Lab 0.1.0-253: Faithful Exchange Grid, AI Grid Search

Lab 0.1.0-253, released on 27 October 2026, plays an exchange grid the way the exchange does: it pauses buys on a band breach, recentres while keeping the width, liquidates on the mark price, adds MEXC funding, reads a bot card and lets your AI search grid settings, every trial counted. It is free, for Windows and Linux.

Published on 27 October 2026. Research on past data, not investment advice.

At a glance

Version and download

• Version 0.1.0-253, build 20261027-253-n1, released on /download on 27 October 2026.
• Windows: EXE or MSI installer, unsigned (SmartScreen may warn), SHA-256 hashes published on the page.
• Linux: AppImage and .deb package, CPU engine without GPU.
• Manifest: strategy-arena-desktop-release.json.

Paste your grid, the Lab plays it like the exchange

An exchange grid bot does more than place orders between two bounds. When the price leaves the band, it stops, pauses or recentres. 0.1.0-253 adds these rules:

• buy pause (band_breach: soft_buy_pause): on a band breach, pending buys are cancelled, lots and their sells stay, and buys resume after N bars spent in the band;
• recentring with a constant width (recenter: keep_width): the new band keeps the same price width, centred on the close;
• MEXC funding next to Binance and OKX; the source of the funding applied is always written.

Without these options, a 0.1.0-252 grid gives a byte-identical result.

Why it matters: on the 2022 study, the Lab's static grid lost −20.45% where our replica of the bot lost −3.38%. With the band exit rule, the Lab gives −3.94% on 4h bars and −3.58% on 15-minute bars. The gap falls from 17.07 points to 0.56 point (0.20 on 15-minute bars).

How long the grid actually worked

Every grid verdict now states how many days the grid worked, what share of the period that is, and when and why it stopped. Under 25%, a warning is shown: a return earned over a few active days says almost nothing. On the 2022 static grid, the Lab measures 9.32% active time, against 9.35% for the replica.

Liquidation on the mark price

Exchanges liquidate on the mark price, not on the last trade. A leveraged grid is now liquidated on the Binance BTCUSDT mark price (public hourly history) when it covers the bars. Otherwise the Lab falls back to candle prices and writes the reason: mark missing, different market, or bars finer than one hour. liquidation_price: last gives back the 0.1.0-252 rule.

The difference shows on real cases, on 4h bars since 2020: a ×20 grid between 6,000 and 70,000 is liquidated on 19 May 2021 at 08:00 on the mark price, against 04:00 on the last price, 4 hours later.

Reading an exchange card

A bot card shows a percentage, often annualised, without saying what margin sits behind it. research_project {action:"exchange_card"} takes the card's percentage, gain, duration, leverage and side, and returns the margin committed, the notional, the move of the underlying and the liquidation distance.

On an example OKX card from 21 October 2026 (+8.96%, 5 days 20 h, ×10 long):

The Lab never annualises a window shorter than 90 days: it shows the card's figure, then brings it back to its real window. How to read a card, with a computed example: leveraged grid bot, what the card does not say.

Your AI searches grids, every trial counted

research_project {action:"exchange_grid"} lets an AI declare a space of settings: bounds, levels, leverage from 1 to 125, base position, band exit rule, funding. It also sets a trial budget. Then:

1. each setting is written to the trial registry before it is computed;
2. it is run over the history with margin, mark-price liquidation and dated funding;
3. it is judged against holding by the 0.1.0-252 judge (SPA and StepM, the stricter of the two);
4. a setting liquidated at least once is never ranked "survives".

Measured example: 50 settings drawn around the OKX card, 45 evaluated and counted (5 invalid), 18 liquidated at least once, 0 "survives". The best one gets an adjusted p of 62.0% (SPA) and 63.6% (StepM). Same seed, byte-identical receipt.

Reserves

• The grid search describes the past over the whole history: no period is sealed, it is not a proof.
• No button in the interface: these tools go through the local MCP server, which your AI calls.
• The mark price is Binance BTCUSDT, also used for an OKX or MEXC bot; the note names the instrument.
• MEXC history starts on 5 May 2025, and the tables stop on 26 October 2026. Beyond that, a computation is refused by name, never invented.
• CPU only: the GPU and the Pine export refuse these rules by name.
• Only long grids are simulated. A short card is refused by name.

What does not change

• The Pine exports of the test battery are identical to 0.1.0-252 (68 out of 68).
• The judge bench gives the same lines as 0.1.0-252 (3,460 out of 3,460), in free and certified modes.
• exp and ln stay byte-identical between Windows and Linux (166,615 reference values).
• The /verified registry now accepts certificates from Lab 0.1.0-236 to 0.1.0-253; those from 251, 252 and 253 replay identically, 15 out of 15 each.
• The trade-by-trade comparison with TradingView remains the published method: identical trade lists (258/258, 97/97 and 26/26 on 13 October 2026), method described in reproducible TradingView backtest, trade for trade.
• The Lab places no orders. The MCP server only listens on 127.0.0.1, with a session token.

What comes next

The next version of the Lab is in preparation. It will be described here when it is released, not before. The previous version: Lab 0.1.0-252.

FAQ

When was Lab 0.1.0-253 released?

On 27 October 2026 on /download (build 20261027-253-n1), for Windows (unsigned installer, SHA-256 published) and Linux (AppImage and .deb, CPU engine).

What does the grid search by AIs do?

Your AI declares a space of settings and a trial budget through the local MCP server. Each setting is counted in the trial registry before it is computed, run with margin, liquidation and dated funding, then judged against holding. A setting liquidated once is never ranked "survives".

Is the grid search a proof?

No. It describes the past over the whole history, with no sealed period. On the measured example, 45 settings counted, 18 liquidated, none "survives".

Does liquidation follow my exchange's mark price?

It follows the Binance BTCUSDT mark price (public hourly history) when it covers the bars, including for an OKX or MEXC bot; the note names the instrument. Otherwise the Lab falls back to candle prices and writes the reason.

Sources

• Public changelog, entry "Lab 0.1.0-253" of 27 October 2026: /changelog (data: /api/changelog).
• Release manifest: /static/downloads/strategy-arena-desktop-release.json.
• TradingView parity figures: /llms.txt, "Positioning" section, dated readings.