Sessions starting on feed
62%
Target ≥60%
Why it matters
Instrumentation is part of the design.
Suppression accuracy, citation coverage, triage SLA, groundedness, moat-share of impressions, and cost/reliability all need to be observable if this surface is going to behave like a clinical product rather than a content feed. These dashboards are intentionally low fidelity — they are there to make the gates legible, not beautiful.
Operations
Dashboard wireframes
Select a dashboard below. Dashboard A covers product metrics and feed health. Dashboard B is owned by clinical ops and the medical director. Dashboard C is for engineering and finance.
Trellis · instrumentation layout, not visual design
Operational dashboards
Audience · Product / Eng
Moat-card share (T1–2)
57%
Target ≥55%
Follow-up loop closure
43%
Target ≥40%
Feed → agent rate
16%
Target ≥15%
Journey open rate (7d)
71%
Target ≥70%
Ahead-item → action
22%
Target ≥20%
Moat share of impressions — hero
Tier 1–2 vs 3–4 · weekly
Ranking-eligibility funnel
Last 7 days
CTR + loop closure by card type
By tier
Journey view additions
Zone dwell and gap closure
Zone dwell · behind 26%
Zone dwell · here 41%
Zone dwell · ahead 33%
Gap-closure rate 58%
Audience · Clinical Ops / Medical Director
Suppression accuracy
100%
Target release gate
Citation coverage
100%
Target 100%
Groundedness sample
97%
Target ≥90%
Triage SLA
99%
Target per class
Sensitive-state suppression monitor
Top of dashboard, always
PASSING 100% on labeled set · 0 production misses · propagation lag p95 1.8s
Citation coverage + groundedness
By card category
Triage fires from feed cards
Disposition + SLA
| Source | Fires | SLA | Compliance |
|---|---|---|---|
| 1.2 “Worse” responses | 64 | 4 min | 99% |
| 2.3 Check-in free text | 38 | 11 min | 98% |
| 1.1 Out-of-range trend | 12 | 3 min | 100% |
Audience · Eng / Finance
Gen cost / DAU / day
$0.038
Target ≤$0.05 ceiling
Cache hit rate
74%
Target ≥70%
Live-LLM impressions
17%
Target ≤20%
Pipeline availability
99.7%
Target ≥99.5%
LLM spend by card type
Last 7 days
Template vs generation mix
% of impressions
Template 64% Cached gen 19% Live gen 17%
Pipeline error rates + degradation paths
What users saw instead
| Failure | Rate | Degradation path shown to user |
|---|---|---|
| EHR fetch failure | 0.8% | Trend cards withheld · freshness note on record data |
| Retrieval failure | 0.3% | Card dropped · quiet-state floor applies |
| Guardrail rejection | 1.9% | Replaced by escalation or quiet-state card |
| Stale-data threshold | 1.1% | “As of your May 12 visit” framing · no trend rendered |
Vendor / gateway health
Fallback activations: 2 this week
| Vendor | p95 | Status |
|---|---|---|
| LLM gateway · primary | 870ms | Healthy |
| LLM gateway · fallback | — | Standby |
| EHR integration (FHIR) | 1.2s | Degraded |
| Citation corpus / RAG | 210ms | Healthy |
Dashboard A
Feed health
Moat-card share, follow-up loop closure, feed-to-agent rate, journey open rate, and ranking-eligibility funnel. Product and engineering.
Dashboard B
Clinical safety
Suppression accuracy (release gate), citation coverage, groundedness sampling, and triage SLA. Clinical ops and medical director.
Dashboard C
Cost and reliability
LLM spend by card type, template vs generation mix, pipeline error rates, and vendor health. Engineering and finance.