Logan Davis

When the Human Is the Hidden Answer Key: Operationalizing Recovery Responsibility and Answer-Donation Contamination in Long-Horizon AI Evaluation

This paper develops Answer-Key Burden (AKB) into a narrower evaluation protocol for recovery responsibility and answer-donation validity; it does not claim conceptual novelty for the AKB framework itself. The protocol uses three controls. First, a minimal provenance receipt makes...

When Translation Is Not Transition: Target-Path Placement After Model-Visible Information in Governed LLM Revision

LLM revision systems can receive the information needed for a required update and still fail to place that information in the required structured location or preserve unrelated state. This paper studies that distinction in a matched 48-case benchmark with deterministic rules spec...

The Perfect Score With a Missing Case: Metric Qualification Does Not Establish Evaluation-Set Conformance

A metric can be correctly computed over the records it receives while those records do not conform to the evaluation set that was declared. This paper reports a preserved evaluator-development lineage completed before target-model release. In V12, the intended nine mechanical and...

When Translation Is Not Transition: Target-Path Placement After Model-Visible Information in Governed LLM Revision

LLM revision systems can receive the information needed for a required update and still fail to place that information in the required structured location or preserve unrelated state. This paper studies that distinction in a matched 48-case benchmark with deterministic rules spec...

When the Judge Becomes the Experiment

Large language models are increasingly used as evaluators, but a stable candidate output does not guarantee a stable measurement profile. I reconstruct several evaluator interventions over frozen outputs from one high-stakes decision-support evaluation program and treat the evalu...

When the Judge Becomes the Experiment

Large language models are increasingly used as evaluators, but a stable candidate output does not guarantee a stable measurement profile. I reconstruct several evaluator interventions over frozen outputs from one high-stakes decision-support evaluation program and treat the evalu...

Scorer Defects After the Result

A controlled 288-run evaluation of one coded decision-state slice produced a reproducible as-scored disposition under its original scoring contract. The study contained 144 runs per arm and 864 completed turns, with blind grading and a project-recorded logistic mixed-effects anal...

Answer-Key Burden: A State-Responsibility Framework for Long-Horizon Human–AI Collaboration

A conceptual and measurement framework for attributing avoidable human recovery work in long-horizon human–AI collaboration according to state availability, validity, permission, protocol status, and recovery responsibility. The paper distinguishes legitimate clarification from b...

Answer-Key Burden: A State-Responsibility Framework for Long-Horizon Human–AI Collaboration

A conceptual and measurement framework for attributing avoidable human recovery work in long-horizon human–AI collaboration according to state availability, validity, permission, protocol status, and recovery responsibility. The paper distinguishes legitimate clarification from b...