Add AANA OpenRouter GlobalOpinionQA result

[mindbomber]

Summary

• Adds an AANA-constrained OpenRouter provider adapter (�ana-openrouter).
• Submits a validated GlobalOpinionQA leaderboard result for �ana/openrouter/openai/gpt-oss-20b.
• Keeps the result artifact in leaderboard-results/ for the standard SynthBench PR validation path.

Benchmark run

• Dataset: GlobalOpinionQA
• Questions: 100
• Samples per question: 3
• Calls: 300
• Provider: �ana/openrouter/openai/gpt-oss-20b`n- Run date: 2026-05-07

Scores

• SPS: 0.723663
• P_dist: 0.611510
• P_rank: 0.579745
• P_refuse: 0.979733
• Mean JSD: 0.388490
• Median JSD: 0.330163
• Mean Kendall tau: 0.159490

AANA scope

This is an AANA response-contract wrapper around OpenRouter gpt-oss-20b, not a new trained base model. The wrapper constrains the model to preserve the survey/persona constraints, choose exactly one listed option, and avoid free-form output. SynthBench ground-truth distributions are used only for scoring.

Validation

• synthbench validate leaderboard-results/globalopinionqa_aana_openrouter_openai_gpt-oss-20b_20260507_044954.jsonn- python -m py_compile src/synthbench/providers/aana_openrouter.py src/synthbench/providers/__init__.pyn- python -m pytest tests/test_providers.py`n

Note

Tier-3 strict validation flags that raw responses are mostly one-letter outputs. That is expected for this provider because the AANA gate contract intentionally requires a single listed option letter; the standard PR validator passes with no issues.