Unit. A listed (run_id, server_name) and, on a successful unauthenticated remote
tools/list, a (run_id, server_name, tool_name) definition. SQLite has runs with enrollment
and completion, server_observations with listing, reachability, endpoint and reason, and tool
observations with timestamp, description, input schema and content hash. JSON snapshots can
contain additional raw response fields and are preserved as plain JSON or verified gzip.
Selection. The original sample took one in twelve current registry names using
md5(name) % 12 == 0. A 900-attempt budget began skipping sampled servers on August 7 at
12:20 UTC. Of the 198 completed historical runs through September 5, 107 hit the cap. A recorded
cap skip represents no endpoint attempt. An unsuccessful attempted read and an active listing
without a remote endpoint have separate coverage states.
Fixed panel. From 20261115T050427Z, the runner checks the same 897 names each cycle. These
were remotely listed in at least one completed sample from July 9 through August 8. A name missing
from the active registry is recorded as delisted. The collector attempts each listed remote
member without an attempt cap, accounts for each package-only and delisted entry, and records no
newcomers. The first run had 865 remotely listed members, 426 successful observations, two
listed without a remote endpoint, 30 delisted, and zero cap skips. The committed sorted panel
has SHA-256 25bb0a25be38852d85de29c4367e590ab3da8bb765885c44216215dcd0e61ec5.
Period and vantage. Two June baselines precede the gap through July 1. Incomplete runs are excluded. The stable July 9 anchor starts the historical series. Collection was scheduled every six hours through the September 5 cut and the initial November recovery, and is daily from the first fixed-panel run, from the same US vantage. Six scheduled runs were missed on August 19 and 20. A storage issue stopped collection after September 5 until November 14. See the readout for the coverage comparison.
Hash scope. SHA-256 of canonical JSON containing tool name, description and inputSchema only.
Changes in result envelopes, capabilities, output schemas, HTTP headers and server behavior are
outside the fingerprint. The current probe uses read-only initialize and tools/list, and never
invokes the tools. A failed fetch does not establish that a definition is gone.
Supports. Dated metadata differences observed on the same reachable server. Use completed run IDs and the reachable intersection of the boundaries; skip unreachable observations when comparing a server and state the gap between the two compared runs.
Does not support. Attribution to the July specification without independent evidence, population prevalence, behavioral safety, authentication-only or runtime-gated changes, or identification of malicious intent from text alone. Verbatim tool text is untrusted input.