# How we rate a venue for automated trading

Updated: 24 September 2026 / Signal Grid research

Eight criteria, fixed on 23 September 2026 before a single venue on this site was marked under them. They
are weighted to a hundred, and the eight marks they produce are reported as one total out of
40, rounded to one decimal place. This is version 2.0 of the
scale. The weights are the same on every page here: what a page changes is which venues appear on it
and, sometimes, which reading sets their sequence. Moving a weight would make a new version with a new
date, not a quiet edit, and every total on the site would be worked out again with it.

That is what happened on 23 September 2026. Version 1 scored six things and none of them was the subject
this site is named for: what a venue automates for a trader who writes no code, and what its order
interface tells a strategy that writes its own. The defect was a site about automated trading that
measured neither. It was fixed by adding two readings, not by moving the six already there.

The weights were settled before any venue was read against them, and every total was rebuilt from the
new set rather than nudged. The board moved at both ends: Lighter led version 1 and is third here,
Aster leads now on a grid tool of its own, and EVEDEX rose from seventh to second as the one venue
here that wrote its own strategies — while still documenting neither order control.

## What each dial holds of the hundred

| Criterion | Weight | The reading behind it | Where that reading comes from | Marks it actually spanned |
|---|---|---|---|---|
| What the venue automates for you | 24 | What a trader who writes no code is handed — a strategy the venue itself authored and the trader parameterises, a grid tool the trader must design a rule with, capital deposited into an algorithm the venue runs, or nothing of the venue's own. | Each venue's own documentation and, where it has one, its live product page, read on 23 September 2026. | 0.0 to 10.0 over the ten marked, 3.0 to 10.0 over the eight on the bot-platform grid |
| Programmatic access | 14 | What the documentation gives a strategy that places its own orders — whether a rate limit is published in the venue's own units, and which of post-only, reduce-only and TWAP the order interface documents. | Each venue's own developer documentation and published interface specification, read on 23 September 2026. | 6.0 to 10.0 over the ten marked, 6.0 to 10.0 over the eight on the bot-platform grid |
| Measured liquidity | 22 | What the depth gauge reads on the BTC perpetual — dollars resting inside 10 basis points of mid, bids and asks together, taken as the median of our own polls of the public book. | Our own snapshots of each public BTC perpetual book, taken at ten-minute spacing, the last reading dated 22 September 2026. | 0.0 to 6.9 over the ten marked, 3.9 to 6.9 over the eight on the bot-platform grid |
| Round-trip cost | 16 | What one automated cycle pays the venue before the book has its say — the entry-tier taker rate on the BTC perpetual, charged once to get in and once to get out, shown in basis points. | Each venue's own fee schedule, read on 18 September 2026. | 3.0 to 10.0 over the ten marked, 6.0 to 10.0 over the eight on the bot-platform grid |
| Security and record | 12 | Three warning lamps rather than a reading — what outside auditors have published, whether a bug can be cashed in at the venue, and what the public record of the last 24 months holds. | Audit reports, bounty pages and the public incident record of the 24 months to 18 September 2026, and CertiK Skynet where it carries a listing for the venue. | 4.0 to 10.0 over the ten marked, 7.0 to 10.0 over the eight on the bot-platform grid |
| Custody and settlement | 8 | Where a strategy's collateral stands while it keeps sending orders — how often the venue commits to a chain, what route out exists without the operator, and who holds the deposits, all from its own documentation. | Each venue's own architecture and withdrawal documentation, read on 18 September 2026. | 5.0 to 10.0 over the ten marked, 5.0 to 10.0 over the eight on the bot-platform grid |
| Market coverage | 3 | How long the instrument list runs for a strategy rotating out of one margin balance — perpetual markets counted on the venue's own public pages, and the asset classes among them that are not crypto. | Each venue's own public market endpoint or trading terms, counted on 18 September 2026. | 5.2 to 10.0 over the ten marked, 5.2 to 10.0 over the eight on the bot-platform grid |
| Access | 1 | The gate in front of the first order — how a trader signs in, what identity papers are wanted before trading, and how far the venue's own list of places it will not serve extends. Five points at most, doubled onto the ten-point scale the other dials use. | Each venue's sign-in flow, onboarding pages and terms of use, read on 18 September 2026. | 8.0 to 10.0 over the ten marked, 8.0 to 10.0 over the eight on the bot-platform grid |


The rightmost column is the part a weight cannot tell you on its own. Whether a criterion lifts one
venue over another depends as much on the distance between two marks as on the share of the
hundred it holds, so a heavy weight laid over marks that cluster does less work than a light weight
laid over marks that do not. On the bot-platform grid the effect is plain. What a venue automates
holds twenty-four of the hundred and its marks there run seven apart; measured liquidity holds
twenty-two and its marks run three apart; programmatic access holds fourteen and its marks run four
apart but sit at the ceiling for six of the eight. Nothing is rebalanced in response: a weight
altered after the marks exist is an answer chosen rather than computed. It is published so the grid
can be argued with on the numbers.

