GroupieSteven/fullstackarena-gpt56sol-cross4-refresh-v432-20260909

FullStackArena cross-site refresh qualification

This gated dataset contains the complete browser evidence for the four ride_vs_active_mode cross-site instances affected by the RideApp refresh contract correction. It is a focused qualification, not a new run of all 12 cross-site instances and not a full-pack pass-rate claim.

Immutable configuration

• Model: openai/gpt-5.6-sol through OpenRouter.
• FullStackArena source: 0910cc2085b1872238dd5f302ef3305035e309a0 from a clean checkout.
• Cross-site schema/generator: 5 / 4.3.2.
• Instance-pack SHA-256: cd7891abc1270735313195096308c797c0c380bb33ba94c23d8527cab4bb82ee.
• Template-pack SHA-256: 5284a72eea0468521ca89ee8b30ba2526ad4441eee7b8ac0be781808e72ab4ee.
• Execution flags: pass@1, uncapped wait, requested include-time, headless, one worker, 900-second task timeout. Cross-site tasks intentionally use the explicit website-time tool rather than automatic elapsed-time injection.
• Deployment: one isolated Yellow cross-site lane with separately scoped Map session and RideApp PostgreSQL/Redis/account state.

Outcome and retry boundary

The original scheduled batch produced three valid strict passes (cross_website_8–10) and one infrastructure error (cross_website_11). The error occurred while opening the local Chrome CDP WebSocket, before any model call or page interaction. It is retained under release_evidence/task11_infrastructure_error and is not counted as a model failure or pass.

Only that zero-exposure infrastructure error was rerun. Its independent retry produced a valid strict pass. Therefore this release contains four valid task-level trajectories, but it does not rewrite the original batch as a 4/4 pass@1 run. The original batch and retry root manifests remain separate under release_evidence.

Every valid task has its complete trace.json, task.log, token summary, step JSON, per-step and final PNGs, readable conversations, exact model request/response records, and canonical model_calls records. The layout under valid_runs/cross_website/gpt-5.6-sol_uncapped_include-time/ follows the project's reader-facing per-task trace convention.

Refresh adjudication

All four valid trajectories have exactly one server-recorded initial Search, one server-recorded on-page Refresh view, and one paired quote_refreshed receipt. The paired payloads agree, each delay exceeds the required 15 seconds, and the deterministic evaluator agrees with the browser action, final answer, and database state. The set includes both booking and do-not-book behavior.

Two complete 12-task deterministic oracle passes are included. A separate live adversarial check retained a correct route, answer, and conditional booking decision but deliberately performed a second Search instead of Refresh. Its event counts were Search = 2, Refresh view = 0, Refresh receipt = 0; the strict evaluator rejected it and the isolated account was reset afterward.

The earlier suspected RideApp SQL failure was not reproduced. The visible Refresh control sends observation_source=refresh; the backend stages the quote_viewed and quote_refreshed rows in one SQLAlchemy session and commits them together. During qualification all five backends were healthy with zero restarts and zero OOM kills, and the relevant logs contained no SQLAlchemy, PostgreSQL, traceback, or HTTP-5xx signature. The observed non-task error was a local CDP opening-handshake timeout, not a RideApp refresh or database failure.

Integrity and interpretation

task_index.json binds every published task to its original execution, source task hash, evaluator evidence, and byte inventory. release_manifest.json hashes every published file; release_validation.json records the fail-closed audit. A pass is authoritative only when deterministic cross_site evidence scores 1.0. Browser-use LLM judgements are diagnostic and cannot override it.

The dataset contains benchmark traces, not secrets: API keys, cookies, authorization headers, raw corpus identities, and private profile data were excluded and scanned before publication. Access is manually gated.