Evidence edition / 23 October 2026 / en-US

Methodology and limits

Bounded, risk-based public evaluation

Priority went to the first useful outcome: discover a game, understand input/access requirements, and recover from a dead end while signed out. Seven route identities, two collection routes, three game details, normal search values, and three viewport sizes formed the sample. One browser owner executed sequentially at human-scale pace; one separate agent reviewed planning and authored report CSS without interacting with the target.

Stop conditions: account or age-verification gates, purchase/submission actions, warnings, bot/security controls, shared-browser ownership, or unsupported host operations. No authentication, age-gate acceptance, gameplay, private API probing, downloads, purchases, stress, fuzzing, exploit attempts or Microsoft contact occurred.

Evidence rules

Passed means the case’s narrow expected behavior was observed. Failed means the stated oracle was not met. Observed means a fact needing interpretation. Partial means only part of the planned behavior was exercised. Blocked/planned/not-offered never count as passed. Repeated trials are not extra distinct checks. Screenshots and DOM were captured sequentially and may differ briefly during transitions.

Toolchain and limitations

CARBON Studio package 0.1.6+codex.20261015. The existing carbon-runtime framework prepare command generated 110 candidates. The native MCP progress, Studio, preferences, intelligence and execution-plan tools were unavailable. Run initialization wrote local partial state but failed when attempting global promotion history outside the writable workspace; no retry of that global write. Native command-report rendering uses the supported per-process CARBON_MUTE_ADS option so report generation does not need that unrelated write.

Browser: Codex task-scoped in-app browser through cua_repl; no standalone Playwright/Selenium, no shared desktop input. Browser engine version could not be retrieved: the host does not support Browser.getVersion through its CDP interface. Session UI was visibly signed out. Underlying cookie isolation was not independently audited. Viewport settings were reset after testing. Network/cache were not controlled.

Performance

No performance benchmark was conducted. Content populated asynchronously and a gallery error recovered, but neither screenshot spacing nor tool-call durations measure page performance. No percentile, Core Web Vital, loading-time threshold or service reliability claim is made.

Eight evidence-qualified confidence dimensions

These are qualitative concern summaries, not claimed runtime quality-confidence scores. Overall product/release confidence: NOT ASSESSED. No absence-of-bugs argument is used.

Next evidence with the most value

1. Continuous keyboard and real-screen-reader journey around modal focus.
2. Confirm intended ownership/subscription rules behind 1000xRESIST messaging.
3. Physical mobile comparison scrolling, zoom, contrast and touch targets.
4. Second engine and controlled performance protocol if separately scoped.

Reproducibility

Use cases.csv with evidence/manifest.json. Public catalog content is dynamic and locale-dependent; observations are dated 23 October 2026. Prices are screenshot observations, not a price promise or purchase recommendation. The package includes public UI images solely to support independent commentary; Microsoft and game marks remain their owners’ property.