Routing Structured Outputs During Generation

A language model generation already contains information that most application pipelines discard: token probabilities, partial structure, confidence changes, segment boundaries, and the moment at which a typed object becomes complete. Conventional systems wait for the entire response, then invoke a parser, classifier, second model, or postprocessor to rediscover structure that was visible during emission.

Route M-Flight is a streaming architecture that routes, validates, records, and optionally entropy-codes typed segments during one generation flight. A model emits into declared channels—claims, actions, deltas, citations, explanations, tool arguments, receipts, or fallback prose. Segment validators operate at cutlines. Completed segments can be sealed with exact context, model/runtime identity, probability quantization, output digest, and raw fallback.

The immediate product value is not compression. It is lower duplication, earlier useful structure, bounded repair, and one-pass population of several surfaces. The compression extension uses the emitting model's own next-token probabilities to code exact output without a second inference pass, but bit-exact decoder conformance remains a serious research constraint.

Current evidence state

SPECIFIED

This label describes the strongest public-safe state of the specific paper claims. Open the paper's claim ledger for claim-level distinctions and limitations.

Publication

• Full paper: /papers/p14-route-m-flight
• Claim ledger: content/papers/p14-route-m-flight/claim-ledger.json
• Source manifest: content/papers/p14-route-m-flight/source-manifest.json
• Diagrams: 3
• Approximate body length: 1,407 words

Demonstrations

• d10-route-m-flight-stream-router

Benchmarks

• b02-fold2-codec-laboratory
• b05-surfacedelta-uplift

Public boundary

The research entry publishes a public-safe architecture and evidence record. It does not grant implementation rights or expose restricted claim-ready material.