Each criterion is marked out of ten first, on the rules in the next section. The weights then turn
eight marks into one number, which is stated out of 40. The two dials
this version added hold between them what the measured book and the fee hold together, and all four
are things a strategy meets on its first order, not things a venue says about itself.

## How a reading turns into a mark

**What the venue automates, 24 of the hundred.** Eight points for a strategy the
venue wrote and the trader only parameterises, six for a grid tool the trader still has to design a
rule with, three for money deposited into an algorithm the venue runs with nothing to configure,
nothing at all where the venue ships none of the three. A published backtest adds a point, because
it is the only thing a trader can weigh before funding anything, and a published capacity cap adds
another, because a venue admitting its strategy stops working above a size is telling the truth
about it. Aster and ApeX Omni each ship a grid and take
6.0; Hyperliquid, Lighter, edgeX, dYdX and Extended run
vaults and take 3.0; Aevo adds a published backtest
to its vault for 4.0; Paradex ships none of the three and
takes 0.0. EVEDEX runs four strategies of its own, a
capacity cap against each and a seven-day backtest against one, which is
10.0 (EVEDEX product page, read 23 September 2026).

**Programmatic access, 14 of the hundred.** Two points for a documented public
interface over both request and streaming transports, four more for a rate limit published as a
number in the venue's own units, and two for each of post-only, reduce-only and a time-weighted
order the order interface documents, to a ceiling of ten. Seven of the ten venues here reach that
ceiling. edgeX loses the rate-limit points because its documentation describes the mechanism and
prints no threshold, and EVEDEX loses the order-control points: it publishes a limit of 30 heavy
requests per 60 seconds per account, but reduce-only is absent from its documentation, its help
pages and its interface specification alike, and post-only exists in that specification only as an
undescribed string field. Both take 6.0.

**Measured liquidity, 22 of the hundred.** Nothing at $10,000 of
resting depth, and two points for every tenfold step above it, with ten as the ceiling. ApeX Omni's
BTC book held a median $1.7 million within 10 basis points of mid and takes
4.5; EVEDEX held $19.0 million and takes
6.6. Ten times the resting depth is worth a little over two
points, because the distance between a thin book and a usable one matters more to an unattended
strategy than the distance between a large book and a larger one.

**Round-trip cost, 16 of the hundred.** Ten points for a round
trip of two basis points or less, half a point off for each further basis point, and nothing left
at twenty-two. Lighter's Standard account charges 0% maker / 0% taker, so two crossings cost
nothing and the mark is 10.0. A 0.045% taker fee makes a round
trip of 9 basis points and a mark of 6.5, which is what EVEDEX,
edgeX and Hyperliquid each carry. Aevo's 0.080% taker doubles that round trip to 16
basis points and leaves 3.0.

**Security and record, 12 of the hundred.** Two points for each
audit on file, to a limit of four; three for a bug bounty anyone can claim against; three for 24
months without a loss of user funds. Lost user money means balances drained in an exploit, a venue-run vault bled by price manipulation, or collateral stranded because the venue stopped answering, with the shortfall never made whole. A position closed out by an ordinary price move is not that.
Hyperliquid has two audits and a bounty of up to 1 million USDC, and loses the last three points
because its own vault was down about $4.9 million after the POPCAT episode of November 2025:
7.0 of ten. Extended has two audits, a bounty of up to
$500,000 and nothing lost in the window, so it takes 10.0.

**Custody and settlement, 8 of the hundred.** Four points when
every trade is written to a chain, two when only position snapshots are; three for a documented
withdrawal that does not need the operator; two for audited contracts holding deposits; one for
wallet sign-in. Lighter takes 10.0 on all four. EVEDEX commits
positions to Arbitrum in batches rather than one write per fill, and documents no withdrawal that
skips the operator, so it takes
5.0 (EVEDEX documentation, read 18 September 2026).

**Market coverage, 3 of the hundred.** Up to six points for the
number of perpetual markets, on a curve where 500 markets earns the full six and 50 earns about
four, plus a point for each asset class beyond crypto, to a limit of four. Aster lists
578 perpetual markets across six asset classes and reaches the ceiling at
10.0; dYdX lists 78 markets and one class
beyond crypto, for 5.2. The two hundredth market is worth far
less than the tenth, which is why breadth cannot carry a page.

**Access, 1 of the hundred.** Two points when nothing has to be
verified before the first trade, one for wallet sign-in, two when the excluded places stop at the
sanctioned ones and one when the list runs wider; the total is doubled onto the same ten-point
scale as the rest. ApeX Omni is the one venue here whose exclusions stop at sanctioned places and
it takes 10.0; every other scored venue takes
8.0. At one point of a hundred this criterion separates
almost nobody, and it is weighted to say so.

