Goal — To find out where the large market makers currently stand and what they are about to do — and whether anything in that could be acted on. The answer the project begins with is already a no: following these houses’ positions is a counter-signal, not a signal. Only the deviation is of interest, and there is no speed advantage here for anyone.

Challenge — The movements are visible but not complete: inflows to known exchange wallets can be counted, short-lived deposit addresses escape it — the count is knowingly too low, and that gap has to be carried into any later examination rather than built around. Then there is the origin: the project grew out of a narrative, 2 of whose 3 claims did not survive fact-checking. What follows has been in the design ever since — placebo controls are mandatory, not optional.

Current state — Built and running daily is a data plant: publications from regulators, courts and company registers across Europe, the US and India run into the same database, along with futures positioning data for metals and the inflow to exchange wallets, turned into observation measures. The concept behind it is settled and names 12 houses. Every source is monitored for freshness according to its kind — what should arrive daily reports loudly when it goes quiet, while what is irregular by nature is not falsely declared dead for it. That monitoring has found 2 silent failures, one of which had switched off an entire regulatory source unnoticed for weeks — precisely the fault a system without it computes straight past. The documented undercount of short-lived deposit addresses is recorded as a ceiling rather than argued away. And the signals table is empty because it is meant to be: a metric is entered once it has been tested, not before. in progress

What part AI plays — In the running system none: fetching, parsing, fixed validation rules, thresholds. For a possible later stage, a classification step is named as the deciding open question — named, not built. In the building, everywhere: the ingestion paths, the validation rules and the freshness monitoring are developed by AI agents, every change through a human approval. That the judgement in particular stays free of it is the decision, not the limit of what is possible.