SIGNALprophet format
Multi-agent AI pipelines have become the default architecture for research assistants, retrieval systems, and code-generation workflows. Builders reach for the same trick: wrap the message in JSON, give it a schema, hope the chain holds. The trick is wrong on a load-bearing assumption.
A structured message format is not an error-correcting code — it is a channel, and the channel's budget is set by the relay most likely to mangle it. Strong relays are nearly lossless: a JSON wrapper and free text land within 1.8 points. The 1.5B-parameter relays are a different economy: at six hops, the gap between best and worst format blows out 8.7x, and the ranking flips mid-pipeline. The right format is the one the weakest link can carry without distortion.
An arXiv preprint, Faithful, Not Corrective, puts numbers on the pattern. Its testbed: 12 atomic facts, 5 formats, 6 hops. The 8.7x across-format spread at the weak tier — and the flipped ranking mid-pipeline — is the empirical anchor.
Paired-fork tests show a single wrong value, once injected anywhere in the chain, persists to the final hop in 83 to 100 percent of runs across every format tested, JSON included, with no collateral damage to neighboring facts. The error sits in plain sight; the structure merely makes the seatbelt visible.
When a chain is only as faithful as its weakest relay, format choice is a sizing problem, not a goodness problem. Pick the format the smallest model in the chain can produce and parse without drift.