## Which venues get a mark, and which only get a row

Venues that run a central limit order book, hybrid or fully on-chain, and publish a BTC perpetual book we could read from a public endpoint. Oracle- and pool-priced venues are listed with their facts and left unscored, because the liquidity criterion measures a resting book they do not have.

Six venues in our files are listed and left unmarked. GMX, gTrade, Jupiter Perps and Ostium fill
an order against a pool at an oracle price, so there is no resting book to poll and no spread to
measure. Orderly runs one shared book behind other front-ends and was not among the books we
polled. Drift has been paused since the exploit of 1 April 2026, with its front end now pointing
at a fork called Velocity. Their facts still appear where they belong, with their dates; what
they do not get is a total built out of a reading nobody took.

## Reported on this site, never scored

Open interest and 24-hour volume are each venue's own figure as republished by CoinMarketCap on
22 September 2026. A venue reporting its own turnover is not a reading we took, so those two numbers
stay in a column of their own and touch no mark.

Part of the automation surface is treated the same way. What a venue publishes about its own
automation is scored: the kind of product, the backtest beside it, the capacity cap, the rate
limit, the order controls. What that automation is worth in practice is not, and cannot be from
here. A latency tier earns nothing and is reported with its date — Lighter's free account carries
300 ms of added taker latency where its paid tier carries 140 ms (Lighter documentation, read
18 September 2026).

No return figure is scored or reprinted. Where a venue publishes what one of its own
strategies has returned, this site records that the figure is published there and leaves it there.
We have run no strategy on any venue here, so we have no reading of our own on what a speed tier
or a shipped strategy earns, and a criterion with no reading behind it is an opinion with a number
attached.

## What a mark cannot see

The liquidity mark rests on the BTC perpetual and on nothing else. No other market was polled, and
a venue that is deep in BTC can be thin in most of what it lists — the market counts in the
coverage criterion say how many markets exist, never how much is resting in any of them.

The cost mark uses what a new account pays by default. Volume tiers, maker rebates, staking
discounts and cashback schemes all lower a real bill, and none of them is in the mark, so a venue
that is dear at the entry tier may be cheap to a desk that trades enough to leave it.

The automation mark reads the shape of a product, not its results. A strategy the venue wrote
scores above a grid because more of the work arrives finished, which says nothing about whether
either one makes money; a published backtest earns its point for existing and being checkable, not
for what it shows. The programmatic-access mark has a narrower blind spot with sharper edges: it
reads documentation. A control a venue has implemented but never written down scores nothing here,
and that is deliberate — a strategy cannot call a field whose accepted values and behaviour are
nowhere stated, whatever the engine underneath would have done with it.

Two audits earn the same four points whether they covered a matching engine or a single contract,
and an audit records that a review happened, not what it found afterwards. The same flatness runs
through the custody rubric: a documented exit that skips the operator earns its three points from
the documentation, not from anyone here having used it.

## What sets the order on each page

A page built around one reading is ordered by that reading; every other page is ordered by the
total. Whichever it is, the criteria, the weights and the marks are unchanged — a venue carries one
score across this site, and a page decides who stands beside it, never what it scored.

| Reading | Pages it orders |
|---|---|
| Where a venue finished across all eight dials, out of 40 | Home, Built-in bots, dYdX alternatives for bots |
| What the venue automates for the trader | Ranking, Automated |
| What the entry-tier taker rate takes out of a $10,000 cycle | Fees, Market making |
| Our own depth reading, median dollars inside 10 basis points of the BTC mid | Algo venues |


More than one page can lead with the same reading. What a page may not do is order itself by one
reading and caption itself with another: the heading above each table names the number that set
the sequence.

## Shared places, and where the rounding happens

Venues that read the same on whatever sets a board's order hold one place on it, shown as a tie. The sequence inside a shared place carries no meaning of its own: where
one of the eight dials sets the order it runs by the site total, highest first; on the fee-ordered
pages, where the taker and the maker rate are both level, it runs alphabetically.

Rounding happens once, at the end. The eight criterion marks are carried into the weighted total
unrounded, and only the total is rounded, which is how two venues arrive at the same number from
different arithmetic. As this version stands no two totals land on the same number, so every
tie printed on the site is a tie in the reading that sets a page's sequence: Lighter and Paradex
both publish 0% maker / 0% taker, so the fee-ordered pages give them one place and say so in
the table, however far apart their totals sit — 28.6 against
20.2 of 40.

Where every figure on this site comes from, and what this desk did not do, is set out on
[About Signal Grid](/about).

Corrections and disputed figures: editorial@ellsworthvane.com.

Signal Grid research, 24 September 2026

These panels are paid placement.