# Xbox — independent CARBON workflow

Local Codex plugin for Jason's Xbox sales discussion. Xbox is the platform under assessment; CARBON supplies the assessment workflow and report. No Microsoft endorsement is implied.

## Use

Requires Codex with supported browser/computer controls, an authorized Xbox session, appropriate existing game entitlement and supported input. Node 20+ runs the evidence utilities; existing FFmpeg encodes captures. No dependencies, API keys, simulators, extensions or game binaries are installed by this package.

Install the local plugin folder using the Codex app's local-plugin interface if available. The manifest is `.codex-plugin/plugin.json`, with `skills/` matching the installed CARBON Games format. This package has not been installed into the host app automatically. If that app build has no local-plugin picker, open this directory in Codex and explicitly invoke `skills/xbox/SKILL.md`; do not edit host registries by guesswork.

1. Open official https://www.xbox.com/play in an identified supported browser. Search the live catalog. Verify account identity without copying credentials/profile data.
2. Check game entitlement, single-player mode, and input requirements. Never buy, change subscription, bypass controls or assume catalog inclusion grants access.
3. Capture the game tab and save dated observations, decisions and outcomes. Stop recording during sign-in and profile inspection. Use fresh screenshots for video-stream controls.
4. Run First Timer, then Casual Player/Settings Explorer only if executable. Bound active play to a few minutes and exit the stream normally. Mark inaccessible scenarios blocked.
5. Append observations to `evidence/observations.json`. Run `npm run learn` to derive local policy memory. This is evidence-based adaptation, not model training. The Codex host model interprets observations; no separate provider or credential access is implemented.
6. Run `npm test` and `npm run validate`. Review report, representative footage and complete video decode. Package public/redacted evidence only.

## Architecture and provenance

Adapted concepts from installed CARBON Games 0.4.13: learn → plan → execute → report, persona-scoped context, evidence persistence, and honest video timing. Reviewed Netflix JEV live report: streamed gameplay can lack semantic DOM; its calibrated Tetris reader must not be reused for Xbox. No Netflix runtime or trained policy is copied.

The unchanged CARBON wordmark and Hero Cut C are copied from CARBON Games assets. Charcoal #101314, panel #171b1c, mint #b9dac9 and text #eff3ed follow its `ui/BRAND.md`. Existing projects remain untouched.

`runtime/policy.mjs` learns only provenance-bearing facts/transitions and proposes controls after safety/freshness checks. It never executes input. `runtime/replay.mjs` deterministically rebuilds memory. Tests use synthetic inputs and are not game results.

## Evidence limits

The first capture contains 24 supported browser screenshots over 10.291 seconds of catalog navigation, with original timestamps; no audio. A failed CDP screencast collected zero frames and is recorded in `video/capture.json`. Never describe catalog screenshots, trailers, or synthetic tests as gameplay. See the report for the actual final run status.

## Completed run

Planet of Lana II 1.1.10 was launched and played in Chrome on 23 October 2026 after verified entitlement. First Timer discovered menus and D/Space movement; Settings Explorer inspected controls without changes; Casual Player reused D and tested A. These were sequential objectives in one session, not independent benchmark runs. The stream was returned to the main menu and its tab closed.

Open `report/index.html` in the review bundle for the report with linked videos. The source plugin package includes the dated ledger and reviewed screenshots; video source samples are retained only in the original local project. It supports `npm test`, `npm run learn`, and structural `npm run validate` without the optional report. Full media reproduction requires the original captures, Python with Pillow, and FFmpeg. Run `python3 runtime/encode.py`, `node runtime/demo-decisions.mjs`, `python3 runtime/report.py`, then `python3 runtime/package.py`.

`policy-replay.json` demonstrates proposals against recorded observations using their original timestamp. It is an offline replay, not additional gameplay. The host model selected the actual live actions. No external API keys were read or created.