# Wrecklight Salvage validation

GPT-6.2 Sol · medium reasoning · single-prompt · Superbench 2.1

Prompt SHA-256 `1553f78fe6f6676307cb78ec85c286993d84646bd5b442eebff2a81aac306ff3`

- **Passed** — CLI response transport. Exit 0; 1 completed turn(s); 1 completed response message(s); 0 error event(s); 0 malformed event line(s); raw response matches message with at most CLI terminal newline: True.
- **Passed** — Tool-free single-message protocol in the CLI event log. 1 completed response message(s). Non-message/non-reasoning item types: none. Shell, browser and web search were disabled by the launcher.
- **Passed** — Exact one-document response format. Raw response begins with the requested doctype and contains only one document with optional trailing whitespace.
- **Passed** — Frozen prompt lineage. Plan, base commit, checkout and CLI prompt receipt match byte-for-byte.
- **Passed** — HTML document structure. Parsed 1 HTML doctype(s), 1 opening and 1 closing html tag(s). Unclosed inline script: False.
- **Passed** — English document language. html lang value: 'en'.
- **Passed** — JavaScript syntax for index.html inline script 1. Node parsed the source without running it.
- **Passed** — Static external-reference scan. No remote URLs found in resource attributes or scanned literal import, fetch and CSS URL expressions.
- **Passed** — Simulation probe — sequential socket installations control the vault. With tug and cells explicitly positioned for the fixture, Interact attached each cell, one socket left Phase 1 active, and the second installed cell switched to Phase 2.
- **Passed** — Simulation probe — reattachment and reactor threshold. The same reactor attached, released and reattached. Phase 2 persisted with the tug across and the reactor only partly across. Moving the entire attached reactor across advanced to Phase 3.
- **Passed** — Simulation probe — Review loads a valid labeled fixture. Review started Phase 3 with both cells installed, the reactor attached outside the beacon, no win and a visible review flag on the mocked HUD.
- **Passed** — Simulation probe — reactor arming, paused countdown and delayed shutter closure. Positioned fixtures activated a four-second gate opening. Pause held the countdown. Occupied doorway stayed Closing, then three closed seconds began only after clearance and reopened afterwards.
- **Passed** — Simulation probe — full detached reactor settles for two seconds. Partial or attached dock placement did not win. Detaching the stationary reactor fully inside awarded victory at 240 simulation steps.
- **Passed** — Simulation probe — Reset clears Review, gate, cells and victory. Reset restored the exact initial state in the fixture harness.
- **Not observed** — Fresh headless browser usability. Chrome 156.0.8078.47 exited before navigation in task-owned headless profiles. Crashpad setsockopt and Chromium SandboxHostLinux shutdown returned Operation not permitted. A fresh Chromium headless-shell profile failed at the same shutdown call.
- **Not observed** — Offline load, resize, keyboard and pointer movement, and solid wall contact. No browser observation.
- **Not observed** — Two distinct physical cell installations and vault state transitions. No browser observation.
- **Not observed** — Attach, turn, brake, release, reattach and check the reactor threshold. No browser observation.
- **Not observed** — Beacon and repeating gate cycle, including closed blocking and safe delayed closure. No browser observation.
- **Not observed** — Reject invalid dock states and confirm detached two-second settle and visible reactor. No browser observation.
- **Not observed** — Pause, reset and deterministic Review fixture through the same gameplay logic. No browser observation.
- **Passed** — Standard packaging check and finalize. Initial check exit 0. Finalize exit 0. Registry generation was confined to model-scoped temporary staging with local coverage placeholders.

The observations above describe the original model-side checks.

## Matched release browser check

Fresh headless Chrome 156.0.8078.47 used software graphics and blocked outside requests. The final HTML was verified by SHA-256. At 1280×720 and 390×844, the scene rendered a visible canvas, fit the viewport width and produced zero console errors, page errors or outside requests. Desktop inputs exercised load review fixture, pause, reset. The sampled inputs cover part of the rubric.
