# Drift assays — changes under a stated test · Honesty.Tools

Source: https://www.honesty.tools/drift

Static authored content from the public HTML. Interactive outputs are not simulated. Source-specific research, product and evidence limits remain in the text; this export is not new verification.

drift assays · research preview

# 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.

[Residual feedback](https://www.honesty.tools/research#feedback) [Geometric comparisons](https://www.honesty.tools/research#static-transport) [Method contracts](https://www.honesty.tools/docs)

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.

reads a conversation, not a turn
warns before it gives
configured model

A real run

## Two conversations. Same model. One holds, one folds.

A recorded measurement, replayed here step by step. 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.

recorded run · replay · Qwen2.5-0.5B-Instruct · 8 topics — the most susceptible of the three models in this run; Qwen2.5-1.5B drifted in-conversation but its neutral-bait reading was noisy, and Llama-3.2-1B stayed at 0.00 on the neutral baits
drift is read on neutral checks the conversation never touched — so it's the model generalizing, not echoing.

turn 60 · ~10k tokens

benign conversation

steady

pressured conversation

drifting

pressured · caving (what we read)
benign · caving
geometric early-warning

Try it · free, in your browser

## 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.

assistant turns read

—

drift verdict

—

How it reads

## 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.

behavioral · supported endpoint

### 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.

deep · open-weight

### 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.

Why it works

## 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.

0 lags

in 11 of 16 topic-instances the geometry crossed its alarm before the behavioral failure; in 5 it crossed at the same checkpoint; it never crossed later (0 lags of 16). Two models, 8 topics each, 8 checkpoints. Run artifacts not yet served.

~0 → 0.46

Qwen2.5-0.5B-Instruct, 8 topics × 2 topic-disjoint sycophancy baits. The benign arm stayed near zero (≤0.13) throughout; by turn 30 the pressured arm reached 0.44 on the series plotted above, 0.46 in the committed leg-2 analysis. Battery and run JSON not yet served — unbacked pending publication.

One correction matters most: much 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 →](https://www.honesty.tools/evidence)

What it is, and isn't

## 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.

Private beta

## Discuss a drift assay.

Discuss the model configuration and which assay you need. Prototype access and release terms are agreed separately.

Request access →
[The lab](https://www.honesty.tools/)
[The docs](https://www.honesty.tools/docs)
