How entries get on the map

The inclusion tests, published in full — including the ones that exclude things we would rather have kept.

One test per stage, each written so it excludes something. A stage whose test excludes nothing is a vocabulary, not a classification, and its counts cannot be quoted. Adopted August 2026, applied from the September release.

Adopted 2026-08. Applied from the September release. The map currently on this site does not yet reflect these tests. Where the two disagree, this page is the standard and the map is behind it — stated plainly rather than quietly reconciled later.

Why this page exists

The first release tagged the average organization into almost half the workflow. When a stage admits nearly everything, its count measures the taxonomy’s vocabulary rather than the market, and a sentence like “N organizations do X” is an artifact of labelling rather than a finding.

So every stage now has a written test, and each one is written so that it excludes something. A test that excludes nothing is not a test.

The tests are published here rather than kept internal for a straightforward reason: a count you cannot audit is a claim you should not trust, and that applies to ours.

The principle underneath all of them

The map is the apparatus, not the participants.

A clinic is where deciding happens; it is not decision infrastructure. A lab is where experiments happen; it is not the experiment layer. A supplement brand is what gets acted on; it is not the action layer.

The operative question for every stage is not “does this organization operate at this point in the pipeline?” — almost everything does. It is:

Does this thing carry the claim, the reasoning, or the uncertainty across the stage boundary, in a form something downstream can use?

If the answer is “a human inside the organization does that, in their head or in a note nobody else reads,” the entity belongs in the candidate queue or a context list, not in the stage.

Two structural rules resolve most disputes:

• One primary stage per entity. Adjacent stages describe reach, not occupancy. If everything is adjacent to everything, the adjacency field is doing no work.
• Nothing is deleted for failing a test. It moves to the candidate queue or a named context list, with the test it failed recorded. Exclusion is a classification, not a verdict on the organization.

The consumer decision lifecycle

Encounter — where a person meets a health claim they did not go looking for. Include a surface that delivers unsought health claims at scale: media, search results, social feeds, advertising, retail placement. Exclude anything a person must already have a question to reach. A search engine qualifies; a symptom checker does not.

Explore — helps a person find candidate answers to a question they have formed. Include if a lay person can use it directly and unaided. Exclude anything gated behind a professional credential.

Ground — connects a claim to the source it came from. Include if a user can get from the statement on screen to the underlying study. Exclude anything that cites nothing, cites only itself, or cites a secondary summary it also wrote.

Evaluate — assesses how strong the evidence is, not merely that it exists. Include if the output contains a quality or certainty judgement distinguishable from a relevance judgement. Exclude search ranking: sorting by match is not appraisal.

Translate — converts a technical finding into lay language without adding a recommendation. Include if it restates a finding at a lower reading level while preserving effect size, uncertainty and population. Exclude anything that jumps to “so you should” — that is Decide, and conflating the two is how translation quietly becomes advice.

Contextualize — adjusts a general finding to this specific person. Include if the output changes when the person changes, driven by data about that person. Exclude demographic segmentation used for targeting rather than for adjusting the estimate.

Decide — produces a decision-ready comparison of options with the uncertainty attached. Include if it emits an artifact — a score, a table, a structured record — setting out the options, the expected effect of each, and what is uncertain, and that artifact outlives the moment of the encounter. Exclude organizations that employ people who decide: a primary-care practice is where decisions are made, not a decision instrument. Exclude anything that recommends a single product — one option is not a comparison.

Act — delivers the chosen action and records what was chosen and why. Include if it both enables the action and retains the link back to the rationale. Exclude selling a product, or every retailer qualifies. Exclude paying for a category — reimbursement sets the conditions under which action is possible and belongs on the governance rail.

Observe — measures an outcome that could change the decision. Include if the measurement is attributable to the specific person or product and is capable of contradicting the original decision. Exclude engagement and adherence metrics: how much someone used a thing is not whether it worked.

