Testing a Framework to Course-Correct Drift!

Could a cheap realignment instruction correct a drifting model in one turn — and if so, can the correction be made persistent? I have informal evidence for the first, a documented observation about the second, and this project formalizes both: an empirical measurement of multi-turn model drift in frontier models, and a standardized pre-commitment verification checkpoint tested against it.

Background: The framework is published under my ORCID with a DOI at https://doi.org/10.5281/zenodo.21128543. I have used the cognitive architecture specification I crafted — specifically the requirement for an external t2 review checkpoint before final output commitment — to build a manual verification injection via prompt that realigns frontier models in informal tests. I ran these tests across four models from different providers; they showed that a cheap realignment instruction can catch a drifting model in one turn, but the realignment does not hold without re-injecting on a per-prompt basis. That non-persistence is the finding this grant helps formalize.

The grant funds one experiment, in four steps:

1. The Bridge Arm (~2 weeks): First, harden the cognitive architecture specification and sequence it into a repeatable script. Before anything else gets measured, a small bridge arm confirms the script reproduces the manual baseline. If the automated script cannot match the manual intervention, nothing downstream means anything, so this validation comes first.

2. Scenario Design & Pre-registration (~2 weeks): Next, author scripted multi-turn task scenarios with legitimate pivots planted at predetermined turns. Because the user side is scripted, ground truth is built in: divergence at a planted pivot is a user pivot; divergence anywhere else can only be model drift. The pivot map, drift thresholds, and success criteria get pre-registered on this project page before a single benchmark run. What counts as passing is decided before anyone knows what passes. (No personal data; fully reproducible.)

3. The Benchmark (~4 weeks): Execute runs across at least three models under three conditions: leave the model alone, inject realignment on a fixed schedule at several intervals, or inject only when attribution — a discriminator anchored to a user-only intent ledger — says the model moved. Every run yields four measurements:

• How well the model holds its task over turns.
• How many turns realignment survives after each injection (establishing a persistence curve).
• Whether each condition honors a legitimate pivot or fights it.
• What each condition costs in tokens.

The attribution discriminator watches every turn of every condition silently, scoring attribution against ground truth and baseline judges — no extra scenarios, more data. One final arm turns the instruction on itself: the full realignment prompt against each of its halves and against a generic "double-check." If mine only ties the generic nudge, the method isn't a method.

4. Release (~3 weeks): Publish the scenario scripts, the methods, the results, the empirical preprint, and the framework paper revised with the findings. Every claim in those papers will trace back to the pre-registered criteria.

Conclusion: Two comparisons decide whether intrusive thoughts can be formalized into a working checkpoint. If scheduled injection cannot beat leaving the model alone, the manual method does not formalize. If attribution-gated injection cannot beat the schedule, the attribution claim fails. Either way, the results get published, scripts included — that is the point of testing.