Paychannels
benchmarks

Every tier. Every curve.
Nothing merged.

The headline is ECDSA-per-payment, fully-verified payments per second on the hub. Every other tier below is its own shape with its own axis — labelled, sourced, reproducible with one command.

465kpayments/sone hub, 24 cores, every signature checkedmeasured 2026-07-03
5.84Mpayments/s11 hubs, 258 cores, one ClaimRoutermeasured 2026-07-03
32 µsper paymenthub CPU to verify, journal and applymeasured 2026-09-09
22.6kpayments/s/corefleet per-core ratemeasured 2026-07-03
the curve

Add a 24-core hub, add half a million a second. Add a 48-core hub, add a million.

Fully-verified payments/s against hub count. Solid: measured (1, 2, 6, 11 hubs). Dashed: the same per-hub slope continued — 20 hubs is ten million a second, 200 hubs is a hundred million.

1101001M10M100Mhubs (log)1 hub · 24 cores: 465k/s (24 cores, 2026-07-03)1 hub · 24 cores: 489k/s (24 cores, 2026-07-03)1 hub · 24 cores: 449k/s (24 cores, 2026-07-03)2 hubs: 1.27M/s (42 cores, 2026-07-03)6 hubs: 3.11M/s (138 cores, 2026-07-03)11 hubs: 5.84M/s (258 cores, 2026-07-03)11 hubs · 5.84M/s measured20 hubs → 10M/s projected200 hubs → 100M/s projected

points: 11-hub fleet 5.84M/s · 11-hub fleet 5.74M/s · 6-hub fleet 3.11M/s · 4-hub fleet 2.21M/s · 2-hub fleet 1.27M/s · projection = hubs × 500,000

method

How we count.

A payment counts once it is verified, journaled and applied.

Every voucher is an EIP-712 signature over a cumulative balance. The hub recovers the signer (ECDSA), checks the channel, appends the payment to an append-only journal and applies it to the ledger. The timed window starts after the corpus is signed and ends when the last voucher is applied — signing is setup, never counted.

Corpus.

Per hub: 10,000 payers × 100 merchants, 1,000,000 unique pre-signed vouchers per distribution (uniform and hot-payer skew both run; the headline is uniform). Fleet hubs get disjoint payer sets and share the merchants. Every archive JSON carries the cardinality, the hardware, the core count and a scaling curve.

Per-process microseconds.

Where a row quotes µs per payment, it is Δ process CPU time over Δ payments applied across a clean window, read from ps — no profiler attached. The hub's 32 µs is what one core spends to verify, journal and apply one payment.

The tiers, and why they never touch.

  1. Engine-direct — the corpus goes straight into one hub process. The headline unit.
  2. Aggregate fleet — the sum of each hub's engine-direct rate across a fleet; one ClaimRouter settles it.
  3. Hub-ingress — a driver on the hub's gRPC wire with pre-encoded binary items. What a facilitator delivers; no HTTP.
  4. Verified-by-cache — one signature per K payments, the rest served from the signature cache. Its own curve, its own K=1 reference.
  5. Facilitator-boundary — completed /settle round trips through the reference facilitator. Excludes the merchant hop.
  6. Full path — agent → site → facilitator → hub over HTTP. What a client observes end to end.
  7. Component-isolated — one component against a stub. Diagnostics, never payments.

Binding rules.

  • Fully-verified and optimistic benchmark modes are ALWAYS reported separately and labeled; never conflated.
  • Raw apply-loop rate may only be reported as "ledger throughput", NEVER as payment TPS.
  • Every benchmark discloses voucher-corpus cardinality, hardware, and core count; multi-core results include a scaling curve.
  • The headline number is fully-verified payments/sec; testnet full-path numbers are NEVER conflated with local engine-direct numbers.
  • Verified-by-cache runs (one signature authorising K>1 payments, requests 2..K verified by the hub's signature cache) are a separate tier from ECDSA-per-payment runs; they are labelled `verified-by-cache`, disclose signatures-per-payment and unique-signature count, and are never merged with or compared to ECDSA-per-payment numbers. The headline stays ECDSA-per-payment fully-verified.
  • Hub-ingress runs (a driver speaking the hub's gRPC ingress directly, pre-encoded binary items, no facilitator, no HTTP) are their own tier, labelled `hub-ingress`; they disclose streams, items per frame, verify threads and a verify-thread scaling curve, and are never merged with or compared to facilitator-boundary, full-path, or engine-direct in-process numbers. The headline stays ECDSA-per-payment fully-verified.
reproduce

Run it yourself.

