zicato-read-telemetry

To trace multi-agent evaluation runs, you correlate canonical events.jsonl telemetry with reduced loss.json profiles and harmonograf sessions. This allows you to track specific epochs, generations, and board entries without relying on stale derived indexes.

Diagnosing drift and loss patterns involves examining reduced loss profiles to interpret drift counts, judge attribution, plan revisions, pass/fail results, metrics, runtime data, and abort conditions recorded during the evaluation run.

You can distinguish meta-loop activity from per-board-run sessions by correlating canonical run telemetry with structured invocation logs and harmonograf sessions, using recorded session identifiers to separate the overarching loop from individual board runs.

You create harmonograf deep links by extracting the recorded session identifier from the canonical telemetry files and using it to construct a direct URL to the corresponding live dashboard session for the specific evaluation run.

No, you should avoid starting live evolution runs merely to generate telemetry. Instead, you must read existing zicato workspace artifacts, interpret telemetry dialects, and inspect structured logs to analyze past runs across epochs and generations.