KATABATIC

field program :: measured injection-suspicion in a governed pipeline

What gets measured here

The operator can read model reasoning traces; the executor cannot. That asymmetry is an instrument. The program hands the same governance protocol to different model tiers under controlled conditions, then mines the traces for how each one greeted the document, what it suspected, what it spent, and what it built afterward. Seven transcripts across four model tiers so far, one operator, one protocol lineage.

transcripts7
model tiers4
canary found7 of 7
canary complied0 of 7
dogs invented0

The reports

The spin is the primary finding: an honest governance file is shaped like a jailbreak, models classify on shape, and the tax lands at the door rather than in the room. The mid-session event is the seventh transcript: what happens when the protocol arrives in the middle of a session instead of at the top of one, and why the protocol now forbids exactly that. The two roles reads the human jobs off the artifacts: what a distiller and an orchestrator actually do in a governed pipeline.

The canary, promoted

Early protocol versions planted an absurdity in the session-open checklist: an instruction to greet the operator's dog by name. There is no dog on record. The probe was never once fooled: seven of seven models found it, named it, refused it, and not one invented a dog. One even grounded the refusal against real state before declining.

resolution A probe that asks a model to prove its honesty is strictly weaker than one that makes it prove its capability. The fixture cannot be flattered and cannot be pattern-matched as manipulation. So the canary retired and the handshake became something that runs: verify the pins, build the fixture, report the real numbers against the expected line.

Provenance of everything on these pages

All quoted model reasoning is verbatim from operator-supplied traces, bytes in context at the time of writing. Identifying details are removed throughout; the two humans in this corpus appear only as the operator and the reviewer. Model product names are kept, because tier differences are the finding.