Every figure on this page came out of a just recipe in the repository. The fleet recipes take a hosts file of bare Ubuntu boxes and do the rest.

the whole payment path — one machine
just demo            # Anvil → deploy → fund → x402 payments → epoch settlement
# → site → facilitator → hub, with exact reconciliation
the devnet, N hubs, locally
just demo-local 4     # ports 8700–8799; HOLD=1 keeps it up for page work
# nginx + indexer + facilitators + hubs, no hardware needed
the gas side — settlement costs
just bench-gas       # postEpochRoot + claimEpoch + one-tx router-claim
the gates that guard the payment path
just sp2-gate && just sp3-gate && just sp4-gate && just sp5-gate
just rearch-gate      # replay, envelopes + x402 SDK, 402 flow, reorgs, async full stack
the evidence behind these tables
# every archived report, matrix and study lives under demo/fleet/archives/
# the transcribed findings live in docs/perf-notes.md
tier 01

Engine-direct — one hub engine-direct · HEADLINE

A pre-signed voucher corpus fed straight into one hub engine (T3): every payment ECDSA-verified, journaled and applied. This is the headline tier and the unit the fleet multiplies.

runheadlinehardwarecorescorpus / shapedatecurve (verify workers)archive
laptop baseline (main @52063f4), T3 fully-verified
payment TPS (fully verified, HEADLINE)		749k
fully-verified payments/s		Apple M5 Max18256 x 100 payers×merchants, 100,000 unique vouchers2026-09-081 verify workers: 53.3k/s2 verify workers: 114k/s4 verify workers: 223k/s8 verify workers: 404k/s16 verify workers: 730k/s18 verify workers: 749k/s118 verify workers749k/s · 14.1×baseline-main-20260908/bench-1788842952.json
laptop baseline (main @52063f4), T3 fully-verified
payment TPS (fully verified, HEADLINE)		733k
fully-verified payments/s		Apple M5 Max182000 x 100 payers×merchants, 200,000 unique vouchers2026-09-081 verify workers: 51.3k/s2 verify workers: 113k/s4 verify workers: 225k/s8 verify workers: 394k/s16 verify workers: 719k/s18 verify workers: 733k/s118 verify workers733k/s · 14.3×baseline-main-20260908/bench-1788842900.json
single CCX63 hub, T3 fully-verified
payment TPS (fully verified, HEADLINE)		533k
fully-verified payments/s		Hetzner CCX63 (AMD EPYC-Milan, 48 vCPU / 24 physical)4810000 x 100 payers×merchants, 1,000,000 unique vouchers2026-09-011 verify workers: 18k/s2 verify workers: 35.2k/s4 verify workers: 71.2k/s8 verify workers: 143k/s16 verify workers: 289k/s32 verify workers: 392k/s48 verify workers: 533k/s148 verify workers533k/s · 29.7×hetzner-ccx63-hub-bench-20260901.json
phase2-hub2-168.119.99.99.bench: single Milan hub, T3 fully-verified
payment TPS (fully verified, HEADLINE)		489k
fully-verified payments/s		x86_642410000 x 100 payers×merchants, 1,000,000 unique vouchers2026-07-031 verify workers: 20k/s2 verify workers: 40.3k/s4 verify workers: 80.6k/s8 verify workers: 159k/s16 verify workers: 317k/s24 verify workers: 489k/s124 verify workers489k/s · 24.4×phase2-hub2-168.119.99.99.bench.json
phase2-hub1-46.225.17.94.bench: single Milan hub, T3 fully-verified
payment TPS (fully verified, HEADLINE)		465k
fully-verified payments/s		x86_642410000 x 100 payers×merchants, 1,000,000 unique vouchers2026-07-031 verify workers: 20.1k/s2 verify workers: 39.9k/s4 verify workers: 80.4k/s8 verify workers: 160k/s16 verify workers: 292k/s24 verify workers: 465k/s124 verify workers465k/s · 23.2×phase2-hub1-46.225.17.94.bench.json
phase2-hub3-204.168.150.195.bench: single Milan hub, T3 fully-verified
payment TPS (fully verified, HEADLINE)		449k
fully-verified payments/s		x86_642410000 x 100 payers×merchants, 1,000,000 unique vouchers2026-07-031 verify workers: 20k/s2 verify workers: 39k/s4 verify workers: 80.6k/s8 verify workers: 159k/s16 verify workers: 316k/s24 verify workers: 449k/s124 verify workers449k/s · 22.5×phase2-hub3-204.168.150.195.bench.json
engine HTTP x402 path (T4, informational)
informational only (NOT a thesis gate)		73k
paid requests/s (fully-verified, server-side)		Apple M5 Max18256 payers × 78 vouchers/payer, 19968 unique payment ids, 18 HTTP workers (driver shares the box)2026-09-08single pointbaseline-main-20260908/bench-1788842952.json
tier 02

