How the router picks a model.

Every LADLE turn goes through a routing decision. The router picks between Fast (Claude Haiku 4.5) and Deep (Claude Sonnet 4.6) based on a handful of signals, evaluated in a specific order.

The order of decisions

**1. Kill switch.** If we've flipped LADLE_PAUSE_SENDS=1 (a global pause we can activate in emergencies), every send is blocked with an honest error pointing at the status page. Rarely used.

**2. Access gating.** If your subscription is inactive (none / canceled / past-due-past-grace), the composer shows a "Set your table" card and no model runs.

**3. Rate limits.** Ten messages per minute per user. The 11th within the window gets a 429 with a retry-after time. Rare to hit; exists to catch accidental hot-loops.

**4. Budget check.** Cycle-level (monthly spend against your plan's API budget) and burst-level (spend in the last 5 hours). Over-budget on either → block with a designed error. Near-budget → the router force-downgrades Deep requests to Fast to preserve capacity (see step 6).

**5. Explicit mode.** If you set the composer mode to Fast, you get Haiku. If you set Deep, you get Sonnet (unless step 6 downgrades).

**6. Auto-routing (default).** No explicit choice: - Design-mode chats always route Deep (Sonnet) — HTML output quality matters. - Scoped-edit turns in the design canvas (element selected via inspector) route Fast (Haiku). Sub-second edits vs 30-90s full regenerations. - Attachments (files, PDFs) force Deep — vision + document reasoning isn't Haiku's strong suit. - Chats already established as Deep (based on prior turn's difficulty signals) stay Deep for continuity. - Everything else defaults to Deep. Sonnet is LADLE's default model; Haiku is opt-in via Fast or auto-triggered by scoped edits.

**7. Budget degrade override.** If the cycle spend approaches the Sonnet-affordable portion of your API budget, the router force-downgrades all requests to Haiku. You'll see the header "Deep model paused — Fast responses continue until [reset date]." This preserves capacity; users would rather have Haiku all week than Sonnet for two days followed by a hard block.

What you can override

The Fast/Deep composer toggle is respected unless budget forces a downgrade. Setting Fast when the router would have picked Deep is fine (you save budget); setting Deep when the router would have picked Fast is also fine (you get the better model).

What you can see

Every reply's kitchen panel (Cmd+K) shows the model that produced it AND the routing reason in one sentence. If you're curious why a specific reply came from Haiku, that panel has the answer.

What we don't do

We don't have a "smart" auto-picker that guesses whether YOUR SPECIFIC prompt is hard enough for Sonnet. Claude's own auto-routing is opinionated in that direction; we chose the simpler pattern (default Deep, budget downgrade to Fast) because it's easier to explain and easier to reason about.