Causal mediation of the Answer Engineering gain

Determine whether TF–IDF recall, All-8 completion, or any other individual measured trajectory feature mediates the downstream Answer Engineering effect of instruction duplication, using a matched controller-on/off factorial design.

Background

The paper finds that duplicating procedural instructions changes several machine-observable trajectory measures, including TF–IDF recall and All-8 protocol completion, while also changing the performance of a downstream Answer Engineering (AE) system. In the SSNHL diagnostic setting, system-only AE achieves 84.2% and AE with a trailing duplicate achieves 97.1%, but the current experimental design does not determine which, if any, measured trajectory feature produces that gain.

The authors identify a matched controller-on/off factorial experiment as necessary to separate the direct effect of instruction duplication from mediation through specific observability metrics. This remains unresolved because the present study measures trajectory changes and downstream performance without experimentally manipulating or isolating the proposed mediators.

References

The present experiments do not establish that TF--IDF, All-8, or any other single measured feature mediates that downstream effect; a matched controller-on/off factorial would be required for that causal decomposition.

— Instruction Duplication as an Inference-Time Control Primitive  (2609.04024 - Lavrenko, 3 Sep 2026) in Section "Why Trajectory Compliance Matters to Deterministic Controllers"; reiterated in Section "Limitations and Ethics"