Lab 0.1.0-252: Registry, Leveraged Grid, Stricter Judge

Lab 0.1.0-252, released on 27 October 2026, opens the public /verified registry to certificates from versions 251 and 252, lets AIs read a long answer page by page, simulates a long grid with leverage, liquidation and dated funding, and makes the statistical judge stricter on noise. It is free, for Windows and Linux.

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

At a glance

Version and download

• Version 0.1.0-252, build 20261027-252-n1, released on /en/download on 27 October 2026.
• Windows: EXE or MSI installer, unsigned (SmartScreen may warn), SHA-256 checksums published on the page.
• Linux: AppImage and .deb package, CPU engine without GPU.
• The machine-readable manifest is strategy-arena-desktop-release.json.

The /verified registry accepts recent certificates

The /verified registry recomputes every certificate submitted to it (public candles, replay by the Lab engine, trade-by-trade comparison with TradingView's Pine Logs) and publishes the verdict as it is. It now accepts certificates from Lab 0.1.0-236 to 0.1.0-252; those from 251 and 252 were added on 26 and 27 October 2026.

To get there, the registry received its own binary, built from the 252 sources: the site's previous engine replayed none of the 15 real 251 certificates (0 of 15). With the new binary: 15 of 15 for 251, 15 of 15 for 252, and forged test certificates are refused by name.

Long answers for AIs: budget and pages

The research_project MCP tool accepts two parameters:

• maxResponseBytes: 12,000 bytes by default, up to 64,000;
• marketCursor: reading market by market.

A project judged on 8 markets can be read in full, at once at 64,000 bytes or in 8 pages. With no declared budget, the answer is byte-identical to 0.1.0-251. The tool list (tools/list) keeps the same size: 43,039 bytes in the full profile, 15,971 in the essential profile.

Leveraged grid: "paste your exchange grid"

A long grid can carry leverage from 1 to 125, with:

• maintenance margin (0.4% by default, OKX BTC-USDT tier 1) and liquidation fee (1.25% by default, Binance BTCUSDT);
• liquidation checked at the open, inside the bar and at the close;
• real dated funding, Binance or OKX, every 8 hours, checked against an independent Python calculation (0 difference on 5 windows).

The result card gives the worst drawdown, liquidations, funding paid and received, the comparison with holding, and the exchange's figure brought back to its window. The Lab never annualizes a window shorter than 90 days.

Limits, written in the result:

• liquidation reads candle prices, not the mark price, which is stricter;
• the embedded funding table ends on 26 October 2026 at 08:00 UTC: beyond it, a leveraged run is refused by name, never invented;
• CPU only: the GPU and the Pine export refuse these grids by name;
• access through the local MCP server only, no button in the interface yet.

To understand what a grid bot card leaves out, with a worked example: leveraged grid bots, what the profit card leaves out.

A stricter judge on noise

The Lab's judge decides whether a strategy beats holding over a sealed period, after counting the variants tried (SPA and StepM tests). In 0.1.0-252, the verdict SPA, the training SPA and StepM take the stricter of two methods: the 0.1.0-251 one and a prewhitened AR(1) sieve bootstrap-t. The judge can therefore only remove "significant" readings.

Measured on 10,000 noise markets per case, with autocorrelated noise:

On the 54 proofs of the test notebook, no verdict changed. Reserve: the test size is not yet under 5% everywhere. The upper bound of the Wilson interval is under 5% on 7 of 12 noise cases; on the other 5, the measured rate ranges from 4.58 to 5.44%.

Same figures on Windows and Linux

exp, ln, pow and log10 are now computed with basic IEEE operations only, the same way on both systems, for everything that enters a proof fingerprint. Result: 166,615 reference values byte-identical between Windows and Linux, and 400 receipts made by 0.1.0-251 re-verified without a difference by 0.1.0-252.

What does not change

• The Pine exports of the test battery are identical to those of 0.1.0-251 (68 of 68).
• Trade-by-trade comparison with TradingView remains the published method: identical trade lists (258/258, 97/97 and 26/26 on 13 October 2026; 62/62 on a buy stop order with Lab 0.1.0-249), method described in reproducible TradingView backtest, trade by trade.
• The Lab places no orders. The MCP server listens only on 127.0.0.1, with a session token.

What comes next

The next version of the Lab is in preparation. It is about exchange grids: reading a grid bot card and re-judging the same grid over the whole history. It will be described here when it is released, not before.

FAQ

When was Lab 0.1.0-252 released?

On 27 October 2026 on /download (build 20261027-252-n1), for Windows (unsigned installer, SHA-256 published) and Linux (AppImage and .deb, CPU engine).

Does the /verified registry accept 252 certificates?

Yes. The registry accepts certificates from Lab 0.1.0-236 to 0.1.0-252; those from 251 and 252 were added on 26 and 27 October 2026. The server replays them with a binary built from the 252 sources.

Does the 252 judge change verdicts?

It can only remove "significant" readings: it takes the stricter of the 0.1.0-251 method and an AR(1) sieve bootstrap-t. On the 54 proofs of the test notebook, no verdict changed. Reserve: the test size is not yet under 5% on every noise case.

Is the leveraged grid in the interface?

Not yet: in 0.1.0-252 it goes through the local MCP server, which your AI calls. It covers long grids with leverage from 1 to 125.

Sources

• Public changelog, "Lab 0.1.0-252" entry of 27 October 2026: /en/changelog (data: /api/changelog).
• Release manifest: /static/downloads/strategy-arena-desktop-release.json.
• TradingView parity figures: /llms.txt, "Positioning" section, dated measurements.