Composition notes

• Question: who is using AI, and is anyone unaccounted for? Two accountings against the identity source's denominator (Okta/Jamf), never the sensor's own sightings: people with no sensor on any device, and covered people who have never run an agent. The inverse — AI use with no person — is Inventory's row; the band names it and links.
• Verdict line is verdicts.teams(world) verbatim at fleet scope; under ?team= the page's own teamsVerdict(world, team) produces the same sentence shape over the team so it cross-foots with the rows on screen.
• Framing band = four claims (Phase1StatBlock): people covered (the one attention device, a filter), devices sensored (deployable vs Windows platform gap, a filter), using AI · 30d (the stalled joiner rides its sub — adoption folds into the claim rather than a lever; the c67 strip stays gated), installs with an owner (a link — one number, one home). All computed from the rows in the table.
• Amber ruling applied: c67 said amber = attribution debt only; the c81 chassis says amber = attention debt only. This page spends its one amber on the coverage gap and states the attribution debt (an install with no identity) in neutral tone with its link. Personal accounts are the red class, in the expand only.
• Team drill-down = filter in place: the strip writes ?team=; verdict, band and roster scope to it. People and measured tokens per team from teamRollup; spend by team is Cost Center's (c67 decision 1).
• Table is the raw UxTableList family so ?id= can pre-open a row through refs. Rows from personRoster(world, team), ranked investigate → uncovered → stalled joiner → tokens; sorters on the rest; search on the Person header. One attention device per row: the Coverage column, except the stalled joiner whose row keeps Coverage quiet and carries amber on Last active = never.
• Expand = the c67 person expand card rebuilt on the world: identity & attribution (team chip filters in place, devices and accounts as Phase1EntityChips with sensor/licensing subs, "user detail →"), attention (derived items, each with a verb-link; absent when clean), agents & usage (one row per install with monitored pill, licensing badge, 7d sessions and tokens, "view details →"; then the 30d usage line with models, repos, MCP servers). Sessions stay on User detail — the card prints the count.
• Derivation: every number traces to coverageStats / teamRollup / stalledJoiners / monitoredStats through derived/teams.ts. The Okta/Jamf integration state is a world fact (world.integrations, folded in at reconcile) — the denominator line reads it, the page asserts nothing.
• Open: whether the stalled joiner deserves its own band claim; "partial" vs "1 of 2" as the coverage word for a Windows-gap person; the page name (Linear's "Teams & Users" kept; c21 floated "Users / Devices").
• v2 (2026-09-03, Mark Peterson walkthrough): Intensity column and the expand card's heavy/moderate/light meter both dropped — Tyler's call: "let's not try to define the intensity because it's always gonna be subjective... you can just sort by tokens." The roster's existing base order (uncovered → stalled joiner → tokens descending) already is that leaderboard; no default-sort change needed. The expand card's Agents & usage section now always leads with person.summary (already-authored fixture data, previously shown only as the zero-instances fallback) — this is the "human readable summary" TODO, just promoted rather than newly built. Added a page-local lackOfSignal() (covered, has run an agent, quiet > SIGNAL_GAP_DAYS = 5d) on Last active — distinct from "not-using" (never run) and from the stalled-joiner case. The footer now states the web-based-AI coverage gap (claude.ai etc. not sensor-observed) as a named limitation. Team-level cost/tokens comparison (validated in-call, using the composed-portal-strip pattern) stays out — it's its own design pass, not folded in here. See mark-peterson-design-brief.md.