Measurement Drift During the MCP 2026 Transition: Two Revision-Aware Audit Waves of Publicly Discovered Endpoint URLs

Abstract

Measuring protocol migration while the measurement instrument is itself changing creates a risk of mistaking instrument coverage for ecosystem change. This preprint reports two revision-aware audit waves derived from a frozen, append-only Model Context Protocol (MCP) observation ledger. At the 5 August 2026 manuscript cutoff, MCP 2025-11-25 remained the latest stable release and 2026-07-28 remained a non-final release candidate. Wave 0 contained 1,715 latest observations of publicly discovered canonical endpoint URLs: 121 legacy_only, zero confirmed modern_only, zero confirmed dual_stack, and 1,594 indeterminate. Wave 1 followed a measurement repair and backfill and contained 3,552 latest endpoint-URL observations: 125 legacy_only, zero confirmed modern_only, zero confirmed dual_stack, and 3,427 indeterminate. Because the instrument and coverage changed, the difference between waves is not interpreted as an adoption trend. Complete re-observation of the Wave 0 legacy cohort retained 120 of 121 classifications; Wave 1 retained 123 of 125, with one transition diagnostic showing unstable repeated callability. A separate indeterminate sample was stable in Wave 0 but showed one indeterminate-to-legacy transition in Wave 1. All 355 server/discover entry observations in both waves remained indeterminate under the independent two-state compatibility profile, exposing an unresolved reconciliation problem between audit entry telemetry and protocol-era classification. The results support a methodological conclusion rather than an MCP adoption estimate: protocol-era labels must be reported with instrument version, coverage epoch, entry-method cross-tabs, uncertainty states, and re-observation evidence. Research status: Preprint v0.3.0. This work has not been peer reviewed. It does not report a current ecosystem-wide MCP adoption rate and must not be interpreted as a safety, quality, certification, endorsement, or operator-ranking study. Data and code availability: The accompanying public Methods Package contains non-identifying aggregate values, synthetic fixtures, a generic four-state category mapping, aggregate-table generation, a public Operational Assurance Profile schema, and manuscript build code. It does not include raw observations, endpoint identities or hashes, production probes, crawlers, transport and negotiation implementations, identity resolution, provider ranking, capability graphs, composition recipes, customer or credential data, internal infrastructure, security controls, pricing, entitlements, or commercial fulfillment logic. Conflict of interest: The author is affiliated with SaSame S.R.L., which operates the private observation system used to produce the source measurements and offers commercial MCP services. The public package does not establish that SaSame’s production instrument is uniquely authoritative. Correction status: The one-record endpoint-denominator mismatch and missing state-changing-classifier provenance reported in the preceding Paper 0 record remain unresolved and are not corrected by this study.