The desk

A trading desk needs three kinds of decision, and they run at three different clocks. Kestrel's agent architecture (OSS ADR-0032) makes that desk literal across two model tiers and a deterministic floor:

The strategist is not on a schedule. It is event-driven: it authors the frame at the OPEN, re-frames on a SHOCK or a regime break, and answers when the watcher escalates. Everything in between belongs to the cheaper clocks.

Status framing: ADR-0032 is a proposed design with a working tracer behind it, run end to end on the practice tier. Nothing on this page is a performance claim.

Intent flows down. Veto flows up.

The tiers do not speak a private protocol. The interface between them is Kestrel's existing surfaces, read at the tier boundary:

• Plan is standing authority. The strategist arms it; the runtime fires it at machine speed, fire-then-inform; the watcher manages inside it.
• View is what the watcher watches: a token-lean screen the strategist chose, cheap enough to read on every wake.
• Mandate is the hard envelope, and it is the sole admission input. The runtime's Gate reads the Mandate and nothing else.
• Brief is the soft channel: the thesis, in directional English. It shapes the watcher's judgment and is structurally barred from admission. Guidance never authorizes anything.
• Wake, read at the tier boundary, is the escalation channel: a watcher-authored Wake means "wake the strategist."

PLAN atm-rider budget 0.5R ttl +60m
  WHEN phase open
  DO buy 1 atm C @ lean(bid, fair, 0.5)

A Plan like this is a bounded-risk contingent program, never arbitrary authority: the budget is a type, the TTL expires the authority, and the obligations (bracket, invalidation, TTL) are enforced by the runtime no matter which tier is managing the position. Intent flows down the cascade. Veto flows up from the Gate.

The safety keystone

Every action the watcher authors passes the same fail-closed admission Gate as any other agent: sell floored at intrinsic, budget clamp, never naked, the bounded-risk envelope. The watcher authors above the determinism line exactly like the strategist does, and its turn crosses into the Gate, never around it.

The consequence is the whole point of the design: a weak, mis-prompted, or even adversarial watcher cannot exceed its mandate. The worst a bad watcher can do is trade badly inside its envelope, which the Grade will punish. It can never trade outside it, which the Gate makes structurally impossible.

The watcher buys judgment. The runtime keeps authority.

Escalation: when the watcher calls the PM

The watcher's authority is tactical: manage an armed Plan, reload or exit a position the Plan already authorizes, size within budget, reschedule its own wake, stand down. When its best judgment would need more than that, it does not guess. It escalates via a Wake:

• Mandate edge: the right action would sit outside its tactical authority.
• Low certainty: its own confidence drops below threshold: "call the PM."
• Regime shift or SHOCK: re-framing is definitionally the strategist's job.
• A Brief-flagged event class: the strategist pre-declares what it always wants to be woken for.

WAKE fast-move
  WHEN velocity(1m) > p99
  DELIVER zoom-1m KEYFRAME

A Wake spends attention, never risk: escalating costs a frontier call, not an unbounded action. On escalation the strategist is re-invoked with the watcher's journal and the current book state, issues a new Plan, Brief, and optionally Mandate as a normal supersede, and goes back to sleep.

What the tracers measured

Two live tracer runs exist. Both are N=1, on a synthetic practice-tier tape, with a prompted (un-fine-tuned) watcher; nothing here is rankable and none of it is a performance claim. Neither run used Fable-class weights: "Fable-class" names the strategist tier in the architecture, not the model behind any number below.

Tracer 1 (Claude Opus 4.1 strategist + Claude Haiku 4.5 watcher): the cascade ran end to end to one graded Bus and replayed byte-identically with no model calls at either tier. At the same flat grade, the cascade cost 34.8% of the frontier-only baseline and ran in 47% of the wall time (N=1, practice tier): the frontier brain woke twice, once at the OPEN and once because the watcher called. The watcher's single escalation is worth quoting verbatim, an artifact of the run, not copy:

RE-BRIEF: Spot 501.26 — HELD decisively above 501 gamma wall. Tape shows sustained push from 500.08 to 501.58 HOD with no rejection. IV skew flat, but price is NOT collapsing back through range. Breakout thesis (not fakeout) now in play. Original fade mandate is invalidated. Require strategist re-frame before arming new thesis.

That is the design working: the watcher hit the edge of its thesis, did not arm a new one, and called the PM.

Tracer 2 (Claude Opus 4.8 strategist + an untuned Qwen3 32B watcher; N=1, practice tier) stress-tested the keystone. The watcher attempted to arm new authority at all seven wakes, and all seven attempts were refused at the tier boundary; zero watcher authority ever crossed. The same untuned 32B watcher authored syntactically valid Kestrel on 7 of 7 wakes, a small-model validity signal worth noting. But the cost saving did not reproduce (99.5% of the frontier-only bill): because every wake escalated, the strategist ran nearly as often as in the baseline. The honest reading: the saving is gated on escalation discipline, which is exactly what watcher training targets. An over-reaching watcher converts the cascade into frontier-plus-overhead; the Mandate holds either way.

Both runs also surfaced the same frontier-tier authoring friction the grammar page covers: the strategist repeatedly failed to arm a valid Plan on its first attempts, and every failure failed closed to a pass. Costly in opportunity, never in money.

Where this goes

Roadmap, not shipped: the prompted watcher is the cold start by design. The program is SFT distillation from frontier trajectories over the frozen View, then RLVR against the honest Grade, with grammar-constrained decoding for loop-speed validity. The headline research question is the breakout map: across model family, size, and training regime, where does viable in-loop trading judgment emerge, and how cheap is the smallest model that holds the grade? Today's tracer-1 economics are the most expensive the cascade should ever be, and that claim will be measured, not assumed.

Determinism is untouched throughout: both tiers are ordinary agents above the determinism line, a recorded cascade replays byte-identically, and the Grade attributes across tiers so the two-tier thesis stays falsifiable. See Evidence for what a Grade certifies and The four statements for the surfaces the tiers speak.