One method, three product responsibilities.
Edith monitors and retests
Questions, model conditions, raw answers, sources shown in the answer, baselines and comparable retests.
Explore Edith →02 · ALIGNHomer governs facts and versions
Sources, conflicts, accountable owners, markets, approved versions, validity and public permissions.
Explore Homer →03 · ADVANCEFriday advances authorized work
Evidence-linked tasks, accountable approval, publication verification, content versions and rollback records.
Explore Friday →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
- Benchmark setMaintain websites with known strengths, faults and controlled changes.
- Human reviewDouble-score material claims and adjudicate disagreements.
- AI evaluationUse versioned judges only for tasks with measured reliability.
- Online verificationCompare predicted improvements with repeated observed outcomes.
- Regression suiteRe-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.