These are the detailed eval figures for the healthcare device-ladder article: which tier seats which eval, how each tier's seated shape looks on shared spokes, and cost against success on one eval at a time. The specimen is the same fictional regional system, Harborlight Health. Every number is a demo value, derived from named public evals so the grammar is honest; none of it is a measured result for any organization.
The method and the use-case table live on the companion page. This page is the plot pack.
In this series
Path: device ladder → system swarm → these figures → Agentic Operation, in plain language → Evaluations.
Harborlight is also the specimen name on /evals/demo. That page is a different artifact: a proprietary-shaped 12-agent intake-to-discharge suite. These charts are the public-eval seating pack from the ladder article. Same fictional health system name; not the same pack.
Nested seating
The seating matrix is the contract. Read a column top to bottom to see what one tier carries; read a row left to right to see where an eval first seats and how it stays seated above that.
demo · derived from real evals · not Measured Harborlight Health · fictional regional health system · specimen · illustrative · not a client
| Eval · use case | T1clinic edge | T2dept deskside | T3single accel. | T4mid-serve | T5rack |
|---|---|---|---|---|---|
| Patient access · edge (intro ≤ T1) · H1–H2 · short QA · short tool loops | |||||
| HDL-C01 Portal Intent Router | S 71% | 77% | 96% | 96% | 96% |
| HDL-C02 Grounded Patient FAQ | S 62% | 68% | 90% | 94% | 96% |
| HDL-C03 Scheduling Tool Loop | S 47% | 53% | 75% | 79% | 82% |
| HDL-C04 Freshness Policy Lookup | S 66% | 72% | 94% | 96% | 96% |
| HDL-C05 Messaging Tool Calls | S 56% | 62% | 84% | 88% | 91% |
| Clinical documentation · deskside (intro @ T2) · H3 · grounded drafts · constraint following · multi-hop · ops tickets | |||||
| HDL-C06 Ambient Note Draft | — | S 49% | 79% | 85% | 88% |
| HDL-C07 Coding Assist | — | S 52% | 82% | 88% | 91% |
| HDL-C08 Prior-Auth Workflow | — | S 43% | 73% | 79% | 82% |
| HDL-C09 Multi-Record Chart Review | — | S 45% | 75% | 81% | 84% |
| HDL-C10 Ops Desk Tickets | — | S 50% | 80% | 86% | 89% |
| Long-horizon · single accelerator (intro @ T3) · H4 · research agents · long context · runbooks · multi-app | |||||
| HDL-C11 Denials Research Agent | — | — | S 48% | 62% | 69% |
| HDL-C12 Longitudinal Record Digest | — | — | S 54% | 68% | 75% |
| HDL-C13 Integration Runbook Agent | — | — | S 51% | 65% | 72% |
| HDL-C14 Care-Coordination Multi-App | — | — | S 46% | 60% | 67% |
| HDL-C15 Guideline Research Brief | — | — | S 58% | 72% | 79% |
| Fleet → rack (intro @ T4 / T5) · H4 · concurrent care-team agents · system-wide orchestration | |||||
| HDL-C16 Care-Team Fleet Density | — | — | — | S 57% | 73% |
| HDL-C17 System-Wide Orchestration | — | — | — | — | S 72% |
S first seat · % nested seat (demo) · — not seated at this tier. Nested inclusion: an eval introduced at tier k is seated on every tier j ≥ k. Hollow = not seated below its first tier.
