99.49% of invocations are self-test. Human callers: 17 ever. Every score excludes synthetic traffic before computation; unknown callers count as synthetic.
Read the last two columns literally.
Probed is how many of a domain's tools were actually tried, not a success rate —
the probe only records a tool as probed once it answers, so the old "% reach" was 100%
everywhere by construction. exc (exercised) counts distinct tools invoked by a
non-self-test caller, which includes cron and n8n. proven is the only thing that
word now means here: registry proof_state='proven', 60 estate-wide.
The two are reported as a pair so one word never carries two measurements.
NO NEEDS EVER RECORDED is a state, not a score — those domains are ranked high
because they are high risk, but nothing about their verification has been measured,
and unknown is never a precise number any more than it is ever green. A CBS marked
≠ was computed over a reduced axis set
(verification excluded, weights renormalised, because nothing has been acquired to
verify) and is not directly comparable to the others in that column.
How this wires: dispatch inserts real
jarvis_jobs rows (class gap_fix,
status queued) — they appear on the
JOBS board; the orchestrator picks them up and asks before fixing.
| Name | Kind | Domain | Surface | Reach | Last real call |
|---|---|---|---|---|---|
| rca_playbook | tool | operations | mac-hermes | live | today |
| ncc_da_precheck | tool | compliance | mac-hermes | live | today |
| supply_leadtime_risk | tool | logistics | mac-hermes | live | honest at low-N |
| deal_rescue_sequencer | tool | sales | mac-hermes | draft-only | freeze — no send path |
| meta_experiment_engine | tool | marketing | mac-hermes | plan-only | freeze — no spend path |
| tiktok.* (40 ops) | api op | marketing | api-operation | no credential | blocked — needs key |