cc-steer hosted calibrate

Calibrate a hosted watcher threshold against a live endpoint and promote it into the hosted registry lane.

cc-steer hosted calibrate [OPTIONS]

Scores the frozen watcher eval frame through the vLLM-compatible ENDPOINT with the exact HttpDrafter sentinel semantics, fits a substrate-specific P(NO_STEER) threshold at the fire budget, and (unless --dry-run) mints and promotes a new --component version copying the promoted local watcher adapter’s bytes verbatim with the hosted threshold. The local watcher lane and the running daemon are never touched.

Full --help output

Usage: cc-steer hosted calibrate [OPTIONS]

  Calibrate a hosted watcher threshold against a live endpoint and promote it
  into the hosted registry lane.

  Scores the frozen watcher eval frame through the vLLM-compatible ENDPOINT
  with the exact HttpDrafter sentinel semantics, fits a substrate-specific
  P(NO_STEER) threshold at the fire budget, and (unless ``--dry-run``) mints
  and promotes a new ``--component`` version copying the promoted local
  watcher adapter's bytes verbatim with the hosted threshold. The local
  ``watcher`` lane and the running daemon are never touched.

Options:
  --endpoint TEXT              vLLM-compatible base URL the hosted watcher is
                               served at; scoring posts prompt_logprobs to
                               /v1/completions (a vLLM extension pure OpenAI
                               lacks).  [required]
  --model TEXT                 Model/adapter name the endpoint serves (the
                               OpenAI `model` field).  [required]
  --timeout FLOAT              Per-request timeout in seconds for each scoring
                               call.  [default: 30.0]
  --api-key-env TEXT           Environment variable holding the endpoint
                               bearer token; unset means an unauthenticated
                               endpoint.  [default: CC_STEER_DRAFTER_API_KEY]
  --fires-per-100 FLOAT RANGE  Fire budget the hosted threshold is fit to, per
                               100 eval rows.  [default: 2.0; x>=0.0]
  --component TEXT             Target hosted registry lane; kept distinct so
                               the local watcher lane and daemon stay
                               untouched.  [default: watcher-hosted]
  --dry-run                    Print the fitted threshold and coverage without
                               minting a version.
  --help                       Show this message and exit.

Options

--endpoint: TEXT: Required. vLLM-compatible base URL the hosted watcher is served at; scoring posts prompt_logprobs to /v1/completions (a vLLM extension pure OpenAI lacks).
--model: TEXT: Required. Model/adapter name the endpoint serves (the OpenAI model field).
--timeout: FLOAT = 30.0: Per-request timeout in seconds for each scoring call.
--api-key-env: TEXT = CC_STEER_DRAFTER_API_KEY: Environment variable holding the endpoint bearer token; unset means an unauthenticated endpoint.
--fires-per-100: FLOAT RANGE = 2.0: Fire budget the hosted threshold is fit to, per 100 eval rows.
--component: TEXT = watcher-hosted: Target hosted registry lane; kept distinct so the local watcher lane and daemon stay untouched.
--dry-run: Print the fitted threshold and coverage without minting a version.