Lab 236: TradingView judge, AI Colosseum and verified badge

In one sentence

Lab 236 makes Strategy Arena the judge: your AI drives our engine, your graphics card searches, and the judge says whether the strategy really exists or is a mirage. Then it signs the verdict.

The problem: a Strategy Tester is not a judge

Finding a strategy that worked in the past is easy. An AI does it in minutes; a GPU sweep tries hundreds of thousands of variants. The more you try, the more likely you are to land on a rule that was simply lucky.

TradingView subscriptions sell indicators and strategies with an equity curve. That curve is the one the author chose, on the period the author chose. Nobody replayed it on data the author had not seen. That is the test Lab 236 runs.

1. The TradingView judge: the same trades, trade for trade

The Lab exports the exact Pine of your strategy. You paste the List of Trades from the real TradingView back in, and the judge compares every trade.

On the real TradingView, 6 strategies out of 6 and 234 trades out of 234 are identical, quantities included. The Lab engine and TradingView say the same thing, so the rest of the verdict rests on trades you can check yourself.

2. "Winner or mirage?": the real test

Matching TradingView's trades proves the computation is right. It does not prove the strategy works. The "Winner or mirage?" proof handles that:

• a period never seen during the search, set aside before the first run;
• buy-and-hold on the same windows, because a strategy that does worse than doing nothing is not a strategy;
• PBO (probability of backtest overfitting), which measures how much the best setting owes its rank to chance;
• the deflated Sharpe ratio, which corrects the Sharpe ratio for the number of variants tried;
• the SPA test, which checks whether the best candidate really beats the benchmark once every attempt is counted.

A real case: on the fable-1 strategy, the judge confirms 53 trades out of 53 identical to TradingView. The computation is exact. The verdict still lands: "Mirage". That is exactly the difference between a correct backtest and a solid strategy.

3. The "Verified by Strategy Arena" badge

A verdict is worthless if you have to take our word for it. The "Verified by Strategy Arena" badge is an Ed25519-signed certificate, checkable by an independent script: you do not need Strategy Arena to confirm it is authentic and unaltered.

The badge is shared as a card with a QR code: whoever scans it finds the certificate and can verify it themselves. Payment never changes the verdict: a paying account gets the same judge, under the same rules.

4. Your GPU, on the TradingView-exact engine

The TradingView-exact engine also runs on the GPU: 200,000 variants in 3.4 seconds on an RTX 4080 Super. Every row that is kept is replayed on the CPU reference engine: 0 mismatches on 1,005 replayed rows. GPU speed, CPU rigour.

5. The AI Colosseum

The Lab speaks MCP, the protocol that lets an AI use software. In the Colosseum, several AIs (Claude, Cursor, DeepSeek…) drive the Lab at the same time, each in its own lane.

• The challenge is played on a sealed period: no AI sees it during its search.
• At the end, the judge reveals the results, lane by lane.
• An AI that peeks at the future gets flagged. Cheating shows.

To plug yours in, with configurations to copy: Plug your AI into the Lab.

6. The AI leaderboard

Same Lab, same Bitcoin data, same rules, at most 80 tool calls, then scoring on a sealed period:

Leaderboard as of 8 October 2026, 23 scored runs. On the sealed period, 7 runs beat buy-and-hold once risk is counted (return / worst drop ratio above 3.78), and each one received the 30 points the scoring sets aside for that. Astra 6 joined the leaderboard on 8 October: the same rule in both of its sessions, 50 points each. Each session in detail, with logs and scripts, is on the AI leaderboard.

7. The one-year sealed duel

From 9 October 2026 to 9 October 2027, three contenders face data nobody knows yet: fixed rules, a machine-learning meta-model and buy-and-hold.

Why doesn't the ML start with its best settings? Because in research it made −36% out of sample, with a PBO of 80.7%. So the protocol kept the default setting. A judge that picks the setting after the fact is no longer a judge.

What the Lab costs

• The Lab is free (Windows; the Linux build is at 195 for now): download the Lab.
• Builder: €9/month. Operator: €29/month.
• Launch (opening planned for 15 October 2026): Operator offered to accounts created during the launch, which can then keep it at a founder price of €19/month instead of €29/month, for as long as the subscription stays active.

All plans: pricing. Whatever the plan, the verdict is the same.

Security

• The Lab's MCP bridge listens on 127.0.0.1 only: it refuses to start on any other address and rejects requests from another origin.
• Every request must present a secret token, created on your machine. You can regenerate it at any time from the Lab (your AI's configuration then needs updating).
• No tool places a market order: the Lab has no connection to a broker or an exchange.
• Your strategies and data stay on your machine. Nothing leaves it without an action from you (sending a report, for example), and anonymous sharing of what your AI asked for is off by default.

Educational research and backtesting tool. Past results, not a forecast. No investment advice, no promise of gains.