Telemetry that fits inside an ε error tube — and proves it does.

What it guarantees

We keep the series with a per-sample bound: |value − reconstructed| ≤ ε, always. The verifier re-decodes and checks — the bound is never violated, and that is a proof, not a promise.

• ▸Certified retention: months of raw data at ±ε for the cost of a downsample — with the tail preserved.
• ▸In-place queries: questions answered touching 0.1–3% of the file, without decompressing (see CLAMP).
• ▸Free detection: when the event arrives the compressor emits more — surprise is a by-product detector.

Where ε lives

Seismic (digitizer counts), subsea DAS (strain noise), emissions (40 CFR Part 75), glucose (ISO 15197), market price (tick). In all of them ε already exists in a document — we just turn it into a guarantee.

Measured on ~15 real sources: where data is integer/impulsive TUBE wins the bytes; where it is smooth it ties or loses to SZ3 — but in all, the bound is provable and the function (query + seal) is the product.

The honest limit

Surprise catches strong, fast transients; it is weak on slow drift and blind to sub-noise signals (raw gravitational-wave strain needs matched filtering). Scope is a feature: we say where it works and where it doesn't.