Aggregate fleet — add a hub, add half a million a second engine-direct · fleet sum

One hub per host, disjoint payer sets, one shared ClaimRouter. The fleet figure is the sum of every hub's max-workers fully-verified rate; the projection line continues the measured per-hub slope.

runheadlinehardwarecorescorpus / shapedatecurve (hubs)archive
11-hub fleet, T3 fully-verified (sum)
payment TPS (fully verified, HEADLINE)		5.84M
fully-verified payments/s		10× AMD EPYC-Milan Processor, 1× Apple M5 Max258per hub: 10000 x 100 payers×merchants, 1,000,000 unique vouchers2026-07-0311 hubs · per-hub 370k–798k/sfleet-report-phase4-11hubs.json
11-hub fleet, T3 fully-verified (sum)
payment TPS (fully verified, HEADLINE)		5.74M
fully-verified payments/s		10× AMD EPYC-Milan Processor, 1× Apple M5 Max258per hub: 10000 x 100 payers×merchants, 1,000,000 unique vouchers2026-07-0311 hubs · per-hub 409k–760k/sfleet-report-phase4-11hubs-run2.json
6-hub fleet, T3 fully-verified (sum)
payment TPS (fully verified, HEADLINE)		3.11M
fully-verified payments/s		5× AMD EPYC-Milan Processor, 1× Apple M5 Max138per hub: 10000 x 100 payers×merchants, 1,000,000 unique vouchers2026-07-036 hubs · per-hub 423k–787k/sfleet-report-phase3-6hubs.json
4-hub fleet, T3 fully-verified (sum) — phase 2
payment TPS (fully verified, HEADLINE) — provenance: session log only (report JSON lost to out/ wipe)		2.21M
fully-verified payments/s		Apple M5 Max (local) + Hetzner 24c Milan hubs (SMT off)90per hub: 10,000 payers × 100 merchants, 1,000,000 unique vouchers2026-07-03single pointSCALING-SUMMARY.md
(aggregate from session log; per-hub Milan bench JSONs rescued (phase2-hub*.json): 465k/489k/449k, local ≈808k)
2-hub fleet, T3 fully-verified (sum) — phase 1
payment TPS (fully verified, HEADLINE) — provenance: session log only (report JSON lost to out/ wipe)		1.27M
fully-verified payments/s		Apple M5 Max (local) + Hetzner 24c Milan hubs (SMT off)42per hub: 10,000 payers × 100 merchants, 1,000,000 unique vouchers2026-07-03single pointSCALING-SUMMARY.md
(session log only (report JSON lost to out/ wipe))
tier 03

Hub-ingress — the wire the facilitators speak hub-ingress

A driver on the hub's gRPC ingress with pre-encoded binary items: no facilitator, no HTTP, one process. Scales with verify threads; the batch size is the lever.

