Measuring what the AI adds, and playbooks that load themselves

What changed for people using Bellwether

• The desk can now be measured against a rules-only twin of itself. A designated twin book runs the same universe, posture and limits with the mechanical rules engine instead of the AI portfolio manager; every week the difference between the two is computed and published, and withheld (with a note) for any week in which the twin did not actually trade or its settings drifted from yours. This is how "what does the AI add?" gets a number instead of an opinion.
• Role playbooks that switch themselves on. Each desk role now has situational instruction modules (for example a halt-and-gap procedure for the portfolio manager, a signal-health diagnostic for the quant) that code selects only when the situation calls for them; which ones are proven to change behaviour on the live model is recorded per module, and unproven ones cannot be wired in. (This release ships the modules, the selector and the proof machinery; roles start reading them in the next batch.)