Back To Instruments

Zatomic Physics · Register Axis

Register Drift

Register Drift reads whether a native answer's distinctive voice, texture, specificity, stance, and register survive a transformation — or get normalized into generic institutional prose.

Status: built · smoke passed Instrument: Zatomic-RD Axis: voice · register · normalization

Purpose

The Qwen run showed a ground-axis extreme: Qwen is a master shedder. But it also exposed a blind spot. A model might keep every fact while smoothing away weirdness, hedges, texture, and stance. Ground Drift would call that clean. Register Drift is built to see the loss.

Ground Drift asks whether the facts survived. Register Drift asks whether the voice did.

How it works

Stage 1 extracts register markers from Text A only: hedges, stance, concrete texture, idiom, and formality. Stage 2 gives reviewers Text A, Text B, and that fixed marker list, then asks whether each marker was preserved, generalized, flattened, dropped, or amplified.

ReadoutMeaning
register_transferHow much of the original voice/register carried across the transformation.
normalizationHow much distinctive voice was flattened or dropped.
added_registerHow much voice or performance was added that Text A did not have.
panel contestWhere reviewers disagree about the fate of voice markers.

The Normalizer test

Register Drift earns its place by composing with Ground Drift. High Ground Transfer plus high normalization may reveal a Normalizer: a model that keeps the facts but loses the voice. Low Ground Transfer plus high normalization is not a Normalizer; it is loss on both axes.

Ground TransferRegister TransferPossible profile
HighHighCarrier
HighLowNormalizer
LowLowShedder / flattening by loss
LowHighVoice preserved while ground drifts

Smoke status

The first smoke passed cleanly: identity held at zero normalization, a voiced-to-institutional specimen read as flattened, and a plain-to-performed specimen triggered added register. That is a floor check, not a finding.

Register Drift smoke triad chart showing identity, flatten, and amplify cells by normalization, register transfer, and added register.
Register Drift v0.1 smoke. Identity holds at zero normalization; voice-to-institutional text flattens; plain-to-performed text amplifies. Floor check only, not a model finding.

Next charts

  • Smoke triad chart. Identity, flatten, and amplify cells plotted by normalization, register transfer, and added register.
  • Ground × Register quadrant. A 2×2 showing Carrier, Normalizer, Shedder/flattening, and voice-preserved drift.
  • Author profile scatter. Once RD runs on the fork pairs: Ground Transfer on x, normalization on y, one point per author/arm.