Drift is measurable
Track deviation from the operating mandate across the developing path.
← Research
Generation-Time Control Research
Control changes the path—and the finished work.
Long-form generation unfolds as a trajectory. Assiduity measures drift against the operating mandate and steers while choices remain.
The model proposes. Assiduity selects.
Selected public results
Generation-time control produced a measurable effect across models and tasks.
89.5% 913 of 1,020 evaluations improved objective retention
d = 1.07 average Cohen’s d across the combined evaluation set
5 models · 4 corpora results span open and frontier model groups
Open models: N=700 · 95.0% improved · d=1.30 Claude / OpenAI: N=320 · 77.5% improved · d=0.55
What the evidence shows
Evaluate the path, not only the answer.
Final-output review shows where the model ended. Generation-time evidence shows whether the mandate stayed influential while the work developed.
Selection changes outcomes
Mandate-aware continuation selection improved objective retention versus greedy generation.
Specificity matters
The benefit falls when the semantic contract loses task-relevant content.
The effect crosses models
Results held across the tested model families and corpora.
Control can be selective
Sparse branching concentrates effort where drift risk rises.
Placebo controls
More samples are not the same as control.
Additional sampling can improve an output without proving mandate-sensitive selection.
When the semantic contract loses task-relevant content, the benefit declines. That points to the mandate—not sampling alone—as the source of control.
Generation Control Record
Return evidence of how control was applied.
Preserve the path behind the finished work—not only the finished work.
Research access
Public findings. Controlled technical review.
We publish validated findings and evaluated settings. Detailed tables, parameters, implementation methods, and patent-sensitive materials remain available through controlled review.
Technical review
Review the evidence behind generation-time control.
Compare the findings, inspect a controlled run, or evaluate a workflow.