Tier 3 relevance gate

After the geo gate and Tiers 1–2, pipe the remaining candidate objects to:

TYPESAFE_API_KEY=onecli-managed python3 /home/node/.claude/skills/tessl__check-cfps/scripts/relevance-gate.py

Each object needs slug. Include available name, description, tags, city, country, is_online, website, cfp_url, organizer, is_user_group, source, status, bot_notes, and user_actioned. Attach Sessionize's verified event metadata from Step 5 in memory; do not invent descriptions or tags. The helper reads the speaker profile from /workspace/trusted/user_professional.md and the operator's criteria from /workspace/group/RELEVANCE-CRITERIA.md.

Use each parallel result's slug, verdict, and reason as Step 6 directs. For a model case, pass its evidence to Sonnet, including the fetched description. metadata_warning names missing website evidence; it is not evidence that a conference is irrelevant. A model result goes to Sonnet 5.5 or newer. Service errors, malformed answers and unknown audiences take that route. Read the helper docstring for the exit and probability contracts. The initial evaluation is recorded in skills/check-cfps/references/jev-calibration.md.

The command supplies a non-secret OneCLI placeholder through the environment; the gateway injects the vaulted key. Outside NanoClaw, supply the real key through the process environment. Do not put it in the NAS project .env. Request and routing contracts live in skills/check-cfps/scripts/relevance-gate.py (docstring, route and named constants); website retrieval bounds live in the skills/check-cfps/scripts/website_meta.py docstring and constants.

Scores and metadata stay in the run artifact. Persist only the existing status and bot_notes fields. The CFP record schema remains version 1; the morning brief continues consuming that schema. The helper preserves user actions and geo-dismissed rows without network requests. It never changes Tier 1 approval or Tier 2 blocking.

Before changing the pinned model, questions or routing thresholds, rerun the helper on candidates from the operator's labelled cohort. Retain the result artifact with the evaluated model and metadata projection. Summarize it through the co-shipped evaluator:

python3 /home/node/.claude/skills/tessl__check-cfps/scripts/summarize-calibration.py

Supply a stdin JSON object with labels (the labelling artifact's items array) and scores (the gate's result array), joined by slug. The JSON summary reports the evaluated and excluded cohort, automatic errors, model fallbacks, preserved rows and missing metadata. Exit 0 emits a JSON summary; exit 2 means invalid or incomplete inputs, with a diagnostic on stderr and no summary. Its complete input/output contract is in skills/check-cfps/scripts/summarize-calibration.py. Record the summary with the run artifact. A small calibration set does not establish accuracy on future conferences.