In Wattenhofer & Kniep's New Constellation (v1.0), every 0.8 s window has k co-located concurrent proposers. A user splits a threshold transaction into n fragments (any t decrypt it) and sends them to n of the k proposers. If the proposer forwards user traffic to an external scheduler/block builder, that builder sees the fragment. Hiding fails when one colluding party receives ≥ t fragments. Validator snapshot: Nov 20 2026, 9:49 AM ET. Alternative strategy: send the whole transaction to one trusted proposer (trusted-proposer explorer).
AgaveBam, FireBAM) advertise the BAM node's TPU;
Harmonic validators (HarmonicAgave, HarmonicFiredancer) forward to all connected builders.
GBX-connected validators stream their transaction flow to every GBX builder (Titan and others) through the Gravity Connector
sidecar. Other JitoLabs (block-engine mode), Agave, Firedancer, Rakurai and others schedule
user traffic locally and only see their own fragment.JitoLabs validators to BAM's party (Jito Labs runs both), or as a separate "Jito" party when BAM's TEE is trusted. "GBX colludes" (on by default) pools every GBX-connected validator's fragments,
since all GBX builders see the flow; there is no TEE to trust. GBX validators use GBX's local block-engine proxy instead of a
public Jito block engine, so they are never in the Jito party.clientIds and puts
at most t−1 fragments on any one party; it fails only when there aren't enough distinct parties to place n fragments that way. GBX isn't in clientId, so a smart
user also reads Titan's public GetConnectedLeaders.getVoteAccounts activated stake and getClusterNodes clientId for 661 validators at
the snapshot time; clientId is self-reported. GBX has no clientId (the Connector runs alongside stock Agave /
Jito-Solana), so the 14 GBX-connected validators (3.2% of stake; 13 JitoLabs, 1 Firedancer) come from
Titan's keyless GetConnectedLeaders at Nov 20 2026, 11:45 AM ET, with every returned slot
matching getLeaderSchedule. They are labelled JitoLabs+GBX / Firedancer+GBX. Hub membership is the nearest enabled hub (no distance cap by default, as in
the paper and slides; optionally only within a radius) using ipinfo.io geolocation of the gossip IP, which is approximate (anycast,
provider HQ addresses). Validators without geolocation, or outside the radius, get no windows in hub mode. This snapshot's
closest-hub stake differs a little from the slides' validators.solutions numbers (e.g. FRA 41.5% vs 46%). Distribution overrides rescale stake inside one hub only (validator count and hub total stay fixed).scripts/gbx-leaders.ts)