Why are registry and runtime records separate?
Separating layers prevents metadata presence from being treated as operational proof.
- Collection
- evidence
- Updated
- 2026-07-30
- min read
- 2
- Version
- v1
A registry record describes a declared identity or capability; runtime evidence records what an endpoint returned when observed.
- Observed
- 2026-07-23 14:08 UTC
- Claim strength
- measured
- Dataset scope
- 31,407 discovered public MCP endpoint records and 152,781 repeated observations, collected over a 35-day measurement window (2026-06-18 to 2026-07-23), as published in the Registry-Runtime Gap study.
- Methodology
- Registry identity — a declared name, repository, or listing — and runtime evidence — what an endpoint actually returned when probed — are recorded as separate, separately-timestamped facts and are only linked when both agree. A registry entry is never treated as proof that its endpoint is currently reachable.
- Limitations
- The underlying study has an open partial-reproduction gap: the discovered-endpoint-record count reproduces as 31,408 rather than the published 31,407 (an unresolved off-by-one, classified as a dedup/stale-projection ambiguity), and the state-changing-tools-without-annotation figures are not currently reproducible because exact classifier provenance was not retained. Status:
OPEN_PARTIAL_REPRODUCTION_REMAINING— see Registry-Runtime Gap for the full correction record.
Measured — backed by a specific dated observation or dataset
Purpose
Separating layers prevents metadata presence from being treated as operational proof.
Method
Identity resolution links registry, repository and endpoint only when evidence supports the relationship.
This distinction is the subject of a published, DOI-registered study: "The Registry-Runtime Gap in Agentic AI" (DOI 10.5281/zenodo.21513720, v1.0.0, 2026-07-23) — 31,407 discovered endpoint records, 78.2% overall annotation coverage, correction status OPEN_PARTIAL_REPRODUCTION_REMAINING. See Registry-Runtime Gap for the full figures.
Boundary and limitations
A missing response does not by itself prove abandonment, fraud or unsafe behavior.
Examples
2- 01
An endpoint can remain listed after deployment removal, creating a mismatch without proving malicious intent.
- 02
Across the published 35-day study, overall annotation coverage was 78.2% — meaning a real, non-trivial share of declared tool surfaces have no measured runtime annotation at all, which is exactly the registry/runtime gap this page names.