Quieter drift pages, cheaper desk sessions, steadier live prices

What changed for people using Bellwether

• Fewer false "positions differ from the broker" pages. When the desk's own order is filling, the books and the broker disagree for a heartbeat by exactly the shares that just filled; that no longer rings your phone. A page now arrives only when nothing of ours explains the difference — and when open orders exist but do not account for it, the message names them and says why (wrong side, too small, only a resting protective stop, quiet for too long) instead of "no order explains it". A broker-side protective stop that sells a position placed on an earlier day now shows up in the books the same day instead of at the close.
• Desk sessions cost less for the same decisions. The part of the desk's instructions that is identical for every book is now sent once and reused across books and across the day's check-ins instead of being paid for again each time; once-a-day sessions keep the cheaper short-lived cache. Nothing about what the desk reads, decides or proposes changed, and no model or thinking budget was lowered.
• Admin › Usage says where the money goes. A new Efficiency section lists each kind of session (research brief, premarket plan, intraday check, Critic review, Consult, the shared market brief) with its count, spend, cost per session, steps per session and how much of its prompt was reused from cache, plus one sentence on how many check-ins were skipped on purpose and why.
• The go-live scorecard judges AI cost sensibly. Cost per filled order punished a desk that correctly sat out; the scorecard now reads cost per decision (proposals, typed decisions not to trade, automatic exits) against a ceiling, and running cost as a share of the book a year against an aim and a ceiling — each "Not measured" until there are enough decisions or end-of-day books to judge, never red on five fills.
• Live prices stay on the stream. The trading engine's five-minute "would a starter buy pass right now" dry run no longer asks the data vendor for bid/ask on every watchlist name of every book; it prices only what the book can act on and lists the rest as not priced. That keeps the vendor's request budget and the trading loop's time for the names you hold, and the "policy gap" alarm now means what it says.

What you may notice

• Overview's status line can read "last looked 10:34 CT, checked again 10:45 CT: nothing changed (prices, fills, news, halts, settings), so that look was skipped, next 11:00 CT" once the intraday change gate is switched on for a book; it ships off.
• The go-live page's rows are worded as sentences with a status word (Green, Amber · passing, Red, Not measured, For information) and figures like "$0.61 each; the ceiling is $0.80" or "2.1% a year · aim under 1.5%"; the thresholds list reads in words rather than setting names.
• Admin › Usage's per-book table is headed "By book"; token counts say their units once under the table; an empty cell reads "n/a".
• A "positions differ from the broker" page that does arrive may add "(still differing, reported earlier: XOM)" rather than repeating numbers it already sent.
• The help article on paper versus live money describes the scorecard's four states and the two AI-cost rows; its screenshot was re-shot.

Known gaps / coming next

• The daily AI budget that scales with the number of books, the money set aside for end-of-day checks, and the Admin panel / Overview line that say a desk is paused for budget re-merge onto this release and follow.
• Three cost levers ship built but off, each with a dated promote-or-kill test: one shared market brief per day read by every book, skipping an intraday check-in when nothing the desk tracks has moved, and reading several names' signals in one call. The order of the desk's instructions (book mandate after the shared part) is itself on a one-week verdict.
• After the release, the interim setting that widened the live-price stream to the whole focus list is removed once a probe at the open confirms nothing polls those names any more; until then the stream carries a few more symbols than it needs.
• The first trading day after the release is checked for "drift settling" log lines and zero drift pages naming our own orders before the escape is marked fixed.