SHUORENHUA / MODEL EVALUATION

Chinese communication.
Compared, model by model.

Explore how models handle the audience, tone, and intent of everyday Chinese communication. Compare preference estimates, evidence coverage, and reported cost.

Model-judged screening · Human validation pending

Loading the model screen…

Model leaderboard

One shared study. Coverage and uncertainty alongside every estimate.

Loading…

EXPLORE THE EVIDENCE

Loading charts…

01 / MODEL COMPARISON

Wins, ties & losses

Accepted comparisons, ordered by the study’s point rank.

WinTieLoss

Longer bars mean more accepted comparisons. These are raw counts against different opponents; the ranking adjusts for opponents. Rejected and missing comparisons are not losses. Select a model to see its table row.

02 / RESPONSE COVERAGE

How much evidence is available?

Usable responses out of scheduled model–scenario responses. Missing or failed responses leave gaps in the comparison evidence.

03 / JUDGE AGREEMENT

Which comparisons count?

A comparison counts only when preference and both usability labels agree in both display orders. Agreement does not establish that the judge is correct.

04 / RANK UNCERTAINTY

How certain is the order?

Point rank 95% interval

Lower ranks are better. Overlapping intervals leave the order uncertain. Select a model to find it in the table. These intervals capture scenario sampling, not judge bias.

BEHIND THE NUMBERS

Read the evidence,
not just the rank.

Two orders. One accepted comparison.

Each scheduled model has two opponents per scenario. The judge sees both display orders. Preference and both usability labels must agree for the pair to count.

Exploratory, with uncertainty.

Ranks come from a Davidson model within the connected comparison graph. Missing responses are not losses. Public, AI-authored scenarios and one model judge do not establish human preference or an overall winner.

Run & cost Inspect usage and coverage

Generation costs sum reported candidate-call charges, including failed calls with known charges; they exclude judging and unknown charges. The local allowance includes imported records and reservations. Reported charges exclude credit purchase fees and taxes. No automatic expansion beyond the frozen study.

Study limits & exclusions Understand what counts