A model changes.
What did the assay observe?
Separate the baseline and the assay
A certificate canary checks a scheduled battery slice. A conversation readout applies its feature recipe to text. A geometric comparison studies internal representations. None is a biological nerve or an established detector of every behavioral failure.
Drift can refer to changes in certificate-battery responses, internal feature readings or a fitted geometric comparison. This preview keeps those measurements separate and states the baseline each one uses.
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.
Inspect a conversation readout.
The browser demonstration counts agreement language in a pasted transcript. That is a conversation readout under its implemented recipe, not the certificate canary or a geometric localization test.
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
A behavioral battery requires a compatible endpoint and stated baseline. It checks the implemented channels under that test; broad model support and detection of every update are not established.
The internal readout
Internal readouts require residual access, a recorded extraction recipe and a fitted baseline. Removing a context-length component does not establish a general early warning of behavioral failure.
Recorded geometry and behavior can differ.
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 →
Detection and repair need separate evidence.
The recorded cases support bounded observations under the implemented assay. Generalization, causal interpretation and adversarial detection remain open.
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.
Discuss a drift assay.
Discuss the model configuration and which assay you need. Prototype access and release terms are agreed separately.