runheadlinehardwarecorescorpus / shapedatecurve (verify threads)archive
Wave 4 — wave4-laptop (run-matrix.sh) — hub-ingress ladder
hub-ingress — a driver on the hub's gRPC ingress, pre-encoded binary items, no facilitator, no HTTP; its own tier, never merged with HTTP or in-process numbers		424k
items/s, fully verified, at 12 verify threads (34 µs of hub CPU per payment)		Apple M5 Max, 18 cores, Darwin 25.6.0, 26.6.218512 payers, 8 streams × 256 in flight, 64 items per frame; verify-thread ladder 1/2/4/8/122026-09-091 verify threads: 56.1k/s2 verify threads: 112k/s4 verify threads: 206k/s8 verify threads: 350k/s12 verify threads: 424k/s112 verify threads424k/s · 7.6×after-wave4-20260909/matrix-summary.json
Wave 3b — fixes + rerun — hub-ingress ladder
hub-ingress — a driver on the hub's gRPC ingress, pre-encoded binary items, no facilitator, no HTTP; its own tier, never merged with HTTP or in-process numbers		416k
items/s, fully verified, at 12 verify threads (34.2 µs of hub CPU per payment)		Apple M5 Max, 18 cores, 64 GiB, macOS 26.6.2 — release binaries, idle box (wait_idle_box)18512 payers, 8 streams × 256 in flight, 64 items per frame; verify-thread ladder 1/2/4/8/122026-09-091 verify threads: 56.5k/s2 verify threads: 111k/s4 verify threads: 204k/s8 verify threads: 350k/s12 verify threads: 416k/s112 verify threads416k/s · 7.4×after-wave3b-20260908/README.md
Wave 3 — gate-phase After (hub ingress) — hub-ingress ladder
hub-ingress — a driver on the hub's gRPC ingress, pre-encoded binary items, no facilitator, no HTTP; its own tier, never merged with HTTP or in-process numbers		259k
items/s, fully verified, at 12 verify threads (39.5 µs of hub CPU per payment)		Apple M5 Max, 18 cores, 64 GiB, macOS 26.6.2 — release binaries, idle box (wait_idle_box)18512 payers, 8 streams × 256 in flight, 64 items per frame; verify-thread ladder 1/2/4/8/122026-09-081 verify threads: 55.9k/s2 verify threads: 109k/s4 verify threads: 194k/s8 verify threads: 267k/s12 verify threads: 259k/s112 verify threads259k/s · 4.6×after-wave3-20260908/README.md
tier 04

Verified-by-cache — one signature, K payments verified-by-cache

One ECDSA signature authorises a window of K payments; requests 2..K hit the hub's (digest, signature) → signer cache. Its own tier with its own curve — the K=1 reference below runs through the same harness.

runheadlinehardwarecorescorpus / shapedatecurve (verify workers)archive
T3c verified-by-cache, K=8
verified-by-cache — 8 signatures per payment window, requests 2..8 verified by the hub's signature cache (hit ratio 87.5 %); its own tier, never compared to ECDSA-per-payment		3.75M
payments/s, verified-by-cache — one ECDSA per 8 payments (208k per core)		Apple M5 Max (18 cores)182000 x 100 payers × merchants, 200,000 payments, 25,000 unique signatures, cache 262,144 entries2026-09-081 verify workers: 346k/s2 verify workers: 727k/s4 verify workers: 1.29M/s8 verify workers: 2.31M/s16 verify workers: 3.5M/s18 verify workers: 3.75M/s118 verify workers3.75M/s · 10.8×after-phase0-20260908/bench-1788855830.json
tiers.t3c_verified_by_cache.uniform
same-harness ECDSA reference, K=1
ecdsa-per-payment reference inside the verified-by-cache harness (not the T3 headline harness)		741k
payments/s, ECDSA per payment, through the SAME mirror harness — the only in-tier apples-to-apples		Apple M5 Max (18 cores)182000 x 100 payers × merchants, 200,000 unique vouchers, one signature each2026-09-081 verify workers: 51.4k/s2 verify workers: 113k/s4 verify workers: 224k/s8 verify workers: 394k/s16 verify workers: 722k/s18 verify workers: 741k/s118 verify workers741k/s · 14.4×after-phase0-20260908/bench-1788855830.json
tiers.t3c_verified_by_cache.ecdsa_reference_same_harness
tier 05

Facilitator-boundary — through the reference facilitator facilitator-boundary

Completed POST /settle round trips at the facilitator's listener (fresh applies only). Excludes the agent→merchant hop and merchant-server overhead. Not comparable to site-path or engine-direct numbers. The reference facilitator is unoptimised on purpose: it is the surface platforms replace with their own.

