Tier the models; the validation boundary is the thing you are actually buying

Two different things are being called validation here, and only one of them costs a model call. The API-layer kind is a shape guarantee: `messages.parse()` with a Pydantic output format hands back a validated object instead of a string you hope survives `json.loads()` ^[6]. That removes the code-fence class of bug for free. It says nothing about whether the number inside the object is true.

The truth-shaped kind lives at L2, which reconciles a record against other sources and checks it for consistency ^[3]. That placement is the design. The layer that catches a cheap model's confident error needs more headroom than the cheap model, so the validation boundary sets a floor under mid-tier spend: you can route volume down to Haiku precisely because something above it is paid to disagree. Make L2 the same model as L1 and you have not economised on validation, you have deleted it, and the symptom is a run that gets cheaper while the brief gets worse ^[5].

L3 is not a backstop. It reasons across whatever L1 and L2 handed up and writes the recommendations ^[4]. Anything false that clears L2 reaches the most expensive model in the stack and leaves it phrased as advice, which is the propagation the author opens with: one agent's hallucination becomes the next agent's input ^[1].

The cost side is worth checking against the code rather than the prose. Ninety percent off the shared prefix ^[15] holds only while the prefix keeps matching, and the stated failure mode is a volatile value such as a timestamp or unsorted JSON breaking it ^[11]. The guard offered is `assert response.usage.cache_read_input_tokens >= 0` ^[11]. A token count is never negative, so that line passes on every call, including the zero-hit calls it exists to catch ^[16]. The comment describes the check; the code does not implement it. What you want is that count logged per call with an alert when it drops to zero on a prefix that should be warm, because the same context repeats across hundreds of L1 and L2 calls in a single run ^[10] and a silent miss surfaces only on the invoice.

On provenance: this is one operator describing his own stack, which he says runs his company ^[12], with no error rates or per-run costs given. The model identifiers are constants in his file ^[13], so treat the version strings as local and the tier boundaries as the portable part. The retry setting is his choice too rather than a default, since the SDK already backs off on connection errors, 429s and 5xx, and he raises `max_retries` to 5 so a transient blip does not kill a long run ^[8].