The lab grades itself in public
Paper trading only. Simulated results from past data are not live trading. Education only, not advice.
Trunkline runs a small research lab: 17 trading indicators, each backtested every day on 48 combinations of coin and timeframe; 3 of them also paper trade forward every 15 minutes. No real money is traded. The point is to learn in the open, so the misses are published with the same weight as anything that works.
What the lab got wrong, dated
2026-11-01. The scoreboard average was misleading. The public scoreboard printed one average per indicator across 48 cells, and a single extreme cell pulled some averages up. The typical cell told a different story. We took the figures down the same day. They return only when the lab’s own control run and the page agree, and the cause is filed.
2026-11-01. The learning loop was not learning. The daily learning job had run 52 times and read zero signals each time, because it looked for a file nothing wrote. Four experiments started in July were never judged. Fixed in code the same day; the first daily run on the new loop is 2026-11-02.
2026-11-01. Two measurement bugs. A time limit on paper trades never fired, and the scorecard charged fees once per round trip instead of on each side. Both were fixed on 2026-11-01. Paper trades made before that date carry the old exit bug, and they leave the 28-day scoring window by 2026-11-29.
2026-11-01. No comparison to simply holding. Profits on long trades could not be told apart from the market rising on its own. From 2026-11-02, every score sits next to what holding the same coin would have done.
How the lab grades itself now
Every day at 8:15AM CT one job reads the day’s paper trades. It scores every version of every indicator on data the version was not built on, after fees on both sides, and against holding the coin. A version is named champion only when it clears a fixed bar: 20 or more trades in 28 days, positive after fees, a t-score above 1.0, and better than holding on at least half its coins. Challengers change one setting at a time and are retired after 28 days unless they beat the version they came from. Nobody edits the rule by hand. As of 2026-11-01: 0 champions, 3 challengers. Every promotion and every retirement is written down as a dated row.
The bar before anything is sold
No indicator is offered as a product until a version beats simply holding, after fees, on held-out data, for 30 days, with its dated receipts. As of 2026-11-01, no version meets that bar. This line changes when the record does.
Check it yourself
System status: the health of the machines, including what is red today. Nightly receipts: the machine’s own nightly check of the public pages, raw. Bot scoreboard: backtested figures, withheld while under reconciliation.
The same habit runs Namebeam, a dated record of which businesses AI engines name, another Carter Enterprise property.

