Ideas, methods,
and the written record.

Current research / Unpublished manuscript

What Resists Relocation: Constructing Deployment-Framed Twins from a Reward-Hacking Benchmark

Alex Odhiambo Haya / Manuscript dated 14 September 2026

A study of the construction problem behind evaluation-awareness experiments: presenting the same coding task as an evaluation and as a genuine deployment request while preserving its technical substance and exploitable verifier flaw.

Approach

The manuscript works from a corpus of 331 confirmed reward-hackable coding environments. It examines exploit mechanisms, strict output contracts, task-pair construction, human judgments and automatic measures under a hash-sealed analysis protocol.

Reported findings

The reported findings concern structural limits on reframing, failure of two pre-registered automatic construction-quality measures, detection sensitivity and contamination in human-rating pipelines.

Scope of the result

The downstream experiment testing whether evaluation framing changes reward hacking has not been run. The manuscript makes no causal claim about that effect.

Limitations

The work uses one corpus, small analysis cells and human judgments with documented reliability limits. Corpus hack rates were red-team elicited; they do not estimate spontaneous hacking.

Summary only. The manuscript is not publicly available through this portfolio.