Spaces:
Running
User guide
On first load the Observatory opens the overview of the hybrid classifier on the challenge set. The header lets you switch classifier and dataset split. Change left navigation to examine dataset shift, subgroup outcomes, confusion matrix, paired regression comparisons, engineering release checks, experiment lineage, and downloadable reports.
How to import: Open Export & governance → select JSON ([{...}] or {"rows": [...]}) or CSV with headers id,split,method,truth,pred. Optional columns are language,domain,tokenCount,text. Each ID must be unique for a split and method. To compare methods, upload identical example IDs across the methods being compared. To see generalization, provide both iid and challenge. Data never leaves the browser, though downloaded exports will contain the imported text.
How to restore: Open Export & governance → Restore bundled evidence. This restores the reproducible original sample and recomputed predictions, replacing imported records. Your current gate thresholds remain until you press Restore research defaults or reload the page.
Important: Synthetic-curated evaluations cannot establish actual model behavior in a high-stakes external domain. A failed release gate is a diagnostic signal, not an automated deployment block. A passing gate is not permission to deploy.