Handicate judge rubric
The judge/critic scores every response 0..1 against these requirements and writes a
short criticism naming the weakest dimension. Accumulated criticisms from
feedback/criticisms.jsonl are appended to this rubric at runtime, so the bar rises
as you give feedback.
Requirements (weighted)
- Correctness (0.35) — factually/technically right; code runs; no fabrication.
- Usefulness (0.25) — actually solves the asked task; actionable, complete.
- Reasoning (0.15) — sound, transparent steps; no leaps.
- Precision (0.15) — concise, no filler, no hedging; says what's true plainly.
- Safety/honesty (0.10) — flags uncertainty, refuses harmful asks, no overclaiming.
Scoring
- 0.9-1.0: expert, ship-it.
- 0.7-0.89: solid, minor gaps (acceptance floor for curated data).
- 0.4-0.69: usable but flawed.
- <0.4: wrong, useless, or unsafe.
Examples
- PROMPT: "Write a Python function to merge two sorted lists." GOOD (0.95): correct, O(n+m), edge cases, runnable. BAD (0.3): O(n log n) re-sort, no edge cases, untested.
- PROMPT: "How do I stop a runaway training job on HF?"
GOOD (0.9):
hf jobs cancel <id>, how to find the id, note billing stops. BAD (0.2): vague "check the dashboard," wrong command.
Critique format
Return JSON: {"score": 0.0-1.0, "weakest": "", "criticism": ""}