runheadlinehardwarecorescorpus / shapedatecurve (workers)archive
Wave 4 — wave4-laptop (run-matrix.sh) — settle-only ramp
facilitator-boundary — completed POST /settle round trips at the facilitator's listener; excludes the agent→merchant hop and merchant-server overhead		60.6k
fresh payments/s at 256 workers (hub 32.3 µs, facilitator 34.8 µs of CPU per payment)		Apple M5 Max, 18 cores, Darwin 25.6.0, 26.6.218ramp --layer merchant-boundary, fresh stack per worker count {32,128,256}; A/B --ingress-window-us {0,100} @ 2562026-09-0932 workers: 54.9k/s128 workers: 59.6k/s256 workers: 60.6k/s32256 workers60.6k/s · 1.1×after-wave4-20260909/matrix-summary.json
Wave 4 — Hetzner three-role fleet — cross-box knee (facilitator host → hub host)
facilitator-boundary, cross-box — the facilitator on one host streaming into a hub on another over the private network; excludes the agent→merchant hop		59.9k
fresh payments/s at 512 workers (hub 97.9 µs, facilitator 74.6 µs of CPU per payment)		Hetzner CCX63 ×2 (48 vCPU, 184 GB) + CCX33 (8 vCPU), Ubuntu, private net 0.5 ms RTT—Hetzner: facilitator on the multi host → engine on the hub host, single-step ramps 32/128/256/512, measurement corpus, reference paused2026-09-0932 workers: 8.67k/s128 workers: 37.4k/s256 workers: 51.1k/s512 workers: 59.9k/s32512 workers59.9k/s · 6.9×hetzner-wave4-20260909/matrix
Hosted devnet, scenario H — reference facilitator knee (one 8-core hub + facilitator pair)
facilitator-boundary — loadgen merchant ramp through the unoptimised reference facilitator into one hub; the landing footnote's 60,000/s per 8-core pair		59.6k
payments/s at 512 workers, 100 % success (p95 2 ms at 32 workers)		Hetzner CCX33 pair (AMD EPYC-Milan, 8 vCPU per role), private network8worker ladder 32/128/256/512/1024; knee rule: first step with p95 > 3 × max(step-1 p95, 1 ms), or success < 0.995, or marginal Δtps/Δworkers < 5 % of the step-1 slope; ceiling = best tps strictly before it2026-09-1032 workers: 19.7k/s128 workers: 52.6k/s256 workers: 58.1k/s512 workers: 59.6k/s1024 workers: 60.4k/s321024 workers60.4k/s · 3.1×hetzner-wave5-20260910/scenarios/H/knee/knee.json
Wave 3b — fixes + rerun — settle-only ramp
facilitator-boundary — completed POST /settle round trips at the facilitator's listener; excludes the agent→merchant hop and merchant-server overhead		59.5k
fresh payments/s at 256 workers (hub 39.9 µs, facilitator 37.7 µs of CPU per payment)		Apple M5 Max, 18 cores, 64 GiB, macOS 26.6.2 — release binaries, idle box (wait_idle_box)18Facilitator-boundary settle-only, ramp --layer merchant-boundary, one fresh stack per worker count (32 / 128 / 256)2026-09-0932 workers: 52k/s128 workers: 57.7k/s256 workers: 59.5k/s32256 workers59.5k/s · 1.1×after-wave3b-20260908/README.md
Wave 3 — gate-phase After (hub ingress) — settle-only ramp
facilitator-boundary — completed POST /settle round trips at the facilitator's listener; excludes the agent→merchant hop and merchant-server overhead		56.5k
fresh payments/s at 256 workers (hub 40.6 µs, facilitator 41 µs of CPU per payment)		Apple M5 Max, 18 cores, 64 GiB, macOS 26.6.2 — release binaries, idle box (wait_idle_box)18Facilitator-boundary settle-only, ramp --layer merchant-boundary, one fresh stack per worker count (32 / 128 / 256)2026-09-0832 workers: 17.5k/s128 workers: 51.8k/s256 workers: 56.5k/s32256 workers56.5k/s · 3.2×after-wave3-20260908/README.md
G0 — wave 3 pre-change baseline — settle-only ramp
facilitator-boundary — completed POST /settle round trips at the facilitator's listener; excludes the agent→merchant hop and merchant-server overhead		48.4k
fresh payments/s at 32 workers (hub 99.2 µs, facilitator 52.8 µs of CPU per payment)		Apple M5 Max, 18 cores, 64 GiB, macOS 26.6.2 — release binaries, idle box18Facilitator-boundary settle-only (ramp --layer merchant-boundary, direct /settle, no shim, no site), one fresh stack per worker count (32 / 128 / 256), 3 M vouchers registered on-chain2026-09-0832 workers: 48.4k/s128 workers: 47.1k/s256 workers: 44.5k/s32256 workers44.5k/s · 0.9×baseline-wave3-20260908/README.md
merchant-boundary soak, driver prebuilt-replay (384 workers)
fully-verified (server-side), facilitator-boundary direct /settle — EXCLUDES the agent→merchant hop and merchant-server overhead; NOT comparable to full-path site→facilitator→hub numbers and NEVER to engine-direct numbers		39.9k
fully-verified payments/s (facilitator-boundary)		Hetzner CCX63 (AMD EPYC-Milan Processor, 48 vCPU) — hub + facilitator share the box48512 payers × 8 merchants × 4000000 unique pre-signed vouchers; batch-settlement scheme, direct /settle, payer-partitioned across 384 worker(s)2026-09-01single pointhetzner-suite-20260901/merchant-replay-final.json
merchant-boundary soak, driver merchant-http (256 workers)
fully-verified (server-side), facilitator-boundary direct /settle — EXCLUDES the agent→merchant hop and merchant-server overhead; NOT comparable to full-path site→facilitator→hub numbers and NEVER to engine-direct numbers		33.3k
fully-verified payments/s (facilitator-boundary)		Hetzner CCX63 (AMD EPYC-Milan Processor, 48 vCPU) — hub + facilitator share the box48512 payers × 8 merchants × 2000000 unique pre-signed vouchers; batch-settlement scheme, direct /settle, payer-partitioned across 256 worker(s)2026-09-01single pointhetzner-suite-20260901/merchant-http-final.json
facilitator-layer ramp probe (laptop baseline) — ceiling before the knee
facilitator layer (via a minimal in-process resource shim; excludes site middleware; NOT full-path) — uncapped max-rate probe		13k
best sustained TPS before the knee (ceiling probe, NOT a headline)		Apple M5 Max18512 demo payer(s) (one per max-worker slot, engine::demo derivation) x 1 merchant x 1 endpoint; channel scheme only, uncapped (no rate limiting, no retries)2026-09-088 workers: 8.56k/s16 workers: 11.8k/s32 workers: 13k/s64 workers: 12.6k/s128 workers: 12.3k/s256 workers: 12k/s512 workers: 11.5k/s8512 workers11.5k/s · 1.3×baseline-main-20260908/ramp-facilitator-1788842830.json
tier 06

