improve-preflightlisted
Install: claude install-skill tomcounsell/ai
# Improve Preflight
Run this before any `experiment freeze`, or whenever asked whether an improvement case can be scored. A freeze pays for known-item generation before the evaluation runner discovers whether its judge can be calibrated at all, so the first real cycle spent a receipt on an experiment whose only reachable verdict was an infrastructure failure with zero trials. These are the reads that would have said so, taken in one pass beforehand.
Every check is read-only: reserve nothing, settle nothing, journal nothing, write no row. The case id is the argument when one is given, otherwise the `case=` value of the dispatch message you are running under. Use the project environment's `valor-improve`; it is not installed outside it.
Check the control namespace with `doctor`: a clean result is the single line naming no paused heads, no stale intents, and no outstanding reservations. Anything else names a held slot that would stop the freeze's evaluation step or a paused head that would refuse the journal transition. Stop there, record what it printed as a `probe` investigation, and never force a resume from a research session. Check budget headroom with the JSON `budget` read, comparing the settled plus reserved amount in the daily paid-inference window against the charter's daily limit and the weekly headroom. State the figures you read together with the window and day boundaries the payload names; the charter asks for the boundaries to be disclosed with every accounting