Trust Model · pre-alpha

Status

Capability is not authority — and where to read it.

EvalGlass is pre-alpha. The canonical, always-current status of every capability — what is shipped versus ahead of the code — lives on the public Roadmap, badged with the four-state register (now / next / planned / experimental). This page keeps the one rule that doesn’t change with any release: capability is not authority.

A feature being present does not mean your run gates on it. Every honest-measurement rule on the other Trust Model pages still applies: a fresh install is informational, thresholds start proposed, and judges start uncalibrated. “Shipped” means the machinery exists, not that your evaluation is enforcing anything yet — that always takes a deliberate, host-owned approval.

Where status lives

The site and docs may run ahead of the framework code: a capability can be documented before it ships. So a single register is the source of truth, and it is public:

StateMeans
nowShipped at HEAD and shown executing only where real.
nextCommitted, building — not yet shipped (e.g. the drift watcher, a hosted sink, the Codex listing).
plannedIntended, not near-term (e.g. a metrics explorer, more worked examples).
experimentalExists but unstable (e.g. the local dashboard sink, a live-judge lane, live trace connectors reached with connect --live).
non-goalA decision, not a deferral — deliberately not built (e.g. a connect --synth generator, per-source-function attribution).

Read the per-capability detail on the Roadmap. Anything documented here that the framework has not yet built is badged there — never shown executing as if it were now.

Stability Pre-alpha means public contracts and JSON shapes are stabilising but may still change. The required path is local and hermetic: no hosted services, provider keys, or network are needed to run an evaluation. Live providers and trace backends are optional, deletable lanes, off by default.

Next steps