Small and specialist models

Narrow the world and a smaller model becomes competitive. The vertical case is made most strongly for domain-tuned models in the ~20B class trained on proprietary data: comparable quality on that vertical, materially lower inference cost. The distillation case runs the other way round — a frontier model generates and verifies traces at scale, and small open models learn from the corpus. Both are bets that the generalist's breadth is being paid for and not used.

The reported floor for complex tool use sits around 5-10B parameters, with dense architectures preferred over sparse ones at small size. More interesting is which capability fails first. Small models pass smoke tests and do most of the work on feature tasks; what they consistently cannot do is close the observe → diagnose → repair loop on their own output. The named deficit is self-verification. That places the failure precisely where a review or verification role would sit.

A weaker executor becomes reliable when the layer above reformulates the goal into steps that fall inside the executor's reliable distribution — "move left until directly over the hammer" rather than "put the hammer on the plate". The leverage sits in the planner's phrasing, not the executor's weights. Underneath it all sits a framing argument worth more than either result: expertise — contractive, situated competence in one micro-world — is largely orthogonal to raw expansive intelligence, so narrowing the world is a real source of capability rather than a consolation prize.

Also: pi-0.7, a vision-language-action model with a lightweight world model; Hermes Agent, Kimi K2.5 and GLM behind the open trace-dataset push.

Assay is context-first by construction: briefs and skills are the 90% layer, and fine-tuning has already been declined as an investment. This evidence does not overturn that — it supplies the trigger condition. Specializing a model is worth considering only at sustained scale in a narrow vertical, which no current Assay workload reaches.

Local 5-10B models look economically tempting for desk roles, and verification is exactly the role small models are reported to fail at. So the honest position is: verification stays on a stronger model, or it becomes mechanical — tests, lints, guards. The second option is the one Assay already prefers, since invariants are closed-form checks rather than model judgment. That makes "run the verifier on a cheap model" a bad trade twice over.

The planner/executor result restates the author-worker gap in mechanical form, and here Assay is shipped rather than designing: brief authoring exists precisely to invest in the reformulation layer rather than in worker capability, and the brief schema is that reformulation written down. Two carries remain unexploited — review comments could coach the reasoning rather than patch the artifact, and "thinking traces need not be textual" suggests verification evidence should use whatever modality fits the claim. The intelligence-versus-expertise frame is the best external justification for the methodology yet sighted: skills are compressed micro-world structure, findings and registers are the reusable structures, and a brief transfers situated competence to a worker with raw capability but no history. The gap it exposes is that this compression is manual today.

• Whether any of the "20B beats frontier on the vertical" claims arrives with a published head-to-head on a named benchmark; today it is a panel assertion plus one white-label vendor's customer count.
• Whether the self-verification floor moves — a small model that reliably closes observe → diagnose → repair would change the desk-role economics immediately.
• Whether the distilled-trace pipeline (frontier model → minimal harness → open trace corpus → small agent) produces a model anyone reports running as a load-bearing production agent rather than a hobby home agent.