Validation

Verified public release — October 4, 2026: source 3641392f85e15f00b02b2049571b15c7494b8725 passed the complete verification and Pages workflow, including dependency audit and production browser checks. Clean local/public main agreed; all 10 deployed files matched the local build, GitHub artifact and live bytes. Pages was enabled before the first push. The full 15 browser journeys also passed against the public URL. Independent review replayed a bounded proceed under the original contract and an exploratory relaxed guardrail that remained blocked under original rules. No browser errors were observed.

The initial publication workflow was expanded during release review to require browser and dependency gates with pinned actions before deployment. Later documentation revisions keep the tested application behavior and receive their own release verification. Software evidence does not establish a real experiment, human comprehension or business value.

Statistical method: NIST difference-of-proportions formula and coverage cautions. Approximate 95% marginal unpooled normal interval, independent-binomial assumption, no multiplicity/sequential correction. An independently calculated Python oracle supplies fixed expected constants for meaningful unit checks. Curated quality signals are not a formal assignment-ratio test. Small-cell inference is suppressed.

Expected EV-E1 primary interval: 6.124944715481708 to 9.875055284518305 pp; complaint interval: −0.3620520004412404 to 0.7620520004412401 pp. EV-E2 complaint interval: 2.2819869639066437 to 3.718013036093357 pp.

No human evaluation was conducted. Proposed evaluation measures remain validity recognition, guardrail reasoning, uncertainty explanation, rule/verdict consistency and recovery comprehension. No commercial outcome or causal impact is claimed.

Executed software evidence — October 4, 2026

Node 24.14.1, locked dependencies. Lint, TypeScript, production build and dependency audit passed; audit reported 0 vulnerabilities. Nine unit tests passed, including fixed independent-oracle numeric constants, only EV-E1 original proceed eligibility, post-result gate separation, snapshot tampering rejection, small-cell suppression and byte/readability binding.

Fifteen production Playwright scenarios passed against the explicit local app origin and /Experiment-Verdict/ path. They cover reviewed proceed and exact JSON export; cancel/focus return; reload and append-only withdrawal; harm/inconclusive/assignment/exposure blocking; exploratory rule preservation; stale verdict/reset/refresh rejection; corrupt-byte preservation and deliberate reset; getItem and localStorage-property read failures with zero writes; write failure with in-memory notice; forward/reverse Tab scope and Escape; 320/390 px width and 1280×633 default viewport; no page errors/external requests; safe literal reviewer text; all seven public document links; explicit-zero vs blank thresholds; visible short-reason and missing-reference validation.

The agent-browser skill verified production and development entry points: meaningful scorecard content, expected controls, no error overlay or page errors. Desktop and mobile screenshots were visually reviewed for hierarchy, readable counts and overflow. Intentionally tracked PNGs in evidence/ are machine-generated local layout evidence, not human-study proof. Fresh repeat-run captures go to ignored test-results/screenshots/; they do not overwrite the published reference PNGs. Generic logs and data artifacts are excluded from source. The React checklist review covered versioned bounded local storage, lazy initial reads, derived assessment state, labels, focus scope, immutable records and component separation. No network service or runtime environment is required.

Earlier browser failures were corrected: exact label selection in the test harness and native-dialog focus wrapping. Final runs include the regression checks and passed. These software results do not establish human comprehension or completed research. All owned development/preview processes and the agent-browser session were stopped after verification.