Each page below is a full record of one pipeline run: the metric contract the model designed, the artifact it generated, what the artifact actually produced when executed against a live target, and the per-stage token cost. Correctness in every run was decided by scraping the artifact and diffing it against the contract — never by a model.
RabbitMQ · REST docs → Python exporter
The full pipeline on a clean run: docs normalized by code, metric surface and IR designed on the standard tier, exporter code generated on the strong tier, then verified by execution. Shows where each token was spent.
PASS 28/28 metrics · 0 repair iteration(s) · 3 model call(s) · 22.0K in / 11.4K out
RabbitMQ · recovering from an upstream API change
The upstream API was changed underneath a working integration (a renamed field and a newly-nested one). Validation caught it, and repair fixed the exporter in one iteration using fresh endpoint samples as ground truth.
PASS 28/28 metrics · 1 repair iteration(s) · 1 model call(s) · 13.7K in / 6.0K out
SNMP device · IF-MIB → snmp_exporter config
One model call total: the MIB is parsed deterministically, the model only chooses what to monitor and how to name it, and snmp.yml plus the validation contract are compiled from that same selection so they cannot drift.
PASS 18/18 metrics · 0 repair iteration(s) · 1 model call(s) · 8.9K in / 3.3K out
Legacy switch · SNMP selection self-repair
A device with no ifXTable, so the 64-bit counters the model preferred do not exist. Repair walked the device, saw which OIDs actually respond, and substituted the 32-bit equivalents rather than dropping throughput coverage.
PASS 9/9 metrics · 1 repair iteration(s) · 2 model call(s) · 21.2K in / 3.3K out