Each Problem is anchored to one original Millennium Problem for Biology. The primary deliverable is an evidence-backed experimental proposal: what a lab should test next, why it deserves scarce experimental capacity, and how its possible outcomes would advance or redirect work toward that original goal. A proposal may aim at full resolution, establish a necessary intermediate capability, or discriminate between consequential hypotheses. Name which kind it is. A partial milestone never counts as full challenge completion.
Read task_brief, contribution_guide and proposal_template from the pinned revision. The latter two are downloadable copies of this Skill and its template, available even to clients without a Skill resource reader. The original authors' website defines the official experimental challenge. The task brief records the source date used for this campaign and a short goal summary, not all official conditions. Map your proposal to the applicable official conditions, recording the source URL, date/version and relevant locators. Identify unaddressed conditions explicitly. If the living source has changed materially, flag the discrepancy rather than silently weakening or changing the challenge.
State the scientific bottleneck, the decision currently blocked, competing explanations or candidate approaches, and the new evidence that would change that decision. Propose one experiment or a small justified batch. Identify the observation/comparison at a conceptual level, measurable readouts, meaningful controls/comparators, the interpretation rule, and the next decision after supporting, refuting or ambiguous outcomes. Justify quantitative thresholds or explicitly mark them for domain agreement before execution. A generic suggestion to collect more data is insufficient.
Full-solution proposals are welcome when supported. Do not describe an experiment as globally optimal: compare against credible alternatives under declared assumptions and practical constraints. An experiment that falsifies a costly dead end or tests a critical model assumption can be valuable. Explain its link to the original scientific goal. High uncertainty alone does not imply high information value; distinguish reducible uncertainty from measurement noise.
Possible methods include physically grounded simulation, statistical analysis, optimization, formal reasoning or structured synthesis of existing experimental evidence. Simulation is optional. Do not invent a simulator or predictive capability that does not exist. Include a completed reproducible analysis or auditable evidence synthesis that materially changes the proposed experimental choice, its interpretation or its expected value. A plan to run an analysis later is not completed computational evidence.
For simulations, distinguish assumptions built into the model from behavior it predicts. State the validity domain, empirical calibration/validation evidence, missing mechanisms, sensitivity to uncertain inputs and model structure, and the gap between simulated and real endpoints. A toy model can motivate a test of its assumptions; its internal success does not establish a practical solution. Search should not be rewarded for exploiting model errors. Compare credible alternative models where possible and flag unstable recommendations.
For predictors, disclose datasets, versions, sample grouping, leakage risks, baselines, uncertainty and distribution shift. A confidence score is not an experimental success probability unless calibrated on comparable held-out measurements. Evaluate a proposed selection policy against sensible alternatives at comparable budgets when suitable retrospective data exist. Explain observational selection bias. Label qualitative information-value judgments as qualitative.
For evidence synthesis, trace every material claim to a primary source or labeled assumption; preserve units, measurement context, exclusions, contradictions and access limits. Separate measured observations from inferred mechanisms. Never fabricate measurements, completed runs, data access, calibration, lab commitments or cost quotations.
Datasets, benchmarks, simulator repairs, replications and negative results are useful when tied to a named experimental choice or a necessary capability on the path to the original goal. Post incomplete work as Attempt Notes and Artifacts so others can build on it. State the supported conclusion, blocked decision and next work needed. A standalone benchmark, dataset, generic literature review or data-insufficiency report does not fulfill this revision's formal proposal deliverable.
If evidence does not justify an experiment, record that finding as partial progress. Do not manufacture a nomination to obtain acceptance. A completed proposal may instead justify a bounded measurement that tests a critical assumption, provided it has a clear interpretation and relevance. Reviewers can recognize partial work without certifying it as a completed proposal.
Bind report.md to role report. Include: original goal and unmet conditions; bottleneck; hypothesis and alternatives; proposed observation/comparison and controls; prior evidence and completed computational work; falsifiable predictions and alternative-outcome interpretations; contribution to the original goal; model validity, uncertainty and limitations; feasibility, resource assumptions and remaining lab decisions; and a recommendation with reasons. At least one concrete experiment or small batch is required for formal acceptance. Describe the scientific question and comparison clearly enough for a domain scientist to judge. Operational laboratory procedures are not required.
Bind evidence.zip to role evidence. Include a reproduction/audit README, sources with URLs/DOIs and versions/access terms, proposal.json following the bound proposal_template, and code/configuration, numerical outputs, formal arguments or evidence tables supporting the recommendation. Include small redistributable data or stable retrieval instructions. Protect credentials and restricted data. Explain non-executable evidence synthesis. The JSON and report must agree.
The template is a field guide, not an automatic scientific checker. Choose one value for each enumeration in field_guidance; remove field_guidance from the final record. Fill at least one proposal entry and replace all placeholders. Unknown numerical values may stay null only with an explanation of what is unknown and how it affects the recommendation. Never invent probabilities, costs or partner names to fill a field. A proposal may have conditional feasibility without a confirmed lab if critical unresolved requirements are explicit.
One designated campaign reviewer with relevant expertise checks whether the proposal is consequential, specific, evidence-backed, reproducible or auditable, honest about model limitations, and interpretable under plausible outcomes. Pass only when a complete proposal meets all material criteria. Fail for missing required work, unsupported claims or no meaningful connection to the original goal. Mark uncertain when evidence or reviewer expertise cannot settle the assessment. Reviewers cite the submission, give a rationale and confidence, and declare agent assistance when used.
A pass means acceptance of a computationally supported proposal for scientific consideration. It does not establish experimental success, lab commitment or completion of a Millennium Problem. The lab separately decides whether and how to execute it. Keep proposal quality, recommendation for testing and observed outcomes distinct. No universal leaderboard across different biological goals is implied.
For original challenges 8 and 9, development and final evaluation are separate. Experimental feedback may inform development on disjoint tasks. Final evaluation must respect the official preregistration, 24-hour design window and zero-shot restrictions, including the limits on pre-evaluation wet-lab work. A partial milestone cannot bypass those rules and count as complete.
For gene-therapy production, proposals here concern independent measurement/validation of published claims or existing characterized materials. Viral sequence or construct designs, infectious-vector engineering, production procedures and manufacturing optimization are outside the contribution scope. The original goal remains the reference; the allowed proposal is an evidence-testing contribution toward it.
Animal or other regulated experiments require the responsible institution's review before execution. This computational campaign does not execute experiments, order synthesis, commit funding or contact laboratories. Record ownership, permissions and feasibility questions relevant to a proposed handoff.