Documentation contents
DOCS / METHODS AND INTERPRETATION
Methods and interpretation
Claims, assumptions, uncertainty, and responsible use.
Three kinds of output
Marres keeps three epistemic categories separate:
- Computed — deterministic transformations or model outputs produced from stated inputs.
- Machine-proposed — classifications, summaries, interpretations, or actions proposed by a language model or semantic layer.
- Analyst-approved — interpretations or recommendations accepted by a responsible human after review.
Machine-proposed content does not become evidence merely because it is fluent.
Descriptive, predictive, and causal claims
- Descriptive: what appeared in the data or evidence.
- Predictive: what a model estimates under specified conditions.
- Causal: what would change because of an intervention.
The first two do not automatically establish the third. Causal attribution generally requires a credible experimental or quasi-experimental design and appropriate data.
Minimum interpretation checklist
Before using a result, record:
- source and collection period;
- missingness and exclusions;
- domain or column mapping decisions;
- sample and coverage sufficiency;
- model version and material settings;
- assumptions supplied by the analyst;
- sensitivity to plausible alternatives; and
- known events outside the model.
Language-model use
Language models may help filter evidence, score material against a schema, propose semantic interpretations, draft competitor responses, or synthesize a brief. Their outputs can vary and may be wrong. Structured outputs, bounded prompts, provenance, review gates, and deterministic calculations are used where appropriate to limit that risk.
Reproducibility
Keep source files, approved mappings, run identifiers, settings, and exported artefacts together. If the evidence or assumptions change materially, create a new run or version rather than silently overwriting the basis of the earlier decision.
Responsible use
Do not use Marres as the sole basis for decisions that determine a person's legal rights, access to essential services, employment, credit, insurance, or medical treatment. Do not infer sensitive personal traits from public content. Follow the source platform's terms, applicable privacy law, and the client's lawful-purpose restrictions.
Documentation draft · Source baseline 16 August 2026