Lab 0.1.0-254: search, and a replay that follows the Pine

Lab 0.1.0-254, published on 1 November 2026, adds search to look for settings while counting every trial, and its replay follows a Pine's levels better, with a stated limit.

Published on 1 November 2026. Research on past data, not investment advice.

In short

• search, an MCP tool to look for settings. An AI declares a script, a market and a budget. The Lab runs elimination rounds, counts every setting evaluated, neighbours included, and never reads the sealed bars (30 % of the period). The proof that follows carries this trial count into its verdict. The tool sits outside the essential profile: call it by name; no button in the interface.
• The GPU wins in search above a measured size, on this machine. Measured on 1 November 2026, one RTX 4080 SUPER, real BTC with 2,558 to 14,000 training bars, 3 rules, total time, results identical to the CPU (same fingerprint). Against 6 CPU threads: the GPU wins everywhere from 2,048 trials; against 16 threads on a free machine: everywhere from 4,096 trials. Below that the outcome varies, and at 256 trials the CPU always wins. auto takes the GPU when trials × training bars ≥ 1,000,000 × CPU threads, and says so (backendUsed, backendReason). Another machine moves the crossover, not the result. Not measured on expression rules or under 2,558 bars.
• := on an entry signal is refused. A name read by an entry order then changed in the script gets a named refusal, PINE_ENTRY_SIGNAL_REASSIGNED. Write the condition as one assignment.
• The false mirage of frozen parameters is fixed. A setting equal to the script as written is no longer drawn or counted as a second candidate, which gave a 100 % PBO through a mere tie. Bench of 1 November 2026, zero fees, Gaussian noise, 6,000 1h bars: at strength 0.0012 a rule with a planted edge goes from 0 % to 84 % survives over 50 cases, the same as without params:. On noise, 0 % survives over 200 cases in each mode. No threshold of the judge moved.
• Levels computed from position_avg_price follow TradingView, in the replay that follows the script. Two modes: next_open_levels_next_bar and signal_close_levels_next_close. Checked on 1 November 2026 on the real TradingView, BINANCE:BTCUSDT 1D, whole history: 916 / 916, 156 / 156 and 10 / 10 trades on the same dates; 8 prices differ in December 2017, because of the candles. Visible limit: without pineSource or replayArguments, the default replay differs, up to the sign of the result, and the Lab says so (in that check, 4,335.19 against 13,968.52 on TradingView, from 10,000). These two modes run on the CPU: the GPU refuses them by name. search has no fillModel.
• The verified_ema and macd_rsi samples no longer have a stop or a target. The default replay follows them: checked on 1 November 2026, 186 / 186 and 30 / 30 trades identical in dates and prices. The other 6 samples (sma, ema, rsi, macd, combo, short_combo) keep a strategy.exit without process_orders_on_close: this form was not measured on TradingView.
• search on a small space never says survives. With 3 settings a single candidate remains: the verdict is inconclusive, even on a strong edge (bench of 1 November 2026: 0 % over 100 cases). Widen the space.
• Other fixes. Pasted Pine goes through a stricter entry: 1 false acceptance over 288 scripts, against 73 in 0.1.0-253 (measured on 31 October 2026). Average-price series are flagged and cannot say survives in certified mode. A proof whose replay fails no longer opens its test period. compose_portfolio follows the test start of a band. A plausibility alert and a power estimate in years sit next to the verdict, without changing it.
• Reserves. The judge was checked on one rule family only on the bench; the GPU was not replayed on the bench. The interface and First Proof were not replayed with these samples on TradingView; only 1D and BINANCE:BTCUSDT are checked there. Windows installers remain unsigned.

Linux downloads

The standard package (.deb, .AppImage) needs no NVIDIA driver. Two separate packages, strategy-arena-desktop-0.1.0-254-cuda-amd64.deb and .AppImage, are built with CUDA and require libcuda. Checksums: SHA256SUMS-254.txt.

What was checked

On 1 November 2026, on the installed build: the interface really opened on Windows, a Pine analysis, a reassignment refusal, a proof that follows its Pine and search on the GPU through a real MCP client. On Linux, search through the real MCP client on the CUDA binary. The full test battery ran on the code just before the version stamp, not on the final freeze. The standard Linux binary only started under Xvfb, with no MCP path, and the Windows path was not replayed button by button in the interface.