Everything published here is the output of an automated research pipeline that has run 45+ pre-registered experiments on Bitcoin data since early 2026 (count as of July 2026). Most failed. We publish the failures with the same prominence as the survivors — because a null result you can trust is worth more than a win you can't.
Every experiment's hypothesis, thresholds, costs, and success criteria are written down and committed BEFORE any computation. After the lock, the spec can only be abandoned — never quietly amended. No moving the goalposts after seeing the data.
Anything that passes its gates then faces inversion tests, random baselines, parameter perturbation, and stability checks. Of 49 machine-learning model-target combinations tested on the full feature universe, the measured verdict was 0 deployable — a result we consider one of our most valuable findings, and one almost nobody else will tell you.
Testing hundreds of ideas and reporting the best one is how most "backtests" lie. Our pipeline keeps a ledger of every statistical test ever run (500k+ as of July 2026) and applies false-discovery control against the whole pool.
Ideas that survive research run as timestamped forward trials with locked evaluation gates — including a leveraged BTC trade call whose every publication is recorded and scored (T+1) at its own entry, stop and target, with its running hit rate, wins and losses alike.
The daily brief is generated by Claude (Anthropic) reading the pipeline's metrics, regime models, and derivatives data each day. It is labeled as AI analysis because that's what it is. The AI is sometimes wrong; the point is that its record is public and continuous — no deleted calls, no cherry-picking.
Public: the daily brief, regime states, metric dashboards, the Alt-Euphoria Gauge, published-call track records, and research write-ups including nulls. Not public: proprietary model internals and execution configurations — including whether and how any published call was traded. The boundary is deliberate and disclosed.
