Consult has one way to tell your desk; the shell lives in the gutters; every algorithm cites its source

What changed for people using Bellwether

• Consult has exactly one way to turn your words into Guidance. When you tell Consult something the desk should act on ("never buy NVDA", "stay in cash today", "lean into energy this week"), it drafts a single "Tell your desk" card carrying your words verbatim, and the same typed confirmation card you see under Strategy › Guidance is read back inline in the chat; once you confirm, the portfolio manager's acknowledgement comes back into the same thread. The older, narrower steering card is gone — there was real overlap and it was costing correctness (the operator asked: "is it bad to have the older, narrow cards? isn't there overlap? if yes, we should remove?"). Measured with the live model on the version that ships, on phrasings written after the wording froze as well as the ones it was tuned on: your words landed on exactly one card in 150 of 150 trials and were carried verbatim with nothing misrouted in 147 of 150 (44/45 tuned, 103/105 held out), no stray steering card in any trial, and one card in 150 over-claimed what the Executor enforces — caught by the grader and reported against a bar of zero. Said plainly: the first run after the removal scored 94 % and reached 50/50 only after two prompt passes on the same ten scenarios, which is why the held-out set exists. Plain nudges that do not map to a typed field are kept as dated notes in your own words instead of being dropped, and interpreting your words now runs in its own lane so it never waits behind a desk session; if the desk is down the card says so.
• The console shell feels alive without anything moving behind what you read. A subtle, slow ambient layer lives only in the gutters and margins of signed-in screens; every card, table and chart sits on a still surface; the backdrop's colour stays inside the same measured calm ceiling the Overview shipped with; dense working screens and anyone whose device asks for reduced motion get a completely still shell. (Operator: "the orange nasty glow is gone, which is great, but i do want some background/ambient effects … living and breathing.")
• Every one of the 14 registry algorithms now cites where it comes from. Each algorithm's spec page names the primary paper and the practitioner formulation it implements (public citations), states rule by rule whether our code MATCHES, is a JUSTIFIED VARIANT, or DIVERGES, and each declared divergence is pinned by a strict test that fails loudly if someone "fixes" the code without a versioned, ledgered change. Nothing about how any signal behaves changed in this release.

Known gaps / coming next

• Four divergences the provenance work surfaced (D1–D4, e.g. the canonical sector-rotation skip) are recorded as proposed versioned changes, not applied; they need evidence under the strategy-change rules before any behaviour moves.
• The PM's tools to think with, the janitor stream's fixes and the five alpha-hypothesis registry rows are building their review fixes and land next; the weekly PM tournament (pm-hillclimb) is running its first cycle.