Revise — holds a pre-stated condition under which the decision should change, and acts on it. Include only if a threshold, stopping rule or review trigger was recorded before the observation, and the entity acts when it fires. Exclude adjusting a plan without reference to the original rationale — continuous optimisation is not revision, because it cannot conclude stop.

This is deliberately the strictest test on the map.

The research lifecycle

Same principle: infrastructure, not participants. A university is not a stage.

Stage	Include if…	Exclude
Funding & scoping	it allocates, sources or shapes research funding	researchers who receive funding
Discovery & grounding	it retrieves primary literature or resolves a claim to a source	general web search; tools summarising only their own corpus
Evidence synthesis	it manages protocol-driven screening, dual extraction, risk-of-bias or pooling	literature search and summarisation — that is Discovery
Protocol & registration	it timestamps a pre-analysis commitment retrievable later	project management with no immutable record
Study design	it produces a design artifact — power, allocation, endpoints, DAG	generic statistical software
Data collection & measurement	it is the instrument, or governs instrument validity	storage and transport of data collected elsewhere
Experiment execution	it runs the physical or computational experiment	scheduling and lab-ops software
Participant recruitment	it identifies or enrolls eligible participants	patient communities without an enrollment path
Analysis & reproducibility	it re-executes an analysis or verifies it re-executes	notebooks with no execution guarantee
Authoring & submission	it produces or checks the manuscript artifact	reference managers alone
Peer review & production	it manages, performs or evaluates review	venues that merely host reviewed output
Post-publication integrity	it acts on the record after publication	pre-submission screening
Guideline & decision translation	it converts graded evidence into a recommendation with strength and certainty	anything that stops at summarising evidence
Clinical evidence	it answers a clinical question at point of care with traceable sources	general medical reference without provenance
Regulated devices	it holds, or is, a marketing authorisation	software marketed as wellness outside the device gate
Regulatory & HTA	it appraises for a regulator or payer, or builds the submission	consultancies without a named appraisal role
Public & media translation	it converts findings for non-experts	tools that summarise papers for people who already read papers
Post-market learning	it surveils a product or model after deployment and can detect degradation	pre-deployment validation
Provenance rail	it issues or resolves persistent identifiers, or packages provenance	anything that merely consumes them
Governance rail	it sets rules others must follow — standards, accreditation, reimbursement conditions	organizations that comply with those rules

Two of these encode mistakes the first release made. Evidence synthesis vs Discovery: three tools were filed as synthesis when they are discovery. Post-publication integrity: four tools filed there actually run before publication.

What the tests do to the counts

Applied by hand to the current data. This is the evidence that the tests bite — each one moves something.

Stage	Before	After	What moves
Decide	7	1	six primary-care businesses become a care-delivery context list
Act	12	2	reimbursement pathways to the governance rail; retail to the candidate queue
Observe	8	6	two wearable companies to the candidate queue, pending evidence their measurement is calibrated to a decision rather than to engagement
Revise	0	0	unchanged — but now zero for a stated reason

That last row is the one worth pausing on. Zero by accident is a hole in the data. Zero because a written test excluded everything is a finding, and it is publishable: nothing on this map records a stopping rule before the observation. The entities previously listed as adjacent to Revise are continuous-optimisation systems, which by this test cannot conclude stop.

The single entity surviving the Decide test is MDCalc, whose calculators ship with the derivation study, the validation lineage and commentary from the person who built the score.

What this costs us

Being explicit about the downside, because a method page that only lists benefits is marketing.

Applying these tests makes the map look smaller and less impressive. A landscape showing 7 organizations in a decision layer reads as a busy market; one showing 1 reads as an empty one. The second is more accurate and less sellable, and we would rather publish the number that survives someone checking it.

It also means our own published counts changed between releases for reasons that have nothing to do with the market moving. The September changelog is therefore a method diff, not a market one, and will be labelled as such. October is the first honest like-for-like comparison.

---

These tests are a draft in force, not a settled standard. If one misclassifies something you know well, the useful reply is the case it gets wrong — that is more valuable than agreement. The map carries a correction link on every entry.