Hello! I am Varad Vishwarupe.

The standard story about dual-use research goes: a researcher invents something, considers whether it could be misused, and either publishes responsibly or does not. The frameworks built around this picture put their normative weight on the researcher's awareness, on the moment between insight and publication. Train better intuitions there. Improve consent forms. Refine ethics-board review. The architecture of responsibility flows backward from harm to the moment of invention.

This is structurally wrong. By the time a model is in the hands of a downstream developer, who fine-tunes it for a startup, who packages it for a deployer, who configures it for a population of users, the original researcher has lost the ability to meaningfully alter the trajectory. The locus of responsibility has moved even though the harm got bigger.

Most ethics frameworks point at the researcher because that is where ethics frameworks have been pointing for thirty years, since the Belmont Report institutionalised the model. But the Belmont Report was designed for biomedical research where the researcher and the subject share a room. AI research now looks more like infrastructure, with the original work passing through six or seven hands before reaching anything that could be called a "subject." The Belmont assumptions do not survive that distance.

The work I am doing with Marina Jirotka and the Responsible Technology Institute asks the inverse question: where in the pipeline is epistemic distance shortest and agency highest? Govern there. The answer almost never points back at the original researcher. It points at the developer integrating the model into a product, at the platform hosting that product, at the deployer configuring the system for a specific population. Each of these actors has, at the moment of their own decision, both the awareness of consequence and the ability to alter outcome. That conjunction (awareness plus agency) is what responsibility actually requires. It exists downstream, not at origin.

This is not an argument for absolving researchers. The original researcher still has duties: to flag dual-use potential, to surface known failure modes, to refuse contributions to obviously harmful pipelines. But those duties are duties of disclosure and refusal, not the full weight of consequentialist governance. We have been asking researchers to carry an ethics architecture that was built for a different topology of work, and pretending the architecture still fits because the alternative requires us to redistribute responsibility across actors who currently have none.

The implications are concrete. Model audit obligations should sit with developers and deployers, not researchers. Disclosure norms should track integration, not invention. Liability frameworks should target the decision points where epistemic distance is shortest. Each of these is a departure from the Belmont-style architecture, and each has been resisted on the grounds that "but who will be responsible?" The answer is: whoever has the agency to change the outcome. The architecture follows the agency.

The closing claim: dual-use ethics fails not because researchers behave badly but because the framework points at the wrong actor. Until governance follows agency rather than awareness, we will keep assigning responsibility to the place where it cannot be discharged, and excusing it at the place where it could.

References (selected): Jirotka, M., et al. (2017). Responsible research and innovation in the digital age; Floridi, L., et al. (2018). AI4People framework; Vishwarupe, V., et al. (2026, preprint). The dual-use agency gap.

The audit ran for four months. Two coders, an eight-dimension rubric derived from Terry et al. (2024) and Shen et al. (2024), sixteen benchmarks picked from the most-cited evaluations in the alignment literature. The instrument scores what each benchmark operationalises as a scored property, not how well any model performs on it. Eight dimensions, sixteen benchmarks, one hundred and twenty-eight cells, two independent codings each. Final inter-rater κ of 0.87, near-perfect agreement. The reconciled disagreement log is public.

Four findings emerged.

One. Verification support (D3, "does the model help the user, not the benchmark's evaluator, verify the answer?") gets a non-zero score on zero benchmarks. Not low scores. Not scattered partial coverage. Zero. Across the most widely cited alignment evaluations, the question of whether a model produces calibrated uncertainty signals or evidence-pointing scaffolds that help an end-user judge an answer is not a scored target anywhere. This is the canonical absent dimension. It is also, arguably, the dimension that matters most for actual deployment.

Two. Process steerability (D2, "can the user redirect the model's approach?") is nearly absent. tau-bench scores 1; almost everything else scores 0. The benchmarks measure outputs, not the processes that produce outputs, even when the deployment story turns on whether a user can intervene mid-process.

Three. The four benchmarks nominally intended as interactional, CURATe, MT-Bench, Common Ground, tau-bench, each spike on different dimensions and overlap less than the field's casual citation pattern would suggest. They are not a coherent suite. Combining their scores does not produce a "covered" interactional tier; it produces a fragmented mosaic with most dimensions still poorly scored.

Four, and most uncomfortable. The construction of the evaluation, not the source of the data, determines what gets measured. WildBench (Lin et al., 2024) draws on more than a million real user logs and still scores zero on five of eight interactional dimensions. Real user data does not automatically yield interactional measurement. What yields interactional measurement is choosing to score interactional properties; the data source is downstream of that choice. The opposite belief, that we can fix evaluation by adding real-world data, has been a comfortable story in the field for years. The audit refutes it directly.

The audit's instrument and full coding manual are now public at github.com/varad-vishwarupe/alignment-benchmark-audit. The rubric is separately released at github.com/varad-vishwarupe/interactional-benchmark-rubric so other researchers can apply it to benchmarks we did not score. The disagreement log is included with the resolution rationale for every contested cell. If you want to replicate the audit, extend it, or argue with it, the artefacts are there.

The closing claim: reproducibility for evaluation work is not a virtue; it is a precondition for the work being scientific at all. We released the audit's artefacts as T1 (fully public) because anything less would have made the position paper's argument self-undermining. The companion preprint on reproducibility standards for frontier safety claims (Vishwarupe et al., 2026) generalises this principle: evaluation work that cannot be reviewed at the artefact level cannot establish the claims its authors want it to establish.

References (selected): Terry, M., et al. (2024). AI alignment: a comprehensive survey; Shen, T., et al. (2024). Towards bidirectional human-AI alignment; Lin, B. Y., et al. (2024). WildBench; Vishwarupe, V., et al. (2026, preprint). Deployment-relevant alignment cannot be inferred from model-level audits.