| # BBEH cost–quality explorer |
|
|
| Open [index.html](index.html) in a browser. It is self-contained and works offline. |
|
|
| Each point is a policy from the calibration frontier, replayed unchanged on 975 final questions (945 families). Always-Luna and always-Sol baselines are included. No final winner or new final frontier is selected. |
|
|
| The default view compares answer-model dollars per 10,000 requests with answer accuracy. The second view adds known encoder API fees and omits native/CPU policies whose hosting is unpriced. Local head hosting, training, and data acquisition are not included. The true-task-metadata method is an assisted control. |
|
|
| Select a point or table row for every supplied 95% paired family-bootstrap interval. These are pointwise, conditional on the observed tasks, and not a guarantee for future requests. The random line is an analytic expectation. A calibration-budget random comparator may have different spend on final data. |
|
|
| Regenerate from the three exact saved artifacts, using a new output directory: |
|
|
| ```bash |
| # Render saved calibration policies and their frozen final evaluation. |
| python -m evaluation.report --calibration calibration.json --frozen frozen.json --final-report final-report.json --output explorer |
| ``` |
|
|
| The renderer verifies artifact hashes, policy identities, population metadata, and the calibration-frontier membership before writing either output. No model, grading, or network call occurs. |
|
|
| - calibration: `b07b0044633ffe314669599b23ff318a1ca8a1b6d473d0de87d3135a98715a8c` |
| - frozen: `2edddada724ee572914e61593c113bee103755b4b3efd939ba218a50ee97fd7f` |
| - final_report: `f8d7fc2438d2254bbc70af8e131940d1caef42154d2858ba6859da22b55ae83f` |
| |
