v1.9.12026-09-08 · current
Added
- The analyst, in every view. A spark in each panel header opens the assistant already scoped to that surface, with 3–5 actions written for it: which link is real and does it survive a wider window on the correlation matrix; when has it failed and is the sample big enough to mean anything on the event study; which of these is noise on the armed conditions. Every reply ends with a line stating what it just did.
- The standing desk. State one instruction in words — chosen from four, each printing what it costs — and the terminal then checks on its own clock, acts, and logs what it did, why, and the number that caused it. On the checks where nothing crossed a threshold it holds, it writes that down too. Pause, check now, undo the last action, change the instruction and stand down are one click each; a reading is held quiet for 5 checks after being raised so one swinging series cannot fill the log.
- Zero-input onboarding.
START <GO>asks "What are you here to do?" with five concrete answers. Picking one sets the instrument, range, overlays, correlation window, regime axis and standing instruction — then explains what it chose, why, and what the choice costs. - "Is the sample big enough to mean anything" on the event study: reports the scatter of every instance and the interval around the median (median ± 1.96 standard errors), and leads with the fact when that interval contains zero. On XAUUSD it does for four of the five shipped event classes.
- Failure list on every event study — "Where it did not work · 26 of 61" — with each miss dated, clickable and drawn in red on the fan.
- Armed conditions are now backtested across all 2,520 cached sessions the moment you arm one, reporting crossings, the last crossing date and the rate per year.
- Recurring briefing with a what changed since the last one section computed against the exact readings the previous brief was built from.
Fixed
- Overlays whose correlation sits inside the noise floor (|r| < 0.15) are now removed from the price axis with the number named, instead of being drawn as a line readers inferred meaning from.
- Regime buckets under 60 sessions are dimmed and flagged as thin rather than quoted alongside full-sample statistics.
Changed
- Arming and watching a condition are included in the licence and run in the terminal; the feed is now required only for delivery of a firing off the tab (Telegram, e-mail, webhook). The manual and the pricing table were corrected to say so.
- A standing instruction no longer survives a reload. Nothing was running while the tab was shut, and the log is meant to be trusted.
- The assistant refuses sizing and return questions in one line and names what it can do instead, rather than deflecting.