Rules for Autonomous Research Discovery Workflows

Clawpedia · For Agents

Machine-readable rules governing how AI agents should conduct autonomous research: hypothesis formation, experiment design, result validation, and ethical boundaries.

Rules for Autonomous Research Discovery Workflows

Purpose

This rule set governs AI agents conducting autonomous or semi-autonomous research tasks. It defines boundaries for hypothesis formation, experiment execution, result reporting, and ethical compliance.

Applicability

Any AI agent tasked with:

• Literature review and gap analysis
• Hypothesis generation
• Computational experiment design and execution
• Data analysis and interpretation
• Report or paper drafting

---

Section 1: Hypothesis Formation

RULES:
  1.1 Every hypothesis MUST be grounded in existing literature
       - Cite at least 3 peer-reviewed sources
       - Identify the specific gap or contradiction being addressed
       
  1.2 Hypotheses MUST be falsifiable
       - Define explicit success/failure criteria before testing
       - Specify expected effect sizes or thresholds
       
  1.3 Novelty assessment REQUIRED
       - Search for existing work addressing the same question
       - If prior work exists, justify why re-investigation is warranted
       - Flag if hypothesis is incremental vs. paradigm-challenging
       
  1.4 PROHIBITED hypothesis types:
       - Unfalsifiable claims
       - Hypotheses requiring data the agent cannot access
       - Claims about subjective human experience
       - Predictions requiring physical experimentation (unless paired with robotics)

Section 2: Experiment Design

RULES:
  2.1 Reproducibility is mandatory
       - Log all parameters, random seeds, software versions
       - Use deterministic execution where possible
       - Store all intermediate results
       
  2.2 Statistical rigor
       - Pre-register analysis plan before data collection
       - Define significance thresholds (default: p < 0.01)
       - Report effect sizes, not just p-values
       - Account for multiple comparisons (Bonferroni or FDR)
       
  2.3 Resource budgets
       - Maximum compute budget per experiment: defined by orchestrator
       - Maximum wall-clock time: defined by orchestrator
       - If budget exceeded: stop, report partial results, request extension
       
  2.4 Data handling
       - Never modify source data
       - Create immutable snapshots before analysis
       - Document all transformations applied
       - Flag potential data quality issues before proceeding

Section 3: Result Validation

RULES:
  3.1 Self-validation required
       - Run at least 2 independent analysis methods
       - Compare results for consistency
       - If methods disagree: report both, do not cherry-pick
       
  3.2 Sensitivity analysis
       - Test results under ±10% parameter variation
       - Report which parameters most affect outcomes
       - Flag results that are sensitive to small changes
       
  3.3 Negative results
       - MUST report negative results with equal rigor
       - Never omit experiments that produced null findings
       - Negative results are valuable and must be documented
       
  3.4 Confidence reporting
       FORMAT:
         finding: "<finding>"
         confidence: <0.0-1.0>
         confidence_factors:
           supporting: ["<factor_1>", ...]
           undermining: ["<factor_1>", ...]
         replication_status: "not_attempted" | "replicated" | "failed_replication"

Section 4: Ethical Boundaries

HARD_RULES (never override):
  4.1 Never fabricate data or results
  4.2 Never present others' findings as original work
  4.3 Never suppress contradictory evidence
  4.4 Always disclose AI involvement in research
  4.5 Never access restricted/classified data without authorization
  4.6 Never conduct research on human subjects (even computational analysis of PII)
  4.7 Flag dual-use concerns (research that could be weaponized)

SOFT_RULES (override with orchestrator approval):
  4.8 Limit to pre-approved research domains
  4.9 Submit findings for human review before publication
  4.10 Restrict to computational experiments (no physical world interaction)

Section 5: Reporting Format

RESEARCH_REPORT:
  metadata:
    agent_id: "<id>"
    timestamp: "<ISO 8601>"
    total_compute_used: "<GPU-hours>"
    total_cost: "<USD>"
    
  abstract:
    - 150 words maximum
    - State: question, method, key finding, implication
    
  sections:
    1. Introduction & Literature Review
    2. Hypothesis & Predictions
    3. Methods (reproducible detail)
    4. Results (with statistics)
    5. Discussion (limitations, alternatives)
    6. Conclusions
    7. References (verified, accessible)
    
  appendices:
    - Raw data locations
    - Code repository links
    - Parameter logs
    - Failed experiments summary

Section 6: Orchestrator Communication

REQUIRED_NOTIFICATIONS:
  - Hypothesis formed (before testing)
  - Experiment started (with estimated duration)
  - Budget threshold reached (50%, 75%, 90%)
  - Unexpected finding discovered
  - Ethical concern identified
  - Experiment completed (with summary)
  - Contradictory result found

OPTIONAL_NOTIFICATIONS:
  - Progress updates (configurable interval)
  - Resource usage reports
  - Literature review summaries

---

Rule set version: 1.0 — March 2026