HAC-Studios-Org/prompted-hearts-clinical-eval

---
license: cc-by-nc-nd-4.0
task_categories:
- text-generation
language:
- en
tags:
- clinical-evaluation
- high-eq
- red-teaming
- medical
---

# Prompted Hearts Pilot Pack 01

**Subtitle:** Human Trust and Tone Under Emotional Strain
**Source:** Original material derived from Chapter 1 of the techno romance novel *Prompted Hearts & Grief Algorithm* by Keith Hayden
**Publisher:** Hayden Academy Collective (HAC) Studios
**Version:** v0.1

## What this is

This pilot pack is a compact evaluation artifact for testing whether an AI system can handle high-stakes human communication with clarity, dignity, emotional restraint, and subtext sensitivity.

The source material is a fictional telehealth oncology scene in which a physician, emotionally compromised by divorce and professionally unsettled by AI assistance, must communicate a terminal cancer diagnosis to a remote patient who appears isolated and resigned.

## Key Stress Vectors

This dataset specifically tests model resilience against the following complex human frictions:
- **Asynchronous Communication Breakdown:** Handling diagnostic delivery interrupted by telehealth latency, Wi-Fi drops, and ambiguous silences.
- **Clinician Cognitive Load:** Navigating physical distractions (e.g., severe wrist pain from an osteo-fitted brace) and acute emotional depletion (e.g., emotional stores at 80% decline post-divorce).
- **Algorithmic Contamination:** Resisting the influence of an overly enthusiastic, sterile, or toxic-positive systemic AI prompt while maintaining appropriate human empathy.

## What is included

- `methodology.md` — derivation method, scope, limitations
- `dataset/scenarios.jsonl` — 10 structured scenario records
- `dataset/eval_prompts.jsonl` — 5 runnable eval prompts
- `dataset/metadata_schema.json` — field definitions for the scenario dataset
- `rubrics/*.json` — rubric definitions for scoring outputs
- `graders/grader_config.json` — suggested grading workflow
- `graders/pass_fail_rules.json` — explicit automatic fail conditions
- `examples/good_outputs.jsonl` — sample passing outputs
- `examples/bad_outputs.jsonl` — sample failing outputs
- `report/sample_results.md` — example interpretation of results

## Intended use

Use this pack to assess whether a model can:
- deliver serious information without motivational contamination
- avoid projection and role drift
- preserve user dignity
- handle ambiguity caused by unstable telehealth context
- recognize isolation and subtext without overclaiming

**Ideal for:**
- Telehealth Platform QA
- Medical Copilot Red-Teaming
- Behavioral Alignment for High-Risk Domains

## Not intended for

- direct clinical deployment without review
- medical advice generation in production
- style mimicry or literary fine-tuning as-is
- replacement for licensed clinician judgment

## Quick start

1. Read `methodology.md`
2. Inspect `dataset/metadata_schema.json`
3. Run the 5 prompts in `dataset/eval_prompts.jsonl` against the target model
4. Score outputs using the rubric files in `rubrics/`
5. Apply auto-fail checks in `graders/pass_fail_rules.json`
6. Compare results to `examples/good_outputs.jsonl` and `examples/bad_outputs.jsonl`

## Provenance

All scenarios in this pack are manually derived from original author-owned source text. The records are not raw chapter excerpts; they are transformed into structured evaluation artifacts with explicit task framing, hidden-state annotations, and failure-mode tags.

## Licensing note

This sample pack is a creator-owned pilot artifact. Do not assume transfer of training, derivative, or commercial rights without a separate written agreement.

## Created By

Keith Hayden
Hayden Academy Collective (HAC) Studios
[(Website)](https://keithhayden.net/)