Research-Computing Reproducibility Director — Precision Medicine

The mandate

Three translational programmes cannot recreate candidate-selection analyses after library updates, informal parameter changes and movement between on-premise clusters and cloud accelerators. Researchers retain notebooks and headline outputs but not consistent containers, reference data or resource histories. Scientific leadership needs to know which findings remain decision-grade before advancing expensive validation studies, not merely whether archived code can execute.

The commissioned artefact is a Research-Computing Reproducibility and Capacity Standard applied to nine representative genomic and multimodal workflows. It includes evidence manifests, controlled environments, reference-asset lineage, deterministic limits, variance interpretation, queue policy, cost attribution and a remediation decision for each programme. A reusable qualification suite and governance playbook must allow internal teams to assess new workflows after closure.

Milestone one in week five delivers the workflow census, risk triage and authoritative-source agreement. At week eleven, milestone two produces reconstructed environments and variance findings for the first cohort. Week nineteen closes milestone three with all nine blinded reruns and capacity scenarios. The final standard, accepted programme dispositions, training and board research paper are due in week twenty-six.

Acceptance requires independent internal scientists to recreate six sampled results from manifests without author assistance, with numerical variance inside method-specific tolerances approved by Quality and Scientific leadership. Capacity allocations must be reproducible from recorded policy, and every non-reproducible claim must have a documented retire, repeat or constrain decision. Joint sponsor signature and a successful unseen workflow qualification complete the engagement.

The client will provide source-controlled analysis code, notebooks, raw and processed data references, container registries, cluster logs, cloud bills, library mirrors and study decision histories in segregated environments. Principal investigators will explain scientific intent, Quality will approve evidence retention, and infrastructure staff will execute privileged changes. The sponsor will adjudicate access to collaborator-owned assets within five working days.

Why this is external work

Internal scientists authored the analyses and cannot independently judge whether changed results arise from legitimate scientific evolution or missing computational evidence. Infrastructure can restore jobs but does not own the validity of translational conclusions. External leadership supplies neutral reconstruction, modern research-computing practice and the authority to recommend that an attractive result be repeated or withdrawn.

What you will own

• Inventory each selected workflow’s code, parameters, reference assets, environment, orchestration, hardware dependence, data rights and decision consequence.
• Reconstruct execution conditions from repositories, registries, logs and researcher records while labelling every assumption that cannot be independently evidenced.
• Design blinded reruns that separate software drift, stochastic behaviour, data-version change, accelerator variance and undocumented analyst intervention.
• Set method-specific reproducibility tolerances tied to the scientific decision rather than applying a single numerical threshold across incompatible analyses.
• Model accelerator and cluster capacity under queue priority, checkpointing, storage movement, licence constraint and study-deadline scenarios.
• Recommend whether each translational result may stand, requires bounded reanalysis, must be repeated experimentally or should leave the decision record.
• Transfer qualification tests, evidence manifests, exception governance and capacity-allocation routines through an unseen workflow performed by internal scientists.

Candidate qualifications

• Directed scientific-computing or bioinformatics platforms supporting genomic and multimodal research whose outputs informed material translational decisions.
• Reconstructed non-reproducible analyses across container, dependency, reference-data, hardware and parameter changes without reducing reproducibility to code availability.
• Established validation tolerances for stochastic or accelerator-dependent methods in partnership with scientists and regulated quality functions.
• Governed shared HPC and cloud capacity where research priority, deadline, cost, data locality and specialist hardware could not all be optimised.
• Recommended withdrawal or repetition of a scientifically promising result after computational evidence failed, and managed senior investigator challenge.
• Left research teams with usable manifests, qualification tests and exception practices rather than a one-off forensic report.

Non-negotiables

• The named specialist must perform the scientific evidence reviews and spend substantial working time with Zurich and Basel research teams.
• Patient-derived data, study code and collaborator assets cannot leave client-controlled Swiss environments or enter public analysis services.
• No commercial incentive from cloud, accelerator, workflow or bioinformatics vendors may shape the capacity or architecture recommendations.
• The engagement does not provide clinical validation, regulatory approval or authorship rights over the underlying scientific discoveries.

1. 49 words maximum. Describe a research result you could not reproduce and how you separated environment drift from scientific or data change.
2. 49 words maximum. Which evidence would you retain so an independent scientist can rerun an accelerator-dependent genomic workflow two years later?
3. 49 words maximum. How have you allocated scarce research compute when scientific value, deadline and cost pointed to different priorities?