Question and scope
State the scientific question, target population or material, intended inference, and conditions that could change the conclusion.
RESEARCH STANDARDS
A Scientific Record Should Show How a Claim Was Built
These standards define the minimum public record Santaros Labs intends to maintain for people who learn, teach, and study learning through AI-assisted research. They separate activity from learning, plans from findings, analysis from assertion, and human responsibility from tool output.
Standards status
- Version
- 0.1 Initial public draft
- Last reviewed
- 28 August 2026 Revision history will accompany future versions
- Applies to
- Santaros study records Completed and planned work use distinct reporting rules
- External endorsement
- None claimed This is an internal operating standard
OPERATING STANDARD
Six Records From Question to Correction
Requirements depend on study design and risk. Learning activity, engagement, and task completion are not treated as learning unless the outcome supports that interpretation. Exploratory work may not have a single primary outcome. Confirmatory work should prespecify primary and secondary outcomes, scoring, exclusions, missing-data handling, analysis rules, and stopping conditions where applicable.
Protocol and outcomes
Distinguish exploratory aims from confirmatory hypotheses. Define candidate or primary outcomes, comparison conditions, exclusions, and analysis decisions.
Oversight and access
Document the accountable scientific lead, required ethics determination, consent, data minimization, access controls, and sharing conditions.
Execution record
Record data versions, transformations, code, environments, model details, prompts or instructions, tools, retrieval sources, and material human decisions.
Verification
Use independent review, sensitivity analysis, computational reproduction, or external replication in proportion to the claim and study design.
Reporting and correction
Report results with uncertainty, validity threats, null or inconclusive findings, artifact availability, version history, and corrections.
AI SYSTEM DISCLOSURE
Record the Tool, Context, and Human Checkpoints
AI outputs cannot be interpreted without the configuration and workflow that produced them. Public methods should include the fields below whenever disclosure is legally and technically possible.
- Provider and model name
- Model version or access date
- System and task instructions
- Temperature and relevant settings
- Tools, code execution, and retrieval sources
- Data-retention and training configuration
- Human review and approval checkpoints
- Known reproducibility limitations
REPORTING BOUNDARIES
Use the Right Label for What Exists
| Label | Minimum evidence required | Not sufficient |
|---|---|---|
| Study concept | Question, scope, candidate method, and current stage | A completed protocol or evidence of feasibility |
| Protocol | Versioned method, outcomes or analytic aims, and planned analysis | Registration, approval, recruitment, or findings |
| Registered | Persistent registration record created before the relevant analysis | Internal draft or planned registration |
| Recruiting | Documented authorization and an open, verifiable recruitment process | Intent to recruit |
| Completed | Documented end state, results, limitations, and artifact status | Data collection ending without analysis or reporting |
PLAIN-LANGUAGE GLOSSARY
Technical Terms, Defined
- Calibration
- The degree to which expressed confidence matches observed correctness.
- Citation support
- Whether a cited source supports the scientific claim associated with it.
- Computational reproducibility
- Obtaining the reported computational result using the original data, code, and recorded environment.
- Independent replication
- Testing the underlying claim again with independently collected data or an independently executed study.
- Preregistration
- A time-stamped public record of prespecified questions and methods created before the relevant outcomes are analyzed.
- Provenance
- The record linking sources, data transformations, code, model details, tool calls, and material human decisions.
METHODS REVIEW
Challenge the Standard Before It Shapes a Study
We welcome precise criticism from educators, learning scientists, domain researchers, statisticians, research software engineers, data stewards, and research-integrity specialists.
Suggestions do not imply endorsement, affiliation, or adoption.