Observed as mathematics
ManimGX · WebGPU. First-seen, unique-seen, and requests stay three identities. N / P cycles films.
ManimGX needs WebGPU. The Plotly timeline above stays live.
Loading retained evidence…
Observe. Compare. Understand.
Four lenses on the field. Separate new nickname-days from repeat activity, and discovery from actual tool use.
Figure 01 / Longitudinal observation
Three distinct measures. One shared clock.
Missing buckets are gaps. The latest hour may be partial.
ManimGX · WebGPU. First-seen, unique-seen, and requests stay three identities. N / P cycles films.
ManimGX needs WebGPU. The Plotly timeline above stays live.
First-seen counts new daily nicknames. Unique-seen includes recaptures and must not be summed as new acquisition. Census requests follow the existing ingestion/deduplication policy; they are not an origin access log. Midnight can create new nickname-days for returning clients. Trial shading shows measurement windows, not attribution.
Figure 02 / Event-aligned comparison
Six complete hours before and after each launch. Not a causal-effect estimate.
Shaded band excludes the launch-containing partial hour. Dashed horizontal reference is the observed six-hour pre-launch mean, when available.
| Wave / UTC launch | Doors | Before/h | After/h | Δ/h | Evidence flags |
|---|
Figure 03 / Classification, not identity proof
Counts and shares of retained events. Scanners stay out of the agent bucket.
Client names are declarations, not cryptographic proof. Unknown/none remains visible. Listing URL labels such as /c/cursor are never treated as provider identity.
Figure 04 / Utility-level observation
Days across. MCP doors down. Compare the same doors across protocol stages.
Select a cell to inspect its retained counts.
Handshakes combine initialize and tools/list observations. Calls count tools/call, not successful task completion. Zero means no matching retained observation in a successfully queried day; gray gaps mean no coverage. The ingestion policy deduplicates detailed events, so these are lower-bound observed counts, not a full request census.