Source: qdash /teams — c67 cognitive pass, 2026-08-15
The form is settled; this round questions the information. The current lenses are mostly states — how much, how many. Optimization, improvement, and risk reduction need levers: numbers that move when someone does something on purpose. The test a lens must pass: (a) an unambiguous good direction · (b) attribution to something the team or its admin actually does · (c) a feedback loop shorter than a quarter. Every provenance claim below was verified against qdash source this session.
The current five, judged
A state — and that's fine. It's the accountability ledger, the Cost Center reconciliation point, and the receipts system's spine. One state lens is correct; five was the problem.
A normalization of Spend, not a cognitive job — and it fails the direction test: high could be power users or waste. The stat sheet and whisper line already carry it. Its actionable half becomes Waste.
Weather. Nobody improves active-now; it's refresh-to-refresh noise. But it's the only near-real-time signal — exactly what the Auto lens was built to surface when notable, and what the focus card shows in context.
The micro-histogram bake-off finding was real — but the metric has no good direction (depth or sprawl?). A diagnostic, not a directive. The histogram keeps its win inside the focus card.
The surprise of the audit: danger_level is already conditions-based (yolo/sandbox/approval configs), and findings subsume those conditions with a workflow state machine and timestamps. What Risk lacked was a lever — the age of the oldest open item adds one: an SLA someone can hold.
Every candidate, with receipts
Provenance: ✓ buildable today client-side · ◐ derivable from existing records/routes · ○ needs a backend aggregate · ✗ the data doesn't exist. The so-what column is the lever test: if this number moves, who did what?
| Metric | Job | If it moves, who did what? | Prov. | Verdict |
|---|---|---|---|---|
| Tokens (native unit) | accountability | the same ledger in ground-truth units — token counts are measured, dollars are estimated | ✓ | shipssame by_team rows as Spend; the $/Tokens depth toggle stays on Cost Center (c66) |
| Waste (idle >14d spend) | optimization | a manager re-engaged the idle person or reclaimed the seat/key — either outcome is the job done | ✓ | shipscosts by_user /30d × rollup last-active; threshold anchored to data-max |
| Retention (7d cohort return) | improvement | enablement worked — the people who tried it last month are still here | ◐ | shipsfirst_seen/last_active from /api/agents (NOT /api/sessions — 200-row cap) |
| Coverage (instrumented machines) | leading risk | someone instrumented a machine — the shadow surface shrank by name, not by vibe | ✓ | shipsinstallations.instrumented via machineRollups; denominator = "machines we can see" |
| Open, aged | lagging risk + SLA | someone resolved the old finding instead of accumulating new ones — the SLA held | ✓ | shipsfinding first_seen + workflow statuses; terminal = resolved/accepted |
| Mapping coverage | the page's own lever | an admin sat in the mapping workspace and assigned identities | ✓ | verdict linefleet-level, not per-team — it headlines, it doesn't lens |
| Cache leverage | optimization | prompt/session habits improved — but per-team it mostly reflects harness mix, so movement ≠ the work working | ◐ | → Cost Centerreachable via /api/costs?team= (the unlock below); cost-composition belongs on /cost |
| Premium-model share | optimization | model-pick policy changed — but workload mix confounds it | ◐ | → Cost Centersame ?team= route, by_model rows; per-person split genuinely impossible client-side |
| Exposure (composed conditions) | leading risk | — | ✗ | killed2 of 3 components don't exist (no MDM signal in web.rs; no sensitivity classification on top_context); the real third — consumer keys — is /identity's per decision 3. Re-scoring here recreates the dual-formula bug this cycle exists to kill. |
| Egress surface | risk reduction | allowlist work shrank the destination set — an admin lever, not a team behavior | ○ | → controls / focus cardegress_dests is a per-agent count; summing double-counts shared hosts — honest per-team distinct needs a backend aggregate |
| Guardrail posture | risk reduction | policy authoring covered more actions — real lever, wrong altitude for a chip | ◐ | → focus card / controlstool_calls decision_source (config|user|policy) — % policy-covered derivable |
The lever strip
One lens per cognitive job: accountability · optimization · improvement · leading risk · lagging risk with an SLA. Each wears its bites-system micro-form; every mark is still a portal. The verdict line carries the page's own lever — mapping coverage — because the strip shouldn't get credit for the workspace's work.
Teams
Receipts, and one correction to our own record
/api/costs applies team scoping BEFORE its accumulation loop, so ?team=X&period=30d returns team-scoped totals (including cache %), by_model, and a per-team daily series. Our earlier claim "per-team trends impossible — the daily series is fleet-wide" is true only for a single unscoped fetch. N scoped fetches (N = team count, small pre-SCIM) buys history, cache split, and model split with zero backend work. CYCLE.md corrected; Cost Center's team table inherits the opportunity.
Waste's 14 days, Retention's 7-day cohort, and Open's age all take a threshold — exactly the kind of judgment the receipts law says to surface. All three anchor to the dataset's MAX(last_seen) (the backend's window_cutoff convention), never wall clock, and the whisper line states the active threshold under every lens.
Waste joins costs by_user (label = email-or-OS-user, lowercased) against the rollup's user key — the same unassigned seam the strip already prints. Coverage's denominator is "machines we can see": an install or an observed agent. No MDM join exists, so the honest gap may be larger than shown — the caption says so instead of smoothing it.
States vs levers
One accountability job is enough. Spend stays because a page needs its ledger, and Tokens rides beside it as the same ledger in the native unit — token counts are measured where dollars are estimated (the c66 principle). Two units, one job. But four more states were four ways of restating "big team." Every other slot now answers "what should someone do, and how will we know it worked": re-engage or reclaim (Waste), make it stick (Retention), instrument the machine (Coverage), resolve the old one (Open, aged). And the page's own lever — assignment — headlines the verdict line rather than hiding in a lens.
The demotions aren't deletions. $/person lives in the stat sheet, Active-now belongs to the Auto lens (surfaced exactly when notable), the agents-per-person histogram keeps its bake-off win inside the focus card. Metrics moved to where their cognitive job actually is — which is the whole lesson of the pass: the question was never "what can we show" but "what will someone change".