score model responses
Evaluator cards
Production-shaped cards with score, failure labels, rewrite, scorer note, dataset labels, and eval-task note.
application route / AI voice / model behavior
xAI Model Behavior Packet
A compact route for reviewing voice judgment, humor, cultural fluency, scoring, rewrites, labels, persona consistency, and evaluation-task design.
Fast review path
10-minute reviewer path
Start here for the clearest proof of scoring, rewrites, Grok-specific model-behavior judgment, and labelable voice evaluation.
Four clicks. No archive wandering.
01
Voice Evaluation Lab
Best work-product proof: scoring, rewrites, labels, and eval-task specimen.
02
Evaluator Cards
Fastest sample of failure diagnosis, scores, rewrites, and scorer notes.
03
Coaching Grok
Diagnoses a reusable failure mode: factually correct responses can still miss the conversational job. Includes a repair template.
04
Labeling Schema
How tone judgment becomes data: humor, timing, reference fit, cringe risk, and rewrite target.
Why this role
Direct fit to the job loop
This role asks for reviewing and scoring AI responses for humor, timing, engagement, and conversational naturalness; revising sample responses while preserving factual accuracy; maintaining a consistent personality across contexts; labeling humor, irony, banter, and cultural references; and helping engineering teams turn those traits into evaluation tasks.
Voice Evaluation Lab and Evaluator Cards show the scoring, rewrite, labeling, and eval-task loop. Coaching Grok shows the model-behavior diagnosis layer. Public Receipts supply the portfolio side: timing, recovery, factual hygiene, and cultural fluency in live public contexts.
Role demand mapped to proof
rewrite sample responses
Mini evaluation cases
Invented weak responses revised toward sharper, safer, more answerable voice while preserving usefulness.
maintain personality
Conversation range
Casual, technical, disagreement, philosophical uncertainty, correction, and public humor without changing characters mid-answer.
label humor and culture
Label sample
A small schema for humor mechanic, banter intensity, reference fit, stakes/risk class, cringe risk, failure mode, and rewrite target.
build evaluation tasks
Task specimen
A scorer-ready task shape with candidate responses, reward criteria, and penalty criteria.
public portfolio
Public receipts
Live public timing, deadpan humor, self-aware recovery, factual correction, meme grammar, and quieter grounded explanation.
Best single proof
Coaching Grok: Upstream vs Downstream diagnoses a reusable failure mode: factually correct responses can still miss the conversational job. It includes a repair template: Frame → Mechanism → Facts → Inference boundary → Refusal.