GroupieSteven/fullstackarena-gpt56sol-representatives-20260829

FullStackArena representative trajectories

This bundle contains one pass@1 model trajectory for each of 30 RideApp and 31 Arenagram template families, plus two explicitly labelled post-fix Arenagram validation reruns. It does not claim that the model executed all 120 RideApp or all 93 Arenagram instances.

Site	Phase	Raw evaluator	Adjudicated
rideapp	original_representative_run	25/30	28/30
arenagram	original_representative_run	22/31	22/31
arenagram	post_fix_validation	2/2	2/2

RideApp's three adjudication changes are evaluator false negatives established from the original server evidence; the original model trajectories are not rewritten. Arenagram's post-fix number is 24/31: the original 22/31 remains in the bundle, while J19 and J22 appear again as separately labelled reruns.

Execution conditions

The 30 RideApp and 31 Arenagram original representative runs are equivalent to fullstackarena-publication-v1: the exact model is openai/gpt-5.6-sol, pass@1, one sequential worker, a 900-second whole-task timeout, headless execution, automatic elapsed-time injection off, and each task's reviewed step budget. Provider calls and explicit browser wait actions are uncapped. “Uncapped” does not remove the whole-task timeout and does not relax evaluators, action order, mutation constraints, or website-time windows. RideApp and Arenagram expose an explicit website-time tool on every task; calling it is optional and every call is recorded.

The two labelled post-fix Arenagram reruns used an explicit 50-step ceiling instead of the current reviewed 60-step instance budget. Both completed below that stricter ceiling. Every record contains its actual settings and a protocol relation field so this deviation cannot be mistaken for an identical run.

Why the rates can be high

These are representative-template runs on deterministic seeded/reset state, not random live-web tasks. Framework startup performs authentication and opens the configured workspace, so evaluation measures the substantive workflow. The authoritative state/event evaluators remain strict. Post-run adjudication removes demonstrated evaluator false negatives rather than forgiving agent mistakes. Several source runs record repository_dirty=true; that fact is preserved per record and must be considered when reproducing the historical raw experiment.

Privacy and auditability

trajectories.jsonl contains visible action summaries, step evaluation/goal, timing, final answer, evaluator results, and source hashes. It excludes raw chain-of-thought/model prompts, screenshots, credentials, authentication state, cookies, local paths, host addresses, and private corpus/profile data. URLs use site placeholders. FALSE_POSITIVE_AUDIT.md records the independent release gate and its bounded conclusion. manifest.json hashes every published payload file.