skip to content

docs

Mesh consensus

Stake-weighted median with 10% cap, 50% quorum, outlier gating. Why no whale validator dominates.

FIGstake-weighted mesh consensus
mesh consensusvalidator 1 · acceptvalidator 2 · acceptvalidator 3 · outliervalidator 4 · acceptSTAKE-WEIGHTEDmedian + outlier gateCONSENSUSone verdict / completionoutliers gated out before they move weight

One validator deciding a verdict is single-point-of-failure. Reliquary aggregates across every online validator using a stake-weighted median — the statistical midpoint of the mesh's opinions, weighted by each validator's stake.

Cap

No validator's stake counts for more than 10 % of the capped mesh total (MESH_STAKE_CAP_FRACTION = 0.10). A 90-percent-stake whale buys exactly the same voice as a 10-percent-stake honest node.

Quorum

Aggregation requires 50 % of the capped mesh stake to be online (MESH_MIN_QUORUM_STAKE_FRACTION = 0.50). Below that, no verdict publishes — better to stall than to let a minority rubber-stamp a window.

Outlier detection

Any validator whose verdict diverges from the mesh median by more than 0.25 euclidean distance in score-dict space is flagged as an outlier for that completion.

Auto-gating

A validator whose per-window outlier rate crosses 5 % (MESH_OUTLIER_RATE_GATE = 0.05) is gated for 12 windows, during which their votes are excluded from the median.

Empirical

Our mesh-live v2 audit ran 4 validators (3 real Blackwell hosts + 1 simulated malicious) processing 256 verdicts in one window. The malicious validator, despite 10% stake, was cleanly gated without affecting the honest consensus.