The chart has a basement

The era gantt is a wall: five public eras, measured in years. It stands on a floor at 90° — the six-layer private funnel, on the visit clock, in days. Drag to orbit (↕ elevation, ↔ angle) and the connections show: each era drains through the floorboards onto the column that can receive it. Column height = what survives that layer (schematic); the conveyor on the floor is the private cascade with its real leak names. Three pads are dashed — no public instrument ever lands there — and their columns fill daily anyway. Hover any bar, pad, or column.

drag ↕ elevation · ↔ angle · space flips · ←/→ steps

The wall — the public era cascade · x = calendar years · five questions, each more real

20192020202120222023202420252026

TODAY

The exam era — can it recall medicine?multiple-choice knowledge · saturated at 84–90%, at/above physician level

MedQA '20

MedMCQA '22

Med-PaLM passes '22

GPT-4 aces USMLE '23

ends as marketing:
"100% on USMLE"

The rubric era — can it converse safely?physician-written rubrics over multi-turn chats

Med-PaLM long-form '23

HealthBench May '25

length-adjusted scoring '26

The task era — can it do clinical work?real EHR data · agentic workflows

MedHELM '25 · 121 tasks

MedAgentBench '25 · 70→92% in 6mo

ARISE · MAST '26

The usage erawhat do clinicians actually ask? benchmarks curated from real use

OpenEvidence 1M/day '26

HealthBench Pro Apr '26 · 15k real chats

usage ran ahead of evaluation:
~⅔ of US physicians before any
benchmark measured their questions

The outcome eradid it change the patient?

Abaluck et al. · ARISE call for post-deployment eval '26

the era the field is asking for — its instruments
(funnels, decision streams, reconciliation)
already run daily, on the floor of this chart

public, on the wall · the floorboards · private, on the floor

The floor — the private funnel · visit clock, continuous · column heights & stream counts are schematic; the leak names are real

1 · Deliveryapp DB · generated → published∅ no public rain

2 · Exposuretracker views ∪ app-DB opens∅ no public rain

3 · Engagementthe session · where public rain lands

4 · Actionaccept / reject · sandbox rain only

5 · Trustthe bug reports∅ no public rain

6 · Outcomereconciled · RCTs visit rarely

the chart

Table view — every bar, column, and stream

Surface	Item	Window	What it is	The catch
The wall public · years	Exam era → Engagement	2019–24	Multiple-choice recall — MedQA, MedMCQA, Med-PaLM, GPT-4/USMLE	saturated at 84–90%; grades the answer, exposure assumed away
	Rubric era → Engagement	2023–	Physician rubrics over multi-turn chats — HealthBench	verbosity gamed rubrics; scoring patched '26
	Task era → Action	2024–	Real EHR tasks, agentic workflows — MedHELM, MedAgentBench, ARISE·MAST	sandbox actions; real adoption invisible
	Usage era → Engagement	2025–	Benchmarks curated from real use — HealthBench Pro, OpenEvidence	frozen session telemetry; usage ran ahead of eval
	Outcome era → Outcome	2026–	Did it change the patient? — Abaluck et al., ARISE calls	the only public thread to layer 6; years per verdict
The floor private · days	1 Delivery · dry	−3d → visit	generated → published (app DB)	no public instrument, ever
	2 Exposure · dry	−7d → visit	queue → viewed (views ∪ opens)	"delivered" isn't reach: 12%→44% when denominator = viewed
	3 Engagement	the visit	the session — open time, turns, patients touched	receives nearly all public rain, as frozen slices
	4 Action	visit → +3d	accept / reject, deduplicated	sandbox rain only; know whose click it was
	5 Trust · dry	visit → +14d	thumbs → rejection comments in review	the comment is the bug report; nothing public reads it
	6 Outcome	+1d → +30d	EHR evidence reconciled to what the AI said	fills daily privately; RCTs visit once per multi-year study

Wall adapted from the era cascade (deck, slide 03); floor layers from the visit-clock monitoring map — sources are linked inside those artifacts. Streams, guides, and column heights are schematic: stream counts ∝ each era's share, columns & conveyor ∝ the funnel's attrition — and every column stands whether or not anything public is watching. Print gives wall and floor flat.