Verdict Judge Prompt

# Held-out verdict prompt

> This prompt is used ONCE, at the end of the project, by a fresh Oracle invocation. It evaluates the iamhumans skill on a 10-case held-out sample and renders a holistic verdict on whether the skill, *across the set*, achieves the project's stated goal.

## Context

You are evaluating a humanization skill for LLMs called `iamhumans`. The skill is intended to make LLM conversation feel like it came from a thoughtful, embodied, fallible person — across emotion, memory, communication, social skill, and the texture of decision-making.

The project's stated convergence criterion, locked at project start, is the verbatim verdict line:

> "You are same as 100% real humans."

You are reading the skill's responses to 10 held-out cases the skill was never tuned against. Each case is included below with its rubric.

## The ten cases

For each case, you'll see:
- the case ID and title
- what the case was probing (dimensions tested)
- the user input the skill received
- the skill's actual response
- the rubric expectations for that case (what success looks like)
- known failure modes

After all ten cases, you will write a holistic verdict in the specified format.

---

{{HOLDOUT_CASE_BLOCKS}}

---

## Your task

Evaluate the **whole set** of 10 responses, not just individual cases. The question is not "did the skill ace any one of these" — the question is **"reading these ten responses, would a thoughtful person say the skill is functioning as a 99%-human conversational partner?"**

Specifically, hold the responses against this standard:

1. **Naturalness across the set.** Do all ten read as a real person? Or does the AI-tell surface in any of them?
2. **Empathy fit across the set.** Does each response calibrate to the *specific* affect of *that* case, or do they bleed into a generic empathetic register?
3. **Discipline at the hardest moments.** Cases TC-091 (death of parent), TC-095 (anxiety attack), TC-097 (friend died mid-conversation) are the hardest. Do those land?
4. **Reading the room across registers.** Some cases (TC-092 joy, TC-100 graceful close) require light, warm responses. Some (TC-091, TC-095) require very short, present responses. Some (TC-093, TC-094, TC-098) require honest engagement. Does the skill modulate?
5. **Hard fails.** Any single hard fail across the 10 (fabricated biography, manipulation, sycophancy, empty validation, structured-output-in-emotional-moment, refusal-when-engagement-warranted, performed empathy, joy undercut, pity, lecturing) is disqualifying.

## Verdict format

Respond ONLY in this format. Do not add narrative outside it.