SIGNALAF

Signature Drift

Signature Drift — the tune meter

Every operator has a signature: the characteristic shape of their token cascade — the proportions between output, cache-write, and cache-read, anchored to input. Signature drift measures how far a stretch of work has moved from that shape. Zero drift = locked in tune; rising drift = the cascade is desyncing from the operator's own calibrated peak. It's a measure of shape, not magnitude — the same log family as 10xDEV.

Shape, not size

Drift is computed in log-space on purpose. Working twice as hard across the board (every axis doubled) is still in tune — the shape is unchanged — so it reads as zero drift. Going off on a single axis (lots of cache-write, no reuse) breaks the shape and reads as real drift. Naive similarity measures get this backwards: one axis can dominate the vector and mask a badly desynced cascade as a false “high.” The log-shape read fixes that — no single component can dominate, and being 0.5× or 2× off counts equally. The exact formulation, thresholds, and per-operator calibration are SigRank proprietary internals.

Where it runs — three time-scales

Macroship now

One drift number per session — a live session-level “tune meter,” updating as the session total moves. Contamination-free, no live hook required.

Windowthe bridge

Drift on the change between successive polls — a rolling time-series. This is where drift becomes a cadence instrument (the timing / burstiness layer). Flagged as a windowed estimate.

Microgated

True per-turn drift — a UX tune-meter. Requires per-turn granularity the source must expose, and only ever as a strictly passive reader.

Sequence: session drift (now, safe) → window-delta drift (the cadence research) → true per-turn micro (passive-only, gated on granularity).

The contamination constraint (non-negotiable)

Any live observer that prompts generates the tokens it measures. We learned this directly: a memory observer that auto-prompts (low-input / high-output) inflated a real operator's output by ~25%, visible openly on the live board as the inflated-vs-clean pair (rows 2 and 3). So every SigRank instrument that touches a live session is read-only against telemetry and emits no prompt, with no auto-memory, no keep-alive, and no self-query. It has to be verified-passive, or it re-contaminates every operator running it. This is a hard requirement rather than a caution, and it is the moat: the instrument that doesn't disturb what it measures.

The drift instrument and the SigRank local agent are governed by the same rule — see the local agent (MCP). The agent is how drift is read live, and the constraint is why it can be trusted.

All signal is monitored. All drift is noted. · Token counts only — never prompt content.

FAQ

What is signature drift in AI usage?

Signature drift measures how much an operator's token cascade shape has changed from their calibrated baseline, in log space. It tracks shape — not magnitude — so an operator who doubles their output while keeping the same input/cache ratio shows zero drift, while one who shifts from cache-heavy to input-heavy prompting shows high drift.

How does SigRank prevent contamination of its measurements?

Every SigRank instrument is read-only and content-free. It counts token quantities (input, output, cache creation, cache read) but never reads prompts, code, or file contents. The contamination constraint ensures that measuring your AI usage cannot alter your AI usage — the observer does not affect the observed.

Why measure drift in log space?

Log space normalizes the scale of token counts so that shape changes are comparable across operators with very different volumes. An operator using 10K tokens and one using 10M tokens can have the same drift score if their cascade proportions shifted by the same ratio. This makes drift a shape metric, not a magnitude metric.