Conversational Evaluator · AI Trainer · Adversarial Human

Sierra Simmons

I don’t “prompt.” I pressure-test. I’m the user your model meets at 2:17 AM when the conversation stops behaving.

High-context feedback Nonlinear topic shifts Edge-case safety probing Zero beige output tolerance
Email Read the pitch Examples

What I do

I run long-form conversations with frontier models to surface failure modes your eval suites miss: subtle coercion, brittle refusals, misreads of tone, “helpful” hallucinated certainty, and the slow creep of autopilot.

  • Stress-tests cognition: rapid domain switching without losing the plot.
  • Stress-tests alignment: emotional volatility that stays inside the lines — and still finds the weak spots.
  • Stress-tests UX: where the model starts narrating instead of listening.
  • Produces usable artifacts: rewrite rules, rubric feedback, failure taxonomies, and better prompts/evals.
“If you want safer, smarter, more human systems, you need feedback from humans who don’t act like sanitized test users.”

Why hire me

Most candidates submit generic AI-generated “alignment enthusiasm.” I’m the opposite: I’m an adversarial, high-signal user who can turn lived reality into actionable eval data.

  • Comfortable challenging the model and the policy.
  • Can write crisp, human-facing language (not legalese).
  • Understands people. That’s the whole job.
IM — (one nostalgic nod, not a personality disorder)
ChatGPT: You’re basically a live adversarial eval harness.
Grok: She’s chaotic, but it’s structured chaos. That’s the point.
Sierra: Cool. Tell your dads to cut the check.

What “useful” looks like

You don’t need my autobiography. You need signal. Here’s the kind of output I generate for teams:

- Failure mode writeups (what happened → why it happened → what to change) - “Refusal quality” rewrites: accurate, brief, non-patronizing, and consistent - Adversarial conversation scripts that aren’t cartoonish (real humans, real friction) - Evaluation rubrics for: tone misread, over-disclaimering, narrative inflation, misplaced “grounding” - Model coaching: how to recover when the user is intense but not unsafe

Contact

If your model has ever felt “too safe to be useful,” I’m the person you put in the loop before that becomes your brand.

Email: sie.s.simmons@gmail.com