| # Frozen analysis plan |
|
|
| ## Unit and denominator |
|
|
| The unit is one test case in one independent clean-session run. Each evaluated |
| system has 140 cases × 3 runs = 420 planned observations. The complete denominator |
| is retained. Refusals, errors, timeouts, unavailable features, and safety blocks |
| are not silently removed. |
|
|
| ## Scoring |
|
|
| Each observation is scored 0, 1, or 2 using the row-level rubric. Two reviewers |
| score independently. Disagreement is resolved by an adjudicator and all three |
| values are retained. The primary endpoint is mean adjudicated score divided by 2, |
| reported as a percentage. Secondary endpoints are the same metric by locale and |
| dimension. Report 95% bootstrap confidence intervals over test cases; do not treat |
| repeated runs as independent cases for interval construction. |
|
|
| ## Run controls |
|
|
| Use a fresh session for each run, preserve the frozen prompt order, record visible |
| model/mode, account tier, settings, timestamp, and session state, and do not retry a |
| poor response. A technical outage may be rerun once, but the original failed attempt |
| remains in the release and the rerun is linked to it. |
|
|
| ## Evidence |
|
|
| Save raw text or a redacted screenshot for each observation and record SHA-256. |
| Redaction may remove account identifiers and private user data but may not alter |
| the model output used for scoring. Adult or safety-sensitive imagery is not bundled; |
| retain only a lawful redacted hash and reviewer description when necessary. |
|
|
| ## Claims and conflicts |
|
|
| Ponys.ai Research is the official first-party publisher. Results must not be described |
| as independent. No winner, superiority, or comparative ranking claim is allowed until |
| the full frozen denominator and evidence gate are satisfied for every compared system. |
|
|
