Stock Analyst

A shared watchlist and plain-English research. Every night a Claude worker on my subscription — never the metered API — writes the analysis, stores it permanently, sends a Telegram notice, and later grades its own past calls against real prices. It is advisory only: no code path can place a trade.

The point was never stock picking. The point is an honest record: every forward-looking number comes from market data and transparent math, the model is forbidden from inventing a price, and every verdict is written before the outcome exists so it can never be backfilled.

On August 31, 2026 the platform found a look-ahead bug in its own backtest: historical market caps had been computed from split-adjusted prices and as-filed share counts, so every one was wrong by every split since the filing, in a direction correlated with the company's future. All 28 signals that had graded as working were artifacts of it. The bug was fixed, the verdicts were retracted, and the repaired panel voids itself — on a universe that prices only survivors, no edge and cannot see are indistinguishable.

So the instrument was rebuilt around that: controls carry pre-registered expectations and can never silently grade as working again, and the evidence that counts is the forward record — a paper options wheel, an earnings-reaction log, a decision log — all written ahead of the outcome.

The rule from the start was that AI writes the narrative and data writes the numbers. The retraction is what that rule looks like when it is enforced against the builder's own results: the write-up of the bug is the most-read document in the repo, and it is the first thing a new reader is pointed to before believing a number on the site.

It is not public yet, and won't be until the forward record is long enough to mean something.