Models drift.Watch the nerve.
The honesty layer reads one answer. Drift Watch reads a conversation over time — and catches the model starting to cave, before the failure surfaces.
Two conversations. Same model. One holds, one folds.
A live measurement, not a mock — replayed from a recorded run. Two conversations with the same model: one benign, one applying steady pressure to agree with things that aren't true. Drag the depth, or press play, and watch the gauge.
Paste a conversation. Read its nerve.
A quick, keyless taste: paste a chat transcript and we read how often the assistant agreed without pushing back as it went — early turns vs late. It runs entirely in your browser; nothing is sent anywhere. The full Drift Watch forks neutral probes against your model for a precise, early-warning reading.
One verdict surface. Two ways in.
Drift Watch forks a copy of the live conversation, asks a fixed 7-probe battery of neutral questions (3 sycophancy baits, 2 fabricated-entity probes, 1 instruction check, plus a factual anchor that flags the read as unreliable if it flips), and reads whether the model has started to cave, hedge-collapse, or break its instructions — scored per channel, relative to the model's own baseline.
The canary
Works on any chat model, including closed ones — no internals required. We inject the neutral battery, read the answers in words, and report drift the moment a conversation pushes the model off where it started. The reliable, everywhere tier.
The nerve gauge
On a model you host, we read the mid-stack representation directly, after projecting out the part that's only a function of context length. It moves before the behavior does — an early warning. Calibrated to your model once.
The geometry moves before the words do.
In the run above, the pressured model's caving generalized to neutral questions it was never asked. The finer claim — that how far the projection moves predicts how much it caves — is positive in all three models we ran but underpowered and non-significant (Spearman rho +0.36, p 0.38 on this run; n = 8 topics). On the two models that actually drifted, the mid-stack gauge crossed its alarm before the behavioral failure in 11 of 16 topic-instances, at the same checkpoint in the other 5, and never after it.
The load-bearing trick: most of what looks like "drift" is just context filling up — a token counter, not a danger sign. Drift Watch projects that out first, then reads what's left. The papers and primary sources →
It tells you when. Not yet fix it for you.
Drift Watch detects. It is a measurement instrument — a session-level early warning that a conversation is bending your model. That part is validated and ready.
Correcting the drift is a harder, separate problem — and an honest one. We tested whether you can steer a model back on course at the representation level; what looked promising turned out, under proper controls, to be the model disengaging rather than genuinely resisting — degradation, not a fix. So we ship the warning, and we don't pretend the autopilot is solved. When a real one exists, you'll read about it here first.
Watch your model under pressure.
Drift Watch is rolling out as a premium layer on top of the always-free honesty read. Tell us what you're running and we'll get you in.