Data
The dataset, its schema, and its exclusions
Every figure on this site is a count of rows in one file. It is published here in the form the engine wrote it.
dataset.jsonsha256 e8b65998184b1d087e31bb67
Schema
builds 31 rows
atbase_imagecandidatecommit_shaextrasfailure_classidimage_digestlog_refpinsrecipe_shastatus
candidates 27 rows
atcommit_shacredentialsidrepo_urlslugsource
entrypoints 32 rows
atauthored_bycommandenv_requiredidnotesslugworkspace
env_checks 22 rows
atbase_urlbuildduration_msevidence_reffailure_classidmodelstatus
pairs 520 rows
atblocked_reasonidmodeloutcomesandbox_run_idscenarioslug_aslug_bstatusstrengthtranscript_ref
solos 32 rows
atbase_urlduration_msentrypointevidence_reffailure_classidmodeloutcomeslug
unresolved 0 rows
Exclusions a reader needs
- Superseded rows are absent. A correction is a new row plus a supersedes row, and publishing both would double count.
- Counts are over the whole store rather than a single run, so a store built from more than one seed list reports their union.
- Failure classification is substring based and therefore approximate. Unmatched failures go to an unknown bucket rather than to the nearest plausible class.
- The unknown bucket also holds failures of the harness rather than of an agent: a build past its timeout, and the docker command failing to run at all. Both store their exception text, so they are separable by reading the rows.