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)