---

name: compare-models-blindly description: Creates seeded blinded model-comparison packets and evaluates candidate outputs twice with reversed order to control identity and presentation bias. Use when two models or prompts produce different valid answers that require qualitative judging.

Compare Models Blindly

Compare usable outputs without revealing model identity, teacher answers, latency, or prior scores to the judge.

Freeze the comparison set

Select cases by predefined metadata and stable input order before inspecting outcomes. Record the manifest. Reuse saved inference when available; do not rerun a model merely to prepare judging.

Validate candidates independently before strategic or qualitative judging. A candidate that fails identifier, replay, terminal-boundary, or other safety gates cannot win.

Build the blinded packet

Input rows should contain shared case context and two candidate fields. Run:

python scripts/blind_candidates.py disagreements.jsonl \
  --id-field id --left-field modelOne --right-field modelTwo \
  --seed experiment-v1 \
  --cases-output blinded-cases.jsonl \
  --mapping-output identity-mapping.json

Store the mapping separately. Do not consult it during judging. The script uses a SHA-256-derived assignment per case, so reruns with the same seed and IDs are stable.

Judge twice

1. Judge Candidate A and Candidate B using only supplied context and a fixed controlled rubric.
2. Create a second pass with candidate order reversed.
3. Keep explanations compact and evidence-grounded; do not request chain-of-thought.
4. Translate both verdicts to underlying candidate identity.

Award a preference only when both orders select the same underlying candidate. Otherwise classify the case as order-sensitive. Preserve ties, both-poor outcomes, insufficient-context cases, and differing valid answers for later review.

Report narrowly

Unblind only after both passes are complete. Report validation failures by model, stable wins, ties, inconclusive and order-sensitive cases, controlled score averages, and relevant metadata groups. Call automated judging what it is; do not claim human preference, product usefulness, or model acceptance.