How the rating is computed

Published in full, including what it cannot see and where it is currently weak. A rating you cannot reproduce is an opinion, and an opinion is not what we are selling. Every figure below is read from the sealed board at the moment this page is served.

Board scoreboard-2026-10-04T13-40-03-531Z.json · captured 2026-10-04T13:40:03Z · 7,070 operators

1 · The population

Every operator that appears in our 6-hourly census of the public MCP registry with a resolvable remote endpoint — 7,070 of them in this board. Operators are not asked to participate, do not opt in, and cannot opt out: this is an outside-in rating, the way a credit bureau rates a borrower who never applied.

Coverage is 41.8% of distinct remote-declaring hosts, and absence proves nothing. A host we could not enumerate is reported unobserved, never "clean" and never "safe". Treating our blind spot as a negative result would be the single easiest way to make this rating lie.

2 · The five dimensions and their weights

Dimension	Weight	What it measures	Distinct values	Cohort median	Signal
Reach	30%	how much of the measured tool surface this operator serves	23	0.0988	graded
Payment readiness	25%	advertises a machine-payable rail (x402)	2	0.0000	binary
Surface stability	20%	tool surface did NOT change under an unchanged version string	19	1.0000	graded
Maturity	15%	serves a version string at all, and keeps it coherent	30	0.3300	graded
Self-identification	10%	names itself in its own server info	1	0.5000	CONSTANT — cannot affect rank

Score is the weighted sum of the five, each already normalised to 0–1 by the census engine. No judgement, no analyst override, no appeal to us — the arithmetic is the whole decision.

Known weakness, published rather than hidden: Self-identification has exactly one value across all 7,070 operators while carrying 10% of the weight. That share of every score is a constant, so it moves no ranking. It is the next fix to the model. We would rather you read it here than find it yourself.

3 · Why grades are relative, not absolute

The highest-scoring operator in the entire economy sits at 0.6796 and the median at 0.3299. An absolute curve (A at 0.85) would hand a D or an F to every operator alive, which is a broken instrument, not a damning finding. So a grade is percentile standing within the measured population, stated as such everywhere, with the raw score always shown beside it.

Grade	Band	Operators	Share
A	top 5%	353	5.0%
B	next 15%	1,061	15.0%
C	next 30%	2,121	30.0%
D	next 35%	2,474	35.0%
F	bottom 15%	1,061	15.0%

4 · Ties are reported as ties

The distributions are heavily tied: 7,065 operators share the same payment-readiness value and 6,904 share the median reach. A percentile computed inside a tie block that large is precision theatre, so a dossier prints "tied with N (x%)" instead of inventing a rank. Where a percentile IS shown it is a midrank percentile — ties get half credit — so a value shared with the whole population lands near the 50th, not the 100th.

5 · Findings are observations, not accusations

A finding on a dossier means we enumerated an operator's tool surface at two different times and it changed while the served version string did not. 18 hosts carry one at or above CLASS_SUSPENDED_CRITICAL — 14 of them are on the current board, the rest are on hosts absent from the latest census window and keep their finding page without a rating. That is a statement about what a client who pinned that version is being served today — nothing about intent, and nothing about whether the change is malicious. Every finding ships the plain MCP handshake to reproduce it with no credentials of ours involved.

6 · Disputes

There is no appeals process, by design, because there is nothing to appeal to. The board is sealed, the method is this page, and the arithmetic is reproducible. If a number is wrong, recompute it and the recomputation is the correction — we do not get a vote.

Verify a sealed figure → What this costs (nothing, if you are the operator) →

7 · What this rating does NOT measure

Code quality, security posture of the implementation, or vulnerability counts — we read the surface, not the source.
Whether an operator is trustworthy, solvent, or legitimate. A high grade is not an endorsement.
Anything about closed card rails: operator-level capture on them is not externally observable and is reported UNOBSERVABLE, never zero.
Tool descriptions as served to a different client — descriptions are addressee-dependent, so a single vantage point cannot speak for all of them.
Any host outside the census. A2A coverage this window: 7,070 hosts probed, 1,176 serving an agent card.