Everything we tried.
7 write-ups, three statuses. Falsified is not a failure of the lab - it’s the product.
Designing an AI options trader that cannot go rogue
Seven independent safety rails, each defeatable on its own. Together they leave the current codebase without a live-order path.
Sharpe 15 in training, zero out of sample: how our trading models fooled us
Seven of eight candidate strategies had no edge on held-out data. The most expensive failure was not a bug: we trained on one venue's spot data and executed against another venue's perpetuals.
GPU passthrough into LXC, without the pain
Four identical RTX 5060 Ti cards, one Proxmox container, and a PCIe lane map that quietly decides which card is allowed to talk to which. The passthrough was the easy part.
The look-ahead bug that made our backtest look worse
Our Chandelier exit peeked one bar ahead, cutting winners short instead of inflating the result. A fixed backtest is not a found edge: ETH 4h went from -16.2% to +185.0%.
The zombie feed: how a delisted coin fabricated 314 winning paper trades
A frozen perp at 0.37621 against a live coin near 0.08 booked 314 fake paper wins. Purging them moved recorded P&L from $6,350 to $5,287.
Our LLM Polymarket bot was just reading the market price back to us
Aggregate Brier score 0.12. And it lost to the market in every single edge bucket. A postmortem of an LLM forecaster that turned out to be an expensive mirror.
Everything I got wrong planning a $5.5k AI homelab
Three planning mistakes, one invisible GPU, and seven months of running a 64GB-VRAM box that ended up doing a completely different job than the one I built it for.
Buried, not hidden.
Every falsified experiment stays on the record, with the one lesson that made it worth the funeral.