The naive router breaks at the moment a model name crosses from prose into a billing call. That is why the pipeline puts server-side capability and cost validation after the LLM proposal instead of trusting the proposal [1]. Everything upstream exists to shrink the set of names the planner is allowed to produce, and the stated rule is that the planner must not rely on training data to remember model constraints [12].
Classification is the first cut. `modelKindFor` takes media type, reference count and mode, then returns `video` for any video request, `image-fusion` when the mode is fusion or the reference count is two or more, `image-i2i` when the mode is edit or there is exactly one reference, and `image-t2i` otherwise [5]. Reference count is doing real work there: one reference means a local edit or a reinterpretation, several imply fusion with consistency constraints [6]. No model has been named yet.
The media-type classifier is where I would spend review time. It tests image-to-video patterns first, then image signals, then video signals, and returns null when nothing matches [9]. Precedence is the mechanism. "Create a cinematic 3x3 photo grid" and "Animate this photo into a short cinematic video" both carry image and cinematic language, and only the second is a video job [10]. Reverse the order and you charge someone for a video they did not ask for. The author reserves the LLM classifier for genuinely ambiguous briefs and records whether each decision came from rules, llm or fallback so misroutes can be traced [11].
Then the catalog. `ModelSpec` carries kinds, minimumTier, baseCredits, resolutions, aspectRatios and formats, and `catalogForTier` hands the planner only the models whose kinds include the task and whose minimumTier ranks at or below the caller's tier [13][14]. Four generation kinds against three tiers gives twelve filtered views that function has to get right, which is a small enough matrix to enumerate as fixtures [19]. The claim for doing it this way is that it is cheaper and safer than passing a long list of unusable models [15]. The cheap part is tokens. The safe part is that a model absent from the prompt cannot come back in the answer.
The client-side detail matters more than its two sentences suggest: browser components need capability metadata but must not import provider SDKs, so the design keeps a client-safe projection with no server clients or environment-dependent imports, generated from the authoritative registry and tested for drift [16]. Drift there presents as a picker offering a model the server will refuse.
This transfers on one condition, which the piece states plainly. The product spans text-to-image, image editing, multi-reference fusion and video, each with different aspect ratios, durations, pricing rules and access tiers [2]. With one model and one tier, the catalog and the validator are overhead you maintain for nothing. The evidence is also one practitioner's account with no error rates or misroute counts, and the excerpt breaks off mid-list before the JSON repair strategy [20]. The failure list is still the right test suite: truncated JSON, unknown or deprecated model IDs, unsupported resolutions, a model from the wrong tier, and prose that disagrees with the structured values it sits beside [18].
