Every experiment carries one of four statuses, and the status is a claim we can be held to. A concept is written, not built: prose and design only. A mockup is built and interactive but runs on curated or specimen content, which is always labeled as such. A backtest is a resolved-ground evaluation: real forecasts on real events whose outcomes were already known when the run was designed. Its numbers are measured, not specimen, but hindsight is contained by protocol (frozen inputs, blind resolution, published logs), not eliminated, and a backtest is never a forward track record. A working demo runs on live events from the production pipeline. Promotion between statuses is one-way and dated.
The experimental writer operates under the same editorial constraint as the production pipeline: every factual sentence must trace to source material, and anything the model concludes on its own must be declared as inference and gated before publication. The Glass Article is this constraint made visible; it is not a feature added on top of the writing but the record of how the writing happened. Corroboration currently counts outlets, which is a first approximation: three publications sharing one wire story are one source. Independence heuristics (shared-wire detection) are on the bench roadmap, and any "corroborated" grade shown on this site inherits that limitation until they land.
Claims made by a demo are checked on the bench: fixed gold sets, published methodology, results reported per run with dates and sample sizes. Until a bench entry exists, any number shown on this site is a specimen and is marked as one. We publish results whether or not they flatter the experiment.
Nothing on labs.pollar.news is journalism. Specimen passages are written to demonstrate an interaction, and are labeled in the figure that contains them. Live demos render real events but carry the experiment's status wherever they appear. The news is at pollar.news; this site is the workshop where its next writer is being tested.