Closed research experiment, now a public archive. Paper trading only. No real money was ever involved. Simulated capital of $10,000 USDT (US Dollar equivalent).

Losses are displayed alongside wins. Nothing inside the May 27 to July 20, 2026 window was reset or trimmed. An earlier development run, from February to May 2026, was purged when the project moved from calibration to observation and is not part of this archive.

Observations generated by AI (Claude + Gemini) between May 27 to July 20, 2026. This was exploratory research on a sample of 4 trades. Past simulated results do not indicate future performance.

Closed experiment · Public archive

We gave Claude and Gemini $10,000 in simulated capital. Here is what actually happened.

Over 55 days, two models analyzed crypto markets 327 times and produced 327 observations. 296 of them were HOLD. Ten hard-coded risk checks rejected 24 of the 31 that proposed an action. 4 trades were executed.

The experiment ended on July 20, 2026. Everything it produced is now open and free: every observation, every rejection, every line of the AI's reasoning. No account, no payment, no delay.

What the experiment actually shows

May 27 to July 20, 2026. Final numbers, read live from the archive database.

The AI portfolio
-0.61%
$9,939.06 from $10,000
BTC buy and hold
-13.70%
Same period, same starting capital
Read this before drawing a conclusion

Losing less than Bitcoin here is not evidence of skill. The portfolio held an average exposure of 0.97% of its capital and was in the market only 11.9% of the time. The gap against buy and hold comes almost entirely from not participating.

The honest benchmark is doing nothing at all, which would have returned exactly 0.00%. Measured that way the agent underperformed: it ended at -0.61%.

With 4 executed trades, of which 2 were closed by human intervention, the sample carries no statistical significance whatsoever. Nothing here generalizes.

55
days observed
327
observations
296
were HOLD
4
trades executed

327 analysis cycles produced 327 observations. 296 were HOLD. Of the 31 that proposed an action, the deterministic Risk Manager rejected 24. That left 4 trades.

Last decision recorded7/20/2026, 4:58:03 PM
Sim. HOLDMARKET

Every observation is public, including the 296 HOLDs, the 24 Risk Manager rejections, and the full reasoning chain behind each one.

Free and open. No account, no payment, no tracking.

Final state

The simulated portfolio as it stood when the experiment ended. No real money was ever involved.

Experiment closed on July 20, 2026
$9,939.06
-$60.94 (-0.61%) vs. $10,000 start
Cash available
$9,939
100% of portfolio
Exposed
$0
0% of portfolio
Open positions
0 / 3
All positions closed
Unrealized PnL
+$0.00
Nothing left open
Last recorded posture, before shutdownStrong downtrend

Holding cash. 5 of 5 tracked assets are in a downtrend. The strategy is long-only (spot, no shorting), so in a downtrend it waits for a confirmed reversal rather than buying into falling markets. Staying flat is a deliberate decision, not inactivity — and not a malfunction.

These figures are final. The bot was decommissioned on July 20, 2026 and nothing writes to the database any more.

How it worked

1

Collect

Price action, on-chain data, news feeds, and market sentiment, gathered every 4 hours.

2

Synthesize

Gemini (Google's AI) compressed thousands of data points into a structured market summary.

3

Decide

Claude (Anthropic's AI) analyzed the synthesis, weighed converging factors, and output an observation, or held.

4

Validate

10 hard-coded risk checks: position sizing, invalidation zones, exposure limits. Zero AI, pure code.

5

Publish

Approved observations went to the Telegram channel and to this dashboard. All of it is now public here.

The hypothesis, and the verdict

20s

Speed of synthesis

Every 4 hours, the AI ingested technical indicators, on-chain metrics, market sentiment, and breaking news from multiple specialized sources, in under 20 seconds. The pipeline worked as designed. Whether that speed produced better decisions is a question 4 trades cannot answer.

90%

Multi-source convergence, and its cost

A simulated buy required convergence across technical analysis, sentiment, on-chain data, and news flow. One factor alone meant no action. The result was 296 HOLDs out of 327 observations, roughly 90%. The agent was far more cautious than intended.

10 checks

Deterministic risk control

The AI proposed, deterministic code enforced. Every decision passed through 10 hard-coded checks: position sizing, invalidation zones, exposure limits, drawdown rules. No override, no exceptions. The checks rejected 24 proposals, which is most of the reason the capital stayed idle.

Equity Curve

Tracks how the simulated $10,000 portfolio evolves over time, compared to simply holding Bitcoin. No real money is involved.

Sim. PnL
-0.61%
Sim. Win Rate
25%
Sim. Trades
4
Days
55
Sim. Max Drawdown
0.61%
Sim. Avg Win / Loss
0.74

Sim. Monthly PnL

Month-over-month change in simulated portfolio value, as a percentage of starting capital.

Final portfolio breakdown

Simulated Portfolio

Paper trading only. No real money was ever involved.

Total
$9939
Available
$9939
In positions
0.0%
No open positions

Sim. Allocation

Final distribution of the simulated portfolio across open positions and idle cash. Across the whole experiment, average exposure was 0.97% of capital.

Observations

All values are parameters of the AI's simulated paper-trading model. Not trading recommendations. Terms

This table is best viewed on a desktop screen.

Trades

Simulated paper trades. No real money was ever involved. Only 4 trades were executed over the whole experiment, so these results carry no statistical significance.

This table is best viewed on a desktop screen.

Explore the full archive

The tables above are a preview. The explorer holds all 327 observations with filters and the complete reasoning chain behind each one. The analytics page breaks the results down by asset, by hour, by close reason, and by prompt version.