Fleet-benchmarking agent
Fleet benches that
reproduce everywhere.
rig-bench measures your edge fleet's throughput claim as a real, repeatable benchmark suite — not a single cherry-picked run. Geometric mean across workload sizes, an explicit pairwise latency matrix, and byte-for-byte accuracy checked on every scenario.
[higgsfield: three gold bar-chart columns rising left to right on a black graph field, thin engraved axis lines]
rig-bench · 3-workload suite · 2048/8192/32768 tokens, seed 42
Why teams install
What rig-bench earns back
Geomean across a suite, not one lucky scenario
1.53×–1.60× across three workload sizes (2048 / 8192 / 32768 tokens), geomean 1.57× — every individual scenario required to clear 1.0×, never regress below baseline.
‹cc90940e›Real pairwise latency, not a flat merge constant
A full K×K latency matrix (LAN between edge nodes, WAN to cloud) drives merge-overhead — the slowest link among contributing nodes, not an average.
‹232a36f2›A planning primitive you can call before you commit
plan_shard is exposed over MCP — get the allocation without running the full benchmark or touching the proof chain.
‹1059e4cc›How it works
Three steps. No faith required.
01
Plan
plan_shard computes an equal-finish-time allocation via divisible-load bisection across the fleet's pairwise latency matrix.
02
Benchmark
Every workload size in the suite runs both the all-cloud baseline and the edgeshard plan, checked token-for-token.
03
Aggregate
Geometric mean throughput ratio across the suite, with a hard floor on every individual scenario, is what the gate actually grades.
Proof band
One packet. Every claim accounted for.
packet rig/rig-bench · suite 3 workloads claim "equal-finish-time shard plan sustains ≥1.3x geomean throughput, byte-for-byte accurate" workloads 2048 / 8192 / 32768 tokens · seed 42 ratios min 1.5316x · mean(geo) 1.5721x · max 1.6005x accuracy PASS — exact match, every scenario coverage PASS — full token coverage, every scenario themis SEALED — e9a0de03
Every rig-bench run emits a packet like this — hash-chained, verifier-signed, and queryable from your own dashboard. How ProofPackets work →
Objections
Answered before you ask
How is this different from the edgeshard router?
That's the dispatch layer — one scenario, one scalar merge cost. rig-bench is the measurement layer: an explicit K×K latency matrix, a full multi-workload suite, and a geometric-mean claim that has to hold across every workload size, not just one.
Isn't a single throughput ratio enough?
A single ratio can hide regressions at other scales. rig-bench requires every individual workload in the suite to clear 1.0x — never worse than baseline — on top of the geomean floor, so a suite can't pass by averaging out a bad scenario.
Can I get the allocation without running the whole benchmark?
Yes — plan_shard is exposed as a standalone MCP tool. Call it for a plan without touching the proof chain, then run the full suite when you're ready to seal a claim.
Buyers
Who runs rig-bench
ML Infra Lead
multi-node local fleet
Head of Platform
on-prem inference cost owner
Systems Engineer
capacity planning for edge deployment
Pricing
Start free. Pay when it pays.
Generalist ships today. Domain Expert adds vertical memory + priors. Custom is built to your stack.
Free
Priced after a short conversation
1 workload size, 2-node fleet.
Team
Priced after a short conversation
Full workload suite, unlimited fleet size, MCP plan_shard access.
Pro
Priced after a short conversation
Dedicated verifier, custom latency matrix import, SSO.
Team pricing is set after a short conversation. No countdown — the rate re-prices once the cohort fills.