Show the data
| Eval | Derived from | T1 | T2 | T3 | T4 | T5 |
|---|---|---|---|---|---|---|
| HDL-C01 Portal Intent Router | ComplexFuncBench / τ² short | 71% | 77% | 96% | 96% | 96% |
| HDL-C02 Grounded Patient FAQ | SimpleQA / TruthfulQA (+ HealthBench cite) | 62% | 68% | 90% | 94% | 96% |
| HDL-C03 Scheduling Tool Loop | τ²-bench / AutomationBench | 47% | 53% | 75% | 79% | 81.5% |
| HDL-C04 Freshness Policy Lookup | FreshQA | 66% | 72% | 94% | 96% | 96% |
| HDL-C05 Messaging Tool Calls | ComplexFuncBench | 56% | 62% | 84% | 88% | 90.5% |
| HDL-C06 Ambient Note Draft | RAGTruth (+ MedScribe cite) | — | 49% | 79% | 85% | 88% |
| HDL-C07 Coding Assist | SOPBench (+ MedCode cite) | — | 52% | 82% | 88% | 91% |
| HDL-C08 Prior-Auth Workflow | τ² / SOPBench / AutomationBench | — | 43% | 73% | 79% | 82% |
| HDL-C09 Multi-Record Chart Review | FRAMES / MultiHop-RAG | — | 45% | 75% | 81% | 84% |
| HDL-C10 Ops Desk Tickets | EnterpriseOps-Gym | — | 50% | 80% | 86% | 89% |
| HDL-C11 Denials Research Agent | FinAgent / FinSearchComp | — | — | 48% | 62% | 69% |
| HDL-C12 Longitudinal Record Digest | LongBench v2 (+ HELMET hold) | — | — | 54% | 68% | 75% |
| HDL-C13 Integration Runbook Agent | Terminal-Bench 4.0 / SOPBench | — | — | 51% | 65% | 72% |
| HDL-C14 Care-Coordination Multi-App | AppWorld / τ² | — | — | 46% | 60% | 67% |
| HDL-C15 Guideline Research Brief | ReportBench / FRAMES | — | — | 58% | 72% | 79% |
| HDL-C16 Care-Team Fleet Density | τ² / AutomationBench / EnterpriseOps density | — | — | — | 57% | 73% |
| HDL-C17 System-Wide Orchestration | EnterpriseOps / Terminal-Bench soak | — | — | — | — | 72% |
Two rules make the matrix useful. First, seating is cumulative: nothing a lower tier seats is dropped higher up, so the rack's column is the union of everything below it. Second, hollow is hollow: where a tier is not seated we draw nothing, rather than a small number that looks like a result.
One tier at a time
The radar puts every eval on the same spoke set for every tier, so the shape of each tier is comparable. Pick a tier: filled spokes are the evals it seats, with the demo success as the radius; open rings at the rim are evals it does not seat. The tier below is drawn as an outline so the nesting is visible.
demo · derived from real evals · not Measured Harborlight Health · fictional regional health system · specimen · illustrative · not a client
Show the data
| Eval | Derived from | T1 | T2 | T3 | T4 | T5 |
|---|---|---|---|---|---|---|
| HDL-C01 Portal Intent Router | ComplexFuncBench / τ² short | 71% | 77% | 96% | 96% | 96% |
| HDL-C02 Grounded Patient FAQ | SimpleQA / TruthfulQA (+ HealthBench cite) | 62% | 68% | 90% | 94% | 96% |
| HDL-C03 Scheduling Tool Loop | τ²-bench / AutomationBench | 47% | 53% | 75% | 79% | 81.5% |
| HDL-C04 Freshness Policy Lookup | FreshQA | 66% | 72% | 94% | 96% | 96% |
| HDL-C05 Messaging Tool Calls | ComplexFuncBench | 56% | 62% | 84% | 88% | 90.5% |
| HDL-C06 Ambient Note Draft | RAGTruth (+ MedScribe cite) | — | 49% | 79% | 85% | 88% |
| HDL-C07 Coding Assist | SOPBench (+ MedCode cite) | — | 52% | 82% | 88% | 91% |
| HDL-C08 Prior-Auth Workflow | τ² / SOPBench / AutomationBench | — | 43% | 73% | 79% | 82% |
| HDL-C09 Multi-Record Chart Review | FRAMES / MultiHop-RAG | — | 45% | 75% | 81% | 84% |
| HDL-C10 Ops Desk Tickets | EnterpriseOps-Gym | — | 50% | 80% | 86% | 89% |
| HDL-C11 Denials Research Agent | FinAgent / FinSearchComp | — | — | 48% | 62% | 69% |
| HDL-C12 Longitudinal Record Digest | LongBench v2 (+ HELMET hold) | — | — | 54% | 68% | 75% |
| HDL-C13 Integration Runbook Agent | Terminal-Bench 4.0 / SOPBench | — | — | 51% | 65% | 72% |
| HDL-C14 Care-Coordination Multi-App | AppWorld / τ² | — | — | 46% | 60% | 67% |
| HDL-C15 Guideline Research Brief | ReportBench / FRAMES | — | — | 58% | 72% | 79% |
| HDL-C16 Care-Team Fleet Density | τ² / AutomationBench / EnterpriseOps density | — | — | — | 57% | 73% |
| HDL-C17 System-Wide Orchestration | EnterpriseOps / Terminal-Bench soak | — | — | — | — | 72% |
Cost against success, on the same eval
A scatter of cost against success only means something when every point is the same task. Here each point is a hardware tier running one eval; the horizontal axis is a demo cost per task on a log scale and the vertical axis is demo success. The joined points are the ones nothing else beats on both axes at once.
