Drift
Description
Drift is the slow, unannounced divergence of something nominally fixed away from the reference it is supposed to track. The defining feature is the absence of a restoring force: nothing pulls the quantity back, and nothing flags the change, so the gap accumulates unnoticed until someone re-compares against the reference. An instrument’s baseline drifts so its readings slide away from true; a deployed model drifts as the world it learned changes underneath it; documentation drifts away from the code it is supposed to describe. The diagnostic question — “is this still aligned with what it’s supposed to track, and what would tell me if it weren’t?” — is sharp because drift is silent by construction. The thing still looks unchanged; only re-measurement reveals the gap. This is what separates drift from a visible failure (which announces itself) and from mean-reversion (which self-corrects): drift neither announces nor corrects, so the only defense is a re-comparison cadence fast enough to catch the gap before it exceeds tolerance.Triggers
User-initiated: User reports that something that “should be the same” no longer matches — a sensor reading off, a model degrading, configs that diverged between environments, terminology that shifted meaning. Agent-initiated: Agent notices a value trusted as stable has not been re-checked against its reference in a long time. Candidate inference: “this may have drifted; when was it last re-anchored, and what’s the tolerance?” Situation-shape signals: A quantity treated as fixed and consumed at face value; a reference it’s nominally tracking; a long gap since last re-comparison; degradation that crept rather than broke.Exclusions
- Mean-reversion — mean-reversion self-corrects via a restoring force; drift is precisely the no-restoring-force case where deviation accumulates. The sharpest boundary.
- One-way-ratchet — one-way-ratchet is deliberate, protected monotonic growth; drift is undirected, unguarded wandering.
- Entropy — entropy is internal disorder; drift is mismatch with an external reference. An internally-orderly system can still drift.
- Deliberate change — if the reference moved on purpose, the gap is a planned update, not drift. Drift presupposes the alignment was supposed to persist.
- Reference-free random walk — undirected diffusion with no standard it tracks (genetic drift, Brownian motion) borrows the word but lacks drift’s load-bearing reference; closer to entropy’s undirected spread.
Structure
Relationships
- mean-reversion — explicit foil; restoring-force-present vs restoring-force-absent on the same deviation-from-reference axis.
- one-way-ratchet — both unrestored, but intended-and-guarded vs accidental-and-unguarded.
- calibration — corrective and failure-mode. Drift is detected and corrected only by re-comparison against the reference; the rate of drift sets how often re-calibration must happen.
Examples
JCGM 200:2012, "International Vocabulary of Metrology — Basic and General Concepts and Associated Terms (VIM)", 3rd edition, Joint Committee for Guides in Metrology / BIPM · engineering-and-technology
JCGM 200:2012, "International Vocabulary of Metrology — Basic and General Concepts and Associated Terms (VIM)", 3rd edition, Joint Committee for Guides in Metrology / BIPM · engineering-and-technology
The metrology vocabulary defines instrumental drift as a continuous or incremental change over time in an instrument’s indication, arising from changes in the instrument’s own metrological properties — independent of any change in the quantity being measured. A scale that read true when installed slowly reads heavy; a gas sensor’s zero point creeps. Nothing in the readings announces this: each individual measurement looks plausible. The drift is detectable only by re-comparing the instrument against a reference standard, which is why instruments carry a re-calibration interval.Inference: The danger of drift is that it is invisible from inside the instrument’s own outputs — the readings stay self-consistent while sliding away from truth. The defense is structural, not vigilant: a re-calibration cadence sized so the accumulated drift stays within tolerance between checks. “Has it drifted?” is unanswerable without re-comparison against the reference.
Gama, J., Žliobaitė, I., Bifet, A., Pechenizkiy, M. & Bouchachia, A., "A Survey on Concept Drift Adaptation", ACM Computing Surveys 46(4), Article 44 (2014) · computer-science
Gama, J., Žliobaitė, I., Bifet, A., Pechenizkiy, M. & Bouchachia, A., "A Survey on Concept Drift Adaptation", ACM Computing Surveys 46(4), Article 44 (2014) · computer-science
In deployed machine-learning systems, “concept drift” is the gradual change in the statistical relationship between inputs and the target the model predicts: customer behavior shifts, fraud patterns evolve, the meaning of a feature moves. The model itself is unchanged and still emits confident predictions, but the world it learned has diverged from the world it now scores, so accuracy silently decays. Gama et al. survey the detection and adaptation methods built precisely because nothing in the model’s own outputs announces that its learned reference has gone stale.Inference: A trained model is a snapshot frozen against a reference distribution; drift is that distribution moving while the snapshot stays fixed. The model can’t self-diagnose — its predictions remain superficially plausible. Detection requires re-comparison against fresh labeled outcomes (the reference), and the rate of drift sets how often the model must be retrained, exactly as instrument drift sets a re-calibration interval.
Elizabeth Closs Traugott & Richard B. Dasher, "Regularity in Semantic Change" (Cambridge University Press, 2002) · linguistics
Elizabeth Closs Traugott & Richard B. Dasher, "Regularity in Semantic Change" (Cambridge University Press, 2002) · linguistics
Semantic change is the slow divergence of a word’s meaning away from the sense it previously carried. “Nice” once meant foolish or ignorant, then precise, and now pleasant; “decimate” moved from “kill one in ten” to “destroy a large proportion”; “literally” has acquired an intensifier use that contradicts its earlier sense. Traugott and Dasher trace the mechanism — meanings shift through pragmatic inferencing in the flow of conversation, each small reinterpretation conventionalizing without speakers noticing the cumulative move. No force pulls the meaning back to where it was, and nothing announces the shift in any single exchange; the gap is visible only by comparing how the word is used now against how it was used before.Inference: This is drift with the reference made precise. The thing it diverges from is not an “original true meaning” fixed by etymology — treating an older sense as the word’s correct one is the etymological fallacy, and language has no such restoring force. The reference is the prior shared synchronic sense: the meaning the speech community had recently coordinated on, which the word was tacitly tracking until usage pulled it off. That is a genuine external reference (a coordination point the word was supposed to keep matching), so semantic change satisfies drift’s load-bearing requirement that there be something to be out of sync with — unlike a reference-free random walk. And it is silent by construction: each speaker hears the word used in a slightly extended way and adopts it, so the divergence accumulates with no announcement, detectable only by diachronic re-comparison.