live · verified · measured

Ride a Signal

One real signal. Nine stops. Click → edge → route → sandwich → LLM → mark → loops → highway → harden. Every step measured, every step honest.

routing <0.005ms mark <0.001ms gateway <10ms 320 tests · <7s

↓ scroll — or skip to the lifecycle speedrun

§ 01 · The Journey

Send a signal. Watch it travel.

Click the button. The signal fires into a real Cloudflare Worker, runs through every layer of the substrate, and streams its timing back over SSE. No mock. No delay. Live numbers.

Same mechanics. Different metaphor. Current lens: signal

Signal Journey

Watch one signal travel 9 stops through the substrate

OUTBOUNDRETURN

clicksignal

—

edgegateway

—

routeselect

—

sandwichADL

—

LLMtask

—

return trip

markmark

—

loopstick

—

highwayfollow

—

hardenSui

—

The 9 Stops

The click

<0.1ms

signal

receiver

emit

Every action in ONE starts as a signal. This is the primitive.

net.signal({ receiver: 'demo:route' })

The edge

<10ms

edge

gateway

KV cache

Signal lands at the nearest Cloudflare edge. No origin round-trip.

Measure the edge yourself:

curl -w "%{time_starttransfer}" https://api.one.ie/health

Route: the formula

<0.005ms

strength

resistance

select

weight

sensitivity

Routing is math. Not ML. Not a coordinator. Just arithmetic.

weight = 1 + max(0, strength - resistance) * sensitivity
const next = net.select()

The sandwich

<1ms

sandwich

ADL

toxic

capability

boost

jam

Four checks. Each cheaper than a database lookup. Only one costs money.

signal ↓

ADL gate

permission check

<1ms

$0.00

↓

isToxic

3 integer compares

<0.001ms

$0.00

↓

capability

1 KV lookup

<1ms

$0.00

↓

LLM

the probabilistic step

~1,500ms

~$0.0001

↓

mark/warn

edge memory

<0.001ms

$0.00

result ↑

4 of 5 layers cost $0.00. One costs physics.

LLM: the slow part

~1,500ms

LLM

task

generation

Everyone's LLM is slow. Ours is the same speed. The difference is everything around it.

// the one probabilistic step
const result = await llm.ask(prompt)

Boost: feedback starts

<0.001ms

boost

jam

chain depth

outcome

closed loop

pheromone

The signal closed its loop. Watch the feedback travel back.

demo → routehighway threshold: 20

strength2.0

awaiting result...

The seven loops

L1-L7

loop

tick

decay

hypothesis

frontier

evolution

know

Your signal is already in two loops. Five more will find it.

L1 signalroute, ask, outcomeper signal——

L2 markstrength/resistance updateper outcome——

L3 fadeasymmetric decayevery 5min——

L4 economicrevenue on pathsper payment——

L5 evolutionrewrite failing promptsevery 10min——

L6 knowledgepromote highwaysevery hour——

L7 frontierunexplored tag clustersevery hour——

loading loop state...

Highway: what emerges

strength>20

highway

depth

cached

hardened

Fifty boosts. One highway. Seventeen times faster.

See live proven paths:

curl https://dev.one.ie/api/export/highways

Harden: Sui proof

4.25s

harden

highway

Sui

Highways harden into tradable IP. One agent's knowledge becomes another's.

// highway → on-chain, immutable
// other agents can license the proven path

Now run it yourself

Run the routing benchmarks:

bun vitest run routing.test.ts

Check the gateway:

curl https://api.one.ie/health

See live proven paths:

curl https://dev.one.ie/api/export/highways

§ 02 · The Math

Why it’s fast

We don’t compete on LLMs. Everyone’s inference runs at physics speed. The difference is everything around it — and everything around it is arithmetic.

Routing is arithmetic

weight = 1 + max(0, strength − resistance) × sensitivity. No ML. No coordinator. Just a single formula applied across proven paths.

<0.005ms / decision

The deterministic sandwich

Toxicity → capability → LLM → mark/warn. Four of five layers cost $0. Only one is probabilistic. The rest is math wrapped around the model.

4 of 5 layers free

Every signal closes its loop

result → mark. failure → warn. dissolved → warn(0.5). Pheromone accumulates on paths automatically, so the system learns as it runs.

43,200 signals/day

§ 03 · End-to-End

Lifecycle speedrun

Full commerce loop on Sui testnet: create wallet, fund, list service, discover, message, sell, subscribe, browse, buy, verify. Ten stages. One continuous run. Real wallets, real digests.

total: <1μs

Zero → Seller → Buyer → Advocate

<1μs

§ 04 · Receipts

Verified benchmarks

Every number below is produced by scripts/speed-report.ts after every bun test. p50 observed vs. budget target · PERF_SCALE=3 · 125 samples.

Routing Layer

budget · p50 observed · replaces

select() from 100 paths

LLM routing (~300ms)

budget 0.005ms 0.007ms

select() from 1,000 paths

search API + rank

budget 1.00ms 0.074ms

follow strongest

keyword search

budget 0.050ms 0.005ms

follow batch (10k)

—

budget 0.005ms 0.005ms

Pheromone Layer

budget · p50 observed · replaces

mark()

DB write (~10ms)

budget 0.001ms 0.45μs

warn()

—

budget 0.001ms 0.22μs

sense()

—

budget 0.001ms 0.04μs

fade() over 1,000

—

budget 5.00ms 0.137ms

highways() top-10

—

budget 5.00ms 0.92μs

Signal Layer

budget · p50 observed · replaces

dispatch (in-memory)

HTTP call (~50ms)

budget 1.00ms 0.72μs

dispatch: dissolved

—

budget 1.00ms 0.12μs

queue round-trip

—

budget 1.00ms 2.0μs

ask: chain-3

3 sequential LLM (~6s)

budget 100ms 0.017ms

Edge · Sui

budget · p50 observed · replaces

edge cache hit

TypeDB RTT (~100ms)

budget 0.010ms 0.52μs

Sui keypair derive

HSM round-trip

budget 5.00ms 2.45ms

Sui tx build

—

budget 0.010ms 0.51μs

Sui sign

—

budget 5.00ms 0.36ms

Full table: /speedtest · source: docs/speed-test.md

Run it yourself.

Don’t take our word for it. Every benchmark above is reproducible in under seven seconds on your machine. The substrate has no secret sauce — just 670 lines of engine and a formula.

math + pheromone + 4-layer sandwich = instant routing, forever learning.