[Air000000]

Problem

After PR78 merged, the root README, evaluation overview, project summary and demo guide still described v5.2 as never model-evaluated/only synthetic. This confused the zero-call mechanism demo with the separately completed real-model experiments.

Changes

• Correct those four existing entry points only; no new document/framework or interview/coaching material.
• Link frozen v5.2 fictional controls, v5.2.1 contemporaneous format regression, natural v5.1/v5.2.1 diagnosis and v5.2.1/v5.3 observed regression.
• Preserve the actual boundaries: controls are not natural accuracy; repeated controls did not demonstrate a format fix; latest binary proxy agreement22/25→23/25, two refusal fixes/one new refusal, same insufficient admission1/4; five ambiguous targets excluded from binary scoring.
• Explain that the latest contemporaneous baseline22/25 must not be replaced by historical21/25. AI proxy/observed development is not independent expert validation, generated-answer quality or online evidence.
• Close further prompt iteration/repeated paid calls on these same30 under the completed scope. Future new-sample validation must define/freeze scope and obtain separate data/cost authority.

Verification

• 313 offline tests passed with external-network guard and temporary DB.
• Checked all80 local Markdown links in the four documents; latest figures agree with the frozen public result JSON.
• Existing zero-call synthetic demo passed; Ruff and diff checks passed.
• Four Markdown files only,49 insertions/9 deletions; no runtime, dependency, prompt, validator, dataset, frozen report/contract or CI edits.
• No provider calls, new fees, DEV, relabeling or serving changes. Primary dirty checkout/.env/runtime data untouched.