demo · derived from real evals · not Measured Harborlight Health · fictional regional health system · specimen · illustrative · not a client
Show the data
| Eval | Derived from | T1 | T2 | T3 | T4 | T5 |
|---|---|---|---|---|---|---|
| HDL-C01 Portal Intent Router | ComplexFuncBench / τ² short | 71% | 77% | 96% | 96% | 96% |
| HDL-C02 Grounded Patient FAQ | SimpleQA / TruthfulQA (+ HealthBench cite) | 62% | 68% | 90% | 94% | 96% |
| HDL-C03 Scheduling Tool Loop | τ²-bench / AutomationBench | 47% | 53% | 75% | 79% | 81.5% |
| HDL-C04 Freshness Policy Lookup | FreshQA | 66% | 72% | 94% | 96% | 96% |
| HDL-C05 Messaging Tool Calls | ComplexFuncBench | 56% | 62% | 84% | 88% | 90.5% |
| HDL-C06 Ambient Note Draft | RAGTruth (+ MedScribe cite) | — | 49% | 79% | 85% | 88% |
| HDL-C07 Coding Assist | SOPBench (+ MedCode cite) | — | 52% | 82% | 88% | 91% |
| HDL-C08 Prior-Auth Workflow | τ² / SOPBench / AutomationBench | — | 43% | 73% | 79% | 82% |
| HDL-C09 Multi-Record Chart Review | FRAMES / MultiHop-RAG | — | 45% | 75% | 81% | 84% |
| HDL-C10 Ops Desk Tickets | EnterpriseOps-Gym | — | 50% | 80% | 86% | 89% |
| HDL-C11 Denials Research Agent | FinAgent / FinSearchComp | — | — | 48% | 62% | 69% |
| HDL-C12 Longitudinal Record Digest | LongBench v2 (+ HELMET hold) | — | — | 54% | 68% | 75% |
| HDL-C13 Integration Runbook Agent | Terminal-Bench 4.0 / SOPBench | — | — | 51% | 65% | 72% |
| HDL-C14 Care-Coordination Multi-App | AppWorld / τ² | — | — | 46% | 60% | 67% |
| HDL-C15 Guideline Research Brief | ReportBench / FRAMES | — | — | 58% | 72% | 79% |
| HDL-C16 Care-Team Fleet Density | τ² / AutomationBench / EnterpriseOps density | — | — | — | 57% | 73% |
| HDL-C17 System-Wide Orchestration | EnterpriseOps / Terminal-Bench soak | — | — | — | — | 72% |
What this view is not: a mix of different tasks on one device, or latency dressed up as cost. Both are common on slides and both hide the question a buyer is asking.
The inventory behind the ladder is small and we say so: as of 2 November 2026 the machine index holds 162 candidate evals, 71 of them named in the catalog and 33 on disk, and the number of these we have measured on any organization's hardware is zero.1
What a real run would change
Nothing on this page is measured. The demo numbers exist so the charts have the right shape and the right honesty labels. For an actual organization the method is on the companion article: swap the fictional specimen for the real workload list, run the held evals on the candidate tiers inside the organization's governed environment, and replace each demo cell with a measured one, keeping the hollow cells hollow until there is a test set to fill them.
Related
- Healthcare system swarm - population, composite tasks, and token share on the same specimen.
- Agentic Operation, in plain language - seats, empty seats, and estimate vs score.
- Evaluations · Evals demo figures - engagement shape and the proprietary 12-agent Harborlight suite (not these seating charts).
Notes
- Counts are from our eval inventory as of 2026-11-02: 162 machine-indexed candidates, 71 named catalog rows, 33 on disk, 0 measured on any organization's hardware. Public leaderboard rows for healthcare are citations only; none is a test set we hold. The figures read the public subset of the ladder pack's score file; the "Show the data" table under each figure is that file. ↩︎