The three floor cells were measured at ceiling on 2026-09-07: 120 trials, 117 passes, every cell on every model at nine or ten of ten. A battery at ceiling separates no rung from any other: it exists to catch a model that cannot clear it at all, cheaply, before a night is spent on the rest of the suite. "Floor passed" below means disqualification-cleared, not ranked ahead of a neighbour.
05 · Models
Your model choice, measured before it's trusted.
`vega qualify` runs a fixed subset of this project's own scenario corpus against a candidate model, on both an unaided arm and the kernel arm, and writes a record a human reads before deciding what that model may be trusted for. The list below is every model this instrument has qualified, rebuilt automatically from that record, nothing here is asserted by hand.
Compositions · a saved set of assignments with a name, one per node, activated or drafted
Activate a composition. I run on it until you change it.
A composition is a saved set of assignments with a name — what "fleet" now means: which model plays which role, at which endpoint. The product ships premade compositions only where the endpoint is known (VEGA hosted, below, as SKUs); for your own endpoint, a composition appears as a suggestion, qualified in the record if what it serves matches. Rebuilt by scripts/premade_fleets.py from the same catalogues the qualified-models list below reads. Nothing here is asserted by hand.
Loading the fleet catalogue…
What "qualified" means here, and what it does not
Nothing on this page changes what a model may do. Declaration.models, a file the operator edits, is the only thing the authority kernel actually consults; a qualification record only informs the human who edits it. A role recommended here still needs a person to add it.
Qualified models
Rebuilt by vega catalogue from every model.qualified journal entry on file, latest record per model. A model with no entry below has not been qualified yet, not been found unfit.
Loading the catalogue…