What’s new in Bellwether

What changed for people using Bellwether

• The PM trades to its posture. Each posture now carries an explicit Mandate in the PM's brief: expected pace (entries per week), churn tolerance, and a "large gap, no trigger" rule so an under-invested Aggressive/Moderate book takes starter positions instead of waiting indefinitely; the number of orders a session may place is derived from posture (Preservation 2 · Conservative 3 · Moderate 3 · Aggressive 5 · Speculative 8) instead of one number for everyone — verified in live evaluation runs on both models.
• Quant suggestions you keep are acknowledged, not silently edited. Desk suggestion cards get a Keep button; auto-applied tuning shows as such and self-reverts after 7 days if it did not help (bounded auto-applier; the strategy history stays append-only).
• Consult → PM requests are fenced. Anything Consult passes to the PM (symbols, question, wake reason) is validated and quoted inside a fenced block the PM treats as a request, never as instructions; Consult can no longer confirm order-producing cards on its own — an owner or operator must.
• One daily AI budget per book shared by PM, Quant and Consult sessions with plain "budget reached" wording; tool handlers never raise into the SDK.
• Citations mean something: research claims the PM cites must resolve to a brief it actually read (citation-fidelity hold).
• Lab backtests blend signals with the same scoring the live desk uses (5 bps default costs, 520-day lookback), so Lab and live agree.

What changed for people using Bellwether

• New landing page and waitlist. The public page is rebuilt (status line, "An AI-native trading desk. You set the line.", product shots, evidence section); "Join the waitlist" now opens as a centred dialog sized to its content (a bottom sheet on phones) with one plain lead sentence.
• Role characters v2 across landing, Your desk, role headers and help (illustrated set replaces the code-drawn marks; a further illustrated set follows).
• Bellwether's mind v3 on Overview: structured watch/trigger rows with live prices and a status line that says what the desk is doing and what it is waiting for; phone layout reworked.
• Say no to a refusal, once. When the desk refuses a trade you asked for, the inbox row now offers a one-time override with a typed reason (never past a hard risk limit; rate-limited; audited).
• Trade ticket and position requests: ask the desk to open/close/resize a position from the symbol page; the request becomes a typed card the PM must answer (do it, decline with reason, or ask back).
• Errors read like sentences. Actions share one error path: no raw reason codes in messages; the Off Trader card no longer renders when the Trader is disabled.
• Accessibility: loading placeholders announce themselves properly (role=status) across 16 screens.
• Invite flow, OTP layout on ultrawide, tour ending on Overview and the static app backdrop are NOT in this release; they arrive in the next one.

What changed for you

• Stops keep being watched through hiccups. If the broker connection breaks, the desk now retries on its own after a minute, tells you loudly while protection cannot run ("guards unwatched"), and re-arms stops on positions that briefly vanished from a broker read instead of forgetting them.
• A vendor outage no longer cancels a session. When Alpaca's snapshot endpoint fails (as it did on 09-11), planning falls back to the live stream / last stored prices and says so, instead of cancelling the desk's session.
• Tight stops are widened, not refused. When the PM's stop is closer than the market's normal noise, the order is placed with the stop widened to 1.5 × the name's typical daily range at the same dollars-at-risk (smaller size), and the card says so. The separate "Trader agent" stays off: a live head-to-head showed it added nothing over this rule.
• Sign-in and safety fixes from the independent code review: step-up required where it was missing, cleaner lockout, idempotent retries across instances, alerts that fold instead of paging every tick, nightly report no longer fails on a read-only role.
• Quant evals ran live and passed; the Quant's proposals now carry their evidence.

What you may notice

• New alert kinds in the inbox if something goes wrong: "guards unwatched", "quote stale for guard", clearer scan-failure text.
• Deploys: a release can no longer roll back onto code that cannot run the current database (the 09-11 incident); the system prefers "forward-only" and says so.

Switched on in this release

No new switches; behaviour changes are in code. Consult's per-turn budget was raised separately ($0.75 → $2.00) after a first-question failure on a new book.

Known gaps / coming next

Console batches (new landing page, illustrated desk, mind card v3, trade ticket + refusal overrides, admin/onboarding polish incl. the invite default, waitlist modal and ultrawide sign-up fix) land next; error messages get their own surface; releases move to a one-command job inside GCP.

What changed for you

• Loss halts and order caps now scale with the size of your book. A $100k Aggressive book halts for the day at −$3,000 (3 % of equity) instead of a flat −$300; Moderate $100k books at −$1,500; small books are unchanged. Every book got an inbox card stating its before/after numbers. You can still set any value yourself in Risk — the allowed range is now wide enough ($10 – $100,000 a day, up to $1M an order) that the number is yours, not ours.
• Sizing knows about liquidity. On thinly traded names the desk caps an order to a small share of the day's volume and passes when the trading cost would eat the expected edge; on liquid names nothing changes. Each algorithm's card now shows the book size at which it stops scaling.
• Markets page. Indices, the market regime, movers, a scanner across the whole tradable universe (top gainers, unusual volume, breakouts …), upcoming earnings, and what your desk is watching — live during market hours.
• One price everywhere. Positions, the mind card, the symbol page and the trade ticket read the same live quote, labelled with its age.
• Faster, broader news. Headlines now stream in as they are published instead of being polled every ten minutes, and cover names beyond your watchlist.
• Trading halts are respected. When an exchange halts a symbol the desk freezes its stop logic for that name instead of chasing a price that is not trading, and the PM is told the name is halted.

What you may notice

• An inbox card per book titled "Platform update: the loss halt now scales with the account".
• The Risk page shows wider allowed ranges; values you had set yourself were kept.
• Sizing lines in proposals may mention a liquidity or cost clause on small-cap names.

Switched on in this release

Halt feed, minute bars for held/proposed/armed/core names, streamed news, nightly corporate-action re-adjustment, OPRA options context for research, liquidity caps (impact scale 0.7). Macro (FRED) data stays off until a key is added.

Known gaps / coming next

New books still need their first session to build a plan; signal blend v2 remains in shadow pending the point-in-time evidence rerun; the console batches with the new landing page, characters, mind card and operator controls land next.