Damiano Bianda

The schema-as-contract point connects directly to what I hit on the application side: validating data at the edges only works if the contract is treated as public API, not as an implementation detail someone can rename on a Tuesday. The idempotency section is the one I'd underline twice. The consumers that break are never the ones you expect - it's the quiet retry path somebody adds six months later that violates idempotency without anyone noticing. Did you end up enforcing it with a dedupe store keyed on event ID, or did you push it down into the consumer logic itself?

The two-readers framing is the part people underrate. I ran into a small version of this from the other direction: I built a local tool that explains official letters, and the design rule ended up being that the model only ever sees the fields the code extracted, never a number it could invent. Your tests are the same gap in reverse - what the agent "read" is whatever the converter decided to keep. Did any of the converters you tried handle display:none differently from aria-hidden? That seems like the first place a checker would disagree with a human reviewer.

Great teardown. The size gap you documented is real on the ground: when I needed a model for a fully offline tool (explaining official letters, the kind of text that cannot leave the machine), nothing in the newest families was small enough for plain CPU, so I ended up on a 1B Gemma. Slower than your L4 numbers, about 40 seconds per document end to end, but the task budget allowed it. Your memory math is the part most "run it locally" tutorials skip: 18 GB before context is not a laptop model for most people. Did you measure how much extra memory the context window adds as you raise it past 4,096?