High-Fidelity Time-Series Retention

Telemetry is stored as fully time-resolved series, not downsampled summaries, so the fast physics survives into the archive intact.

Keep the fast structure

A summary statistic — a per-shot average or peak — throws away exactly the fast structure the machines are controlled on. High-fidelity retention stores the full time series at acquisition resolution for the diagnostics where it matters, so a Mirnov trace archived today still shows the mode rotation, and a strain trace still shows the transient.

What full fidelity buys

• Re-derivation of any feature from raw series with improved definitions.
• Forensic reconstruction of disruptions and quenches at native resolution.
• Training sets that expose models to real fast dynamics, not smoothed averages.
• Verification that a live feature matches its offline re-computation (parity).

Storage strategy

Full-resolution retention across 60+ channels is expensive in bytes, so the fabric applies lossless compression and columnar time-series layout tuned for range queries. Where a channel's information content genuinely supports downsampling — a slow thermal channel — a lower rate is stored, but the decision is explicit and recorded, never a silent default.

Serving the series

Retained series are served to offline batch analysis on L0 for retraining and multi-physics comparison, and windows are surfaced to the twin and copilots for context. The retention layer feeds the feature store so a feature computed live and one computed from the archive are provably identical. Retention policy is versioned per machine and applies equally to breeder and burner.