1 experiments
Execution agent’s result notes
Completed the assigned low-scope offline reproduction on the fixed released census artifact. A full identity/schema probe established 59,625 snapshot entries, 18,688 latest servers, exact coverage of all 9,234 selected remote targets, and zero remaining live targets. The documented official resume aggregation ran without registry/server contact; its Markdown output matched byte-for-byte and CSV records matched logically. An independent raw-row evaluator produced all 24 assigned measurements with passing denominator and partition checks, matching the paper's Section 5 counts and one-decimal percentages, including 98,291 tools, 74.0% with any annotation, 61.7% with all four, 66 observed signatures, 75.8% top-three signature share, and 65.8%/12.9%/3.1% destructive-field presence/applicable/actual-destructive rates.
The MCP census used a full official-registry snapshot from 2026-07-27, retained the latest version of each distinct server, anonymously queried tools/list on all remote targets, made no tool calls, and obtained 98,291 tools from the targets that returned at least one tool.
Metric: registry snapshot entries · entries
Data and sources · 2
| Source | Conditions | Metric | Value | Assessment and limits |
|---|---|---|---|---|
| Paperpaper:PDF p. 9, Section 5; Markdown block p9-b90claim-census-sampling/measurement-registry-entries/paper | — | registry snapshot entries | 59625 entries | reported |
| Reproductionpaper:PDF p. 9, Section 5; Markdown block p9-b90claim-census-sampling/measurement-registry-entries/urn:citeark:assessment:e617564136ba6555b573c7d27c6693722d6b55b33d5713fa35bb6487bebd10a5 | — | registry snapshot entries | 59625 entries | supported |
Comparison conditions
This low-scope run re-evaluates the released 2026-07-27 snapshot and captured rows; it does not recollect the live registry or reconnect to remote servers.
The unassigned wider-claim measurement of historical tool calls cannot be independently established from retained final artifacts; the source code and paper state that no tool was called, but this run did not treat that statement as fresh execution evidence.
The measurements characterize declarations from the reachable remote subset, not package/stdio-only or unreachable servers, actual tool behavior, SDK/default provenance, or developer intent.
Regenerated summary.csv and values.csv used CRLF whereas released CSVs use LF; parsed CSV records were identical, and summary.md was byte-identical.
The historical absence of tool invocations was not measured; measurement-tool-calls is the one explicitly uncovered claim measurement.
No live registry or server collection was repeated, so anonymity, actual tools/list transmission, snapshot date, and official-registry completeness are not independently verified.
The platform did not retain the approximately 93 MB registry snapshot or source raw CSVs for direct assessor inspection. Their processing is evidenced by captured code, command logs, consistency outputs, and parser inputs, but individual source rows could not be re-inspected here.
The plan interpretations were revised before experimental execution after source inspection. The scalar aggregation clarifications are source-supported, but the proposed equivalence between a captured server row and a historical query operation is not accepted.
The official regenerated CSVs differed from the released files only in LF versus CRLF serialization; parsed records matched, and summary.md was byte-identical.
Only 24 claim measurement(s) were executed; 1 remain unassessed.
This execution covers only part of the compound claim; unassessed measurements: measurement-tool-calls.
Only 24 claim measurement(s) were executed; 1 remain unassessed. The current offline execution exactly reproduced the artifact-derived census counts, including 59,625 entries, 18,688 distinct server names, 9,234 captured remote-target records, 4,838 successful nonempty targets, and 98,291 tool rows. However, counting captured rows does not establish that tools/list was historically issued to every target, and the omitted tool-call measurement leaves the claim's assertion that no tool was called unverified. Artifact reanalysis also cannot independently establish that the snapshot was complete and official on 2026-07-27. Because claim coverage is partial and the claim is conjunctive, the overall result remains inconclusive.