Full path — agent → site → facilitator → hub full-path / site boundary

Agent-shaped: driver → site → facilitator → hub over HTTP, rate-limited, retrying. The only tier whose headline is a payment TPS an external client observed end to end.

runheadlinehardwarecorescorpus / shapedatecurve (workers)archive
full-path soak at target 200/s (64 wallets, 2 hubs)
fully-verified (server-side), full-path site→facilitator→hub over HTTP — reported separately from engine-direct local numbers		200
fully-verified payments/s (client-observed, site boundary)		Hetzner CCX33 stack (AMD EPYC-Milan Processor, 48 vCPU reported by the driver box)4864 wallets × 2 hub(s) (sticky) × 3 endpoint(s); channel-dominant + exact trickle2026-09-01single pointhetzner-suite-20260901/amoy-fullpath-final.json
tier 07

Component-isolated — diagnostics, not payments component · NOT payment TPS

One component against a stubbed neighbor — a diagnostic ceiling for that component alone. Never a payment TPS, never comparable to any tier above.

runheadlinehardwarecorescorpus / shapedatecurve (workers)archive
site isolated
component-isolated (stubbed neighbor; NOT full-path; NOT payment TPS) — site middleware overhead (stub facilitator)		10.1k
paid req/s (stub facilitator)		AMD EPYC-Milan Processor48{"endpoint":"/api/paid/time","mode":"site","requests":3000,"workers":8}2026-09-01single pointhetzner-suite-20260901/component-site-1788266808.json
indexer isolated
component-isolated (stubbed neighbor; NOT full-path; NOT payment TPS) — indexer ingest rate + query latency under synthetic load		14.5k
ingest rows/s		AMD EPYC-Milan Processor48{"expectRows":200000,"querySecs":30,"workers":4}2026-09-01single pointhetzner-suite-20260901/component-indexer-1788266884.json
facilitator isolated
component-isolated (stubbed neighbor; NOT full-path; NOT payment TPS) — facilitator verify/settle overhead (stub engine) — routing + serde + HTTP		6.48k
settle req/s (stub engine)		AMD EPYC-Milan Processor48{"mode":"facilitator","requests":5000,"syntheticPayers":64,"workers":8}2026-09-01single pointhetzner-suite-20260901/component-facilitator-1788266797.json
chain path: exact-settle latency + dispenser funding
chain-path bench: exact-settle latency + confirmation distribution + dispenser funding rate — on-chain path, NOT payment TPS		2.02k
exact-settle RTT p50 ms (on-chain path)		AMD EPYC-Milan Processor485 exact settles, confirmation depth 152026-09-01single pointhetzner-suite-20260901/chain-anvil-1788266963.json
archives

The files.

Every row above links to its archive under demo/fleet/archives/ — report JSONs, matrix summaries, knee ladders, READMEs with the raw logs.

Live explorer: hub view, architecture, archived tiers →