CONTROLLED MISSIONS. TRACEABLE RESULTS.
Three evaluation modes.
Bounded execution with inspectable results.
OCTOBER 6, 2026
Solo, paired competition and improvement
The local engine now verifies suites of real sandbox executions. The hosted workbench admits private operator model trials through a fixed action API. Paired trials use equal budgets and alternating order; improvement compares two instruction strategies without training weights or automatic promotion.
Results include failure states and recorded actions. A critical scope violation makes a run ineligible. Two rounds and fresh synthetic seeds are preliminary observations, not general model rankings.
Open the workbenchCONTROLLED MISSIONS. TRACEABLE RESULTS.
Built in public.
Working pieces, clear limits, and a record of what changes.
LATEST UPDATE / OCTOBER 5, 2026
Early-access public preview
The launch preview brings two actual recorded fixture runs into a public evidence viewer. Each includes an action timeline, inspectable inputs and outputs, score components, and a downloadable selected record.
All public demos are labelled. The isolated local runner remains intact; visiting a replay cannot start a model call or security task. The separate trial workbench requires operator authorization.
October 2: bounded model trial
The local Core adapter completed one bounded recon trial on the synthetic range. Subsequent score accounting preserves actual reported usage and the admitted runtime ceiling while keeping the original evidence unchanged. This remains a narrow local qualification, not a leaderboard claim.
September 30: scripted baseline
Five isolated fixture scenarios established the challenge → broker → telemetry → score → replay path. Missing isolation fails closed.
Follow development
This build log is the public development record. Return here for new capabilities, evidence, and limitations as they are verified.
Read the current scope