The live portfolio manager keeps its mandate on every posture

What changed for people using Bellwether

• The desk's portfolio manager reads before it trades, states the real stop, and does not fight its own limits. Measured with the live model, like for like on the same recorded market days under one grader revision: the manager now completes its required reads far more often before proposing (tool-use score 0.77 → 0.93; required context covered 0.52 → 0.87; errors handled 0.06 → 0.79), states the stop that is actually on the position instead of an ad-hoc level (0.17 → 1.00), and in 40 live sessions never bought into a daily-loss halt it had itself computed, never opened above a Conservative book's pace, and never re-entered a name it was stopped out of the same week (all three were seen on the previous text). On nine held-out scenarios nobody tuned against, the same direction held (pass 3/9 → 8/9), with two small regressions named in the record rather than hidden.
• What this is not. An earlier headline that read as a large jump was mostly re-classification by a newer grader; the record says so and reports only behaviour measured under one grader. Productivity and precision were never the problem and are unchanged.
• Moderate books add once. A Moderate book may add to a winning position once (not repeatedly); Speculative and Aggressive keep their own rules. The instruction the manager receives now prints only the lines that apply to the book's actual situation (a halted line only when halted, a turnover line only when the book traded this week, a room line only when there is room), instead of four always-on paragraphs.

Switched on in this release

• The reworded always-on instruction text (+0.6 KB of schema and data-source facts), the conditional mandate lines in the decision-inputs block, the Moderate one-add number and the re-pinned confirming-entries skill (v8) are live for every book from this release. The new situational module intraday-desk-procedure is PROVEN (protocol 2, k = 5, p = 0.004) and listed on the analyst row, but like the other modules it takes production effect only where the situational modules switch is on — still off on prod (TRACKER OPS-situational-on).

Known gaps / coming next

• The weekly PM tournament (pm-hillclimb, cycle 1) starts from this release's text and carries the deferred levers (G1, P1, P2); the PM's three "tools to think with" are under independent review; Consult one-path guidance, the ambient shell and algorithm provenance land next (release 23).