# From model answer to retest, every step keeps its evidence.

> Edith retains questions and answers. Homer governs the approved fact versions used to judge and generate. Friday records authorized work and verified publication results. Edith then retests comparable conditions.

- Canonical page: https://winin.ai/en/methodology/
- Language: English
- Last updated: 2026-09-04

## One method, three product responsibilities.

*01 · OBSERVE*

### [Edith monitors and retests](/en/products/edith/)

Questions, model conditions, raw answers, sources shown in the answer, baselines and comparable retests.

*02 · ALIGN*

### [Homer governs facts and versions](/en/products/homer/)

Sources, conflicts, accountable owners, markets, approved versions, validity and public permissions.

*03 · ADVANCE*

### [Friday advances authorized work](/en/products/friday/)

Evidence-linked tasks, accountable approval, publication verification, content versions and rollback records.

> **Shared protocol:** method versions, evaluators, human review, qualification and stated scope apply across all three products.

## Measurement principles

### Separate evidence types

Audited capability, sampled representation and operating maturity are reported independently.

### Preserve raw observations

Every claim can be traced to a source, response or test record.

### Version everything

Prompt sets, models, judges, facts and scoring rules carry effective dates.

### Show change stability

Repeated sampling shows the strength, consistency and durability of observed movement.

### Use comparisons

Changes are read against a preserved baseline and relevant comparison tasks.

### Define the scope

Coverage, data sufficiency and review status make every conclusion easier to apply.

## The evidence record

| Record | Primary owner | What remains inspectable |
| --- | --- | --- |
| Question and run | Edith | Customer question, locale, model, mode, retrieval state and timestamp |
| Observation | Edith | Raw answer, sources shown in the answer, refusal, tool use and run state |
| Fact decision | Homer | Source, candidate, conflict, accountable owner, approved version and validity |
| Work and publication | Friday | Finding, task, content version, approval, target URL and publication verification |
| Comparable retest | Edith | Matched measurement contract, comparison record and bounded verdict |

## Calibration

1. **Benchmark set:** Maintain websites with known strengths, faults and controlled changes.
2. **Human review:** Double-score material claims and adjudicate disagreements.
3. **AI evaluation:** Use versioned judges only for tasks with measured reliability.
4. **Online verification:** Compare predicted improvements with repeated observed outcomes.
5. **Regression suite:** Re-run stable tasks after model, crawler or rubric changes.

## Change policy

Method versions are never silently overwritten. Material changes receive a new version, migration note and back-test. Historical reports retain the rules and systems under which they were produced.

> Current public method status: **V1 research protocol**. Weight calibration and cross-industry norms remain under active validation.

## Source

This Markdown document is the machine-readable counterpart of https://winin.ai/en/methodology/. The canonical HTML page remains the public presentation source.