Ranking benchmark
150 labeled intents across six digital capability categories measured deterministic top-one and top-three ranking precision.
Evidence methodology
Discovery is not verification.
ResolveMesh keeps ranking quality, registry identity, provider claims, behavior evidence, health, observed outcomes, and manual approval as separate layers. A signal in one layer does not silently become proof in another.
Bounded validation results
86%Precision@1
89.3%Precision@3
823Tools discovered
100Providers attempted
Ranking metrics come from 150 labeled intents across six capability categories. Supply counts come from a deterministic 100-provider MCP study, not a market-size estimate.
Validation sequence
Three bounded studies
Each study answered one narrower question. None authorized tool execution or automatic production eligibility.
Remote discovery
From 2,000 latest active Registry records, 100 category-balanced providers were selected. Forty-nine completed discovery and exposed 823 tools; 151 met the narrow read-only-candidate heuristic.
Evidence pilot
Five candidates received same-origin attestation retrieval and two immediate initialize, ping, and tools/list observations. No provider tool was called.
Trust boundary
What discovery cannot prove
The official MCP tool specification says clients must treat annotations as untrusted unless they come from trusted servers and recommends human control over invocation.
Annotations as proof
Names, descriptions, and read-only hints are provider-controlled metadata. They can prioritize review, not replace it.
Availability as behavior
Forty-nine of 100 providers completed the bounded discovery, yet one of five selected candidates was unhealthy moments later.
Attestation as endorsement
An exact-definition manifest is a provider claim bound to a digest. Independent behavior evidence is still required.
Promotion gate
Layer evidence before eligibility
Payment never changes organic ranking. Failed, stale, drifted, unattested, behavior-untested, or manually unapproved capabilities remain ineligible.
Official identity
Preserve the active Registry record and exact provider endpoint.
Exact-definition claim
Bind provider attestation to the endpoint and normalized tool definition digest.
Behavior and health
Review synthetic, no-side-effect fixtures and require current, definition-stable observations across time.
Outcomes and approval
Keep manual approval until independent evidence and observed client outcomes justify a narrower automated policy.
Five-provider evidence pilot
5Providers selected
0Manifests found
8/10Successful observations
0Health verified
The immediate pilot recorded eight of ten successful observations. Health verification requires at least three stable successes spanning 24 hours, so zero providers satisfied that time-based gate during the pilot.
Demand intelligence
Publish less than you observe
Unresolved intents become public only as allowlisted, rolling aggregates. If either privacy threshold is not met, ResolveMesh publishes no cluster.
20+Unresolved events
The minimum event count for one normalized demand cluster.
5+Independent tenants
A single tenant cannot make a demand signal public.
30 daysRolling window
Counts and trend remain bounded to the current publication window.
Reproducibility
Claims stay attached to their scope
The public numbers are versioned validation results, not forecasts. New studies must state their sample, limits, retention boundary, and exact operations performed.
Integration contract
Use the current routes, schemas, limits, and privacy boundary published for builders.
Read the agent guide.Machine contract
Resolve local references and request bodies from the generated OpenAPI 3.1 description.
Read the OpenAPI document.