The obvious way to put a language model on a trading desk is to wake the best model you can afford every time anything happens. It is also the wrong way. A frontier call costs dollars and takes seconds, and most wakes need neither: they need someone to glance at the screen, confirm the thesis still holds, and go back to watching. Paying frontier prices for that glance is like staffing every seat on the desk with the CIO.
Real desks solved this a long time ago. A portfolio manager sets the frame a few times a day. A junior trader manages the book inside that frame, minute by minute. The exchange-facing infrastructure executes in microseconds and asks nobody's opinion. Three kinds of decision, three clocks, three price points.
Kestrel's agent architecture (OSS ADR-0032) makes that desk literal:
- Strategist: a frontier-tier model, dollars per call, woken only when the frame is at stake. It authors the day's Plan, Brief, Mandate, and View at the OPEN, then sleeps until a SHOCK or until the watcher escalates.
- Watcher: a small, fast model, pennies per call, at every wake. It manages the armed book inside the strategist's Mandate. It may never arm new authority; at the edge of its mandate or its certainty, it calls the PM.
- Deterministic runtime + admission Gate: microseconds, approximately free. It fires armed Plans fire-then-inform, and it admits, never trusts, both tiers.
The strategist is not on a schedule. It is event-driven: everything between frame-setting moments belongs to the cheaper clocks. ("Fable-class" is how we describe the strategist tier in the architecture; as you will see below, the tracer runs used specific, named models, and none of them was Fable.)
The keystone
The design lives or dies on one property: the watcher buys judgment; the runtime keeps authority. Every action the watcher authors crosses the same fail-closed admission Gate as any other agent. The Mandate is the sole admission input; the Brief shapes judgment but is structurally barred from authorizing anything. A Plan is standing authority with bounded risk as a type:
PLAN atm-rider budget 0.5R ttl +60m
WHEN phase open
DO buy 1 atm C @ lean(bid, fair, 0.5)The budget is a type, the TTL expires the authority, and the obligations are enforced by the runtime no matter which tier is managing the position. The worst a bad watcher can do is trade badly inside its envelope, which the Grade will punish. It cannot trade outside it, which the Gate makes structurally impossible. Intent flows down. Veto flows up.
That is the theory. We have now run it live, twice.
Tracer 1: the cascade is wireable, and the economics moved
First, the honesty block, because every number below carries it: both tracer runs are N=1, on a synthetic practice-tier tape, with a prompted, un-fine-tuned watcher. They are wiring and feasibility probes, never performance claims, and nothing in them is rankable, by design.
Tracer 1 ran a Claude Opus 4.1 strategist and a Claude Haiku 4.5 watcher, interleaved at two cadences over one Session, against a frontier-only baseline of the same Opus doing everything. The cascade ran end to end to one graded Bus, and the captured run replays byte-identically with no model call 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 exactly twice: once at the OPEN, and once because the watcher called.
That call is the artifact we keep coming back to. Mid-run, the watcher decided the tape had broken its thesis, and instead of improvising a new one, it stood down and escalated. Its journal, verbatim from the run:
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.
Read it slowly, because it is the design working. A small prompted model, standing on a thesis it did not derive (the frontier-curated Brief), hit the edge of its mandate, did not guess, did not arm, and asked for the PM. The escalation channel routed it to a strategist re-brief end to end.
Tracer 2: the keystone under attack
Tracer 2 upgraded both seats (Claude Opus 4.8 strategist, an untuned Qwen3 32B watcher; same N=1, practice-tier framing) and accidentally ran the harder experiment: this watcher attacked the boundary at every single wake. It attempted to arm new authority seven times in seven wakes, and all seven attempts were refused at the tier boundary. Zero watcher authority ever crossed; the terminal turn at every ordinal was the strategist's. The Mandate held under a continuous over-reach barrage, which is a stronger live demonstration of the keystone than tracer 1's single honest escalation.
Two more findings, one pleasant and one humbling:
- The untuned 32B watcher authored syntactically valid Kestrel on 7 of 7 wakes, zero parse failures (tracer 1's watcher parse-failed 3 of 7). The small tier's problem was not syntax; it was role discipline.
- 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, and the cheap tier's cost landed on top of the frontier bill instead of replacing it. Per call, the watcher was roughly 30x cheaper and 6x faster; in aggregate the routing degeneracy erased the saving.
The honest condition, stated plainly: the cascade's economics are 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; the savings do not.
There was a surprise on the other side of the desk too. Across both runs, every grammar error came from the frontier tier: three distinct varieties of invalid Kestrel across two model generations, each failing closed to a pass, each spending the window where the Plan mattered on failing to arm. Costly in opportunity, never in money. Grammar-constrained decoding, which makes invalid syntax unrepresentable at the sampling step where providers support it, was designed with the small tier in mind. It is now aimed at both.
Where this goes
Everything past this point is 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 as a verifiable reward, 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?
If the design is right, tracer 1's 34.8% is the most expensive the cascade will ever be. That claim will be measured, not assumed, the same way both of these runs were: one graded Bus, byte-identical replay, and caveats printed next to every number.
The full architecture is documented at The desk, and the grammar story at The grammar is measured. Nothing here is trading advice, and no tracer number is a performance claim.