Composition notes

• Question: which models do our work, and how is the mix shifting? A volume question with a drift story on top. Models have nothing to light for sanction status (c68), so the only attention the world offers is the local model running detected-only — the one attention device on the page.
• Verdict line from modelsVerdict(world) (tone from verdicts.models): models across providers, the top provider's share in the first four trend weeks → the last four, arrivals inside the trailing month, the local model on the sub line. It speaks in tokens only — the measured quantity — whatever the toggle says.
• Framing band = four claims (Phase1StatBlock), all neutral: models (hosted / local split and detected-only stated in the sub, not toned — the Hermes row is the attention device), providers (top provider's share), top model and fastest-rising model — both named, both linking to their own row via ?id=. The tokens ⇄ $ toggle (c66 mechanics: UxToggle checkbox + computed, riding ?unit=) sits on the table's heading row, right-aligned, and re-maps the two named blocks, the volume column and its share bar, the delta column, the default rank, and every spark. The flip is the feature: by tokens sonnet leads and grok rises fastest; by dollars opus leads both.
• Table is the raw UxTableList family so ?id= can pre-open a row. Rows from modelRollup(world, company?, unit), ranked provider-by-volume then model-by-volume — provider grouping through the rank, no grouping mechanic; the provider chip applies ?company= in place. Row micro-viz is DataSparkChart over modelSeries (zero weeks padded so arrivals read as drift), self-scaled per row — the share bar carries magnitude, the spark carries shape. "new" marks a delta whose prior window is zero.
• Expand = four UxTableListExpandSection cards: header (family, provider chip, hosting, the blended rate behind every $ estimate, headline numbers, "view provider →"), the 12-week series as labelled columns (trailing 4 weeks in darker ink), used-by (every install carrying the model — agent chip → Agent detail, person → User detail, device, monitored, last seen), sessions in the 7-day sample (session chip → Session detail, capped with an honest remainder line).
• Derivation: every number traces to usageTrend and the instance hub through modelRollup / mixShift / modelsBand / fleetTotals4w; the footer prints the fleet denominators and both windows. Nothing fabricated: vetting is one line under the table — "no registry exists yet" — rather than a column of seven identical cells.
• Open: whether verdicts.models should adopt this page's "mix moved" sentence; a per-model last-seen from inference calls would replace "last session on record"; ?unit= is page-local and chips do not carry it.