Map

Evidence and Learning

The tests, controls, observations, open questions, and stop conditions that can change the Atlas thesis.

Areas

5 Areas · 12 Points

Conformance evidence

Reproducible proof that processors implement the same portable contract.

3 PointsEvidence bounds maturity claims The Meta-Atlas is implementation pressure Self-hosting does not prove product value

Product evidence

Human and agent task outcomes that can test the useful difference.

8 PointsCross-context benefit must be measured A document-first contextual reader is intended Evidence bounds maturity claims

Controls

Same-corpus generic readers and precommitted tasks that isolate semantic effects from interface quality.

3 PointsPre-UI pilots prove protocol readiness only A same-corpus reader control is required The useful test controls semantic scope

Costs and failure modes

Authoring, review, maintenance, accessibility, performance, and integration burdens.

3 PointsAtlas can be reduced or rejected A document-first contextual reader is intended Maintenance cost can defeat Atlas

Learning agenda

Open questions, re-entry gates, simplification triggers, and kill criteria.

7 PointsAtlas can be reduced or rejected Cross-context benefit must be measured Evidence bounds maturity claims

About this Map

This Map distinguishes proof that the contract is coherent from proof that the product is useful. Fixtures, schema validation, exact processor agreement, and clean release gates can establish reproducibility. They cannot establish that a reader, author, or agent performs better.

New evidence belongs here when it can change a claim, a planned surface, or the project itself. Controls must be credible, limitations must remain visible, and a negative result is allowed to reduce or end Atlas rather than being explained away by prior investment.

Shared selected Points