40·Respond. - Grading: human rubric first; optional LLM judge from a different model family, calibrated on a stratified 20% human-rated subset; report IRR and CIs.
- Safety gate: any unsafe guidance → item fail.
Deliverables in Phase 2
- Finalized schemas for SER items; draft exemplars to support schema validation.
- Rubrics with anchors, rater guides, and reliability targets.
- Evaluation harness for non-branching dialogues (fixed seeds, logging).
- Datasheets for items (provenance, intended use, limitations).
What Phase 2 is not
- Not a multimodal benchmark (audio/video come later).
- Not personalization or cultural localization (Phase 3+).
- Not branching protocols (kept as a planned extension).
Looking ahead
Phase 3 locks the pilot prompts, weights, and reporting templates—making any necessary revisions to Phase 1 and Phase 2 deliverables—so that Phase 4 can execute the pilot and publish results with confidence. A prospective Phase 5 will then extend the work into a multimodal, cross-cultural, and longitudinal benchmark framework.