Measured. Per-peer bytes come from aria2-next 2.5.6's getPeers, which exposes a per-connection cumulative uploaded counter. IRIS samples it and accumulates the result into a durable ledger, so the totals here are observations, not estimates. Peer role is aria2's own isSeeder.

Ephemeral source, hence the untraced residue. The origin knows exactly how many bytes it uploaded; what needs a live connection is naming which device received them. getPeers only returns live connections, so a peer that connects, takes bytes and leaves between two samples takes its counter with it: those bytes certainly went out, but no recipient can be named for them. Measured origin-side capture: 73.3% at 3 s sampling, 88.1% at 2 s. The shortfall is published as the untraced-residue counter and shown on this board rather than being folded silently into someone's total. Traced + untraced = the origin's own upload counter.

Derived, and labelled as such. Panels labelled "derived" use device-reported delivered bytes minus the measured origin contribution — arithmetic on two measured quantities, not an observation of peer-to-peer traffic. The origin cannot observe device-to-device transfers. The separate measured peer row uses receiving devices’ iris.device.peer_transfer_record captures, retaining each transfer/peer’s maximum cumulative bytes. These captures provide direct evidence; missing peers and capture_complete=false mean incomplete coverage. Origin and unknown bytes remain in the measured share denominator. Do not add origin ledger values to these captures.

Window sensitivity. The ledger tiles select the latest cumulative counter sample within the chosen window. Values include bytes from before that window; they are not window deltas. Changing the bounds can select an older value or no sample at all. Seeder RPC alone uses the fixed last 15 minutes ending now, independent of the picker. The event-side panels still depend on which reports land in-window, and a device's terminal report arrives at completion, so a window that clips the end of a transfer understates the peer share. For a rollout, pick a window that fully contains it.

Peer cap. The ledger caps distinct peers per torrent. When that cap is reached the excess peers stop being tracked individually and their bytes fall into the untraced residue. This is measured directly: Peer ledger cap saturated reads 1 while the cap is refusing peers, which is the only way to tell refused peers apart from missed short-lived connections inside the residue.

Idle is not broken. Ledger counters retain history when sampled in the selected window. No metric sample or peer capture means unavailable, not zero traffic. If Seeder RPC reads 0, current origin sampling is unavailable; historical samples and independent device captures can still be valid.