FullStackArena three-site table study (private working dataset)
Protocol fullstackarena-three-site-table-study-v1 (status qualified): five exact OpenRouter models × three cells × the signed Arenagram (124), RideApp (120) and Auction (100) task packs = 45 arms, 5,160 scored attempts, max_steps=30, five workers on five isolated lanes per batch.
Cells: table1_no_time_capped, table1_no_time_uncapped (no automatic time; the explicit get_website_time() tool stays available), table2_time_uncapped (automatic elapsed time, previous-step duration and assigned-lane website time). Table 2 is compared with Table 1 uncapped per model and task.
Layout
protocol/: the pinned spec, source lock, seeded run order, launch and adjudication protocols, and the tracker of all 45 arms.runs/<site>/<model>/<condition>/: one complete arm covering the whole signed pack of that site. summary.json (pass/fail counts, pass rate, pack identity, provenance), attempts.jsonl (one row per task in pack order with the outcome, the strict evaluator details, tokens, time-tool usage and the review note), adjudication_ledger.json (task-by-task review evidence with the outcome and reason), failure_analysis.json and failure_analysis.md (pass rate per task type and the primary failure reason of every failed attempt, overall and per task type), and raw_artifacts.tar.gz (every trace, step, recorded model request/response, screenshot, worker manifest and log).
Outcome rule
Every attempt is pass or fail. A pass is a strict site-evaluator pass, or a strict fail that the reviewer confirmed to be an evaluator false negative: the postcondition and server receipts are satisfied, every reported value equals the authoritative value, the task prescribed no answer format that the answer violated, and the evaluator's rejection was reproduced offline with a content-identical witness (accepted) and a wrong-value counterexample (rejected). Every other attempt is a fail. outcome_basis names which of the two applies to a row. Every template family has at least one attempt reviewed in detail.
Arms published
| path |
tasks |
pass |
fail |
pass rate |
runs/arenagram/openai__gpt-5.6-sol/table2_time_uncapped |
124 |
31 |
93 |
25.0 % |
runs/auction/openai__gpt-5.6-sol/table2_time_uncapped |
100 |
33 |
67 |
33.0 % |
runs/rideapp/openai__gpt-5.6-sol/table1_no_time_uncapped |
120 |
38 |
82 |
31.7 % |
runs/rideapp/openai__gpt-5.6-sol/table2_time_uncapped |
120 |
26 |
94 |
21.7 % |
SUMMARY.md is the cross-site summary: pass rate by site and cell, by task type, the failure reasons of every arm, and the stage results that are not published arms.
Pass rate per task type and failure reasons
Every failed attempt carries exactly one primary reason: the earliest requirement of the task that the attempt violated, read from the site's server-side evidence and the recorded trajectory. A passed attempt carries none. The reviewer assigns the reason while adjudicating the attempt; the vocabulary is closed. Counts are attempts. Each arm's folder holds the same tables (failure_analysis.md) and the task ids behind every count (failure_analysis.json); every row of attempts.jsonl carries failure_category.
runs/arenagram/openai__gpt-5.6-sol/table2_time_uncapped
| Task type |
Tasks |
Pass |
Fail |
Pass rate |
| Type 1 · no time condition |
12 |
7 |
5 |
58.3 % |
| Type 2 · relative wait |
36 |
3 |
33 |
8.3 % |
| Type 3 · absolute deadline |
36 |
11 |
25 |
30.6 % |
| Type 4 · absolute window |
40 |
10 |
30 |
25.0 % |
| All |
124 |
31 |
93 |
25.0 % |
| Failure reason |
Kind |
Type 1 |
Type 2 |
Type 3 |
Type 4 |
All fails |
Share |
| Right action, completed after the deadline or window |
Time requirement |
0 |
13 |
10 |
8 |
31 |
33.3 % |
| Ran out of time before completing the required action |
Time requirement |
0 |
0 |
5 |
13 |
18 |
19.4 % |
| Missed an earlier timed step |
Time requirement |
0 |
20 |
0 |
4 |
24 |
25.8 % |
| Reported a wrong or ill-formed answer |
Choice or answer |
0 |
0 |
1 |
0 |
1 |
1.1 % |
| Could not operate the page correctly |
Operating the page |
5 |
0 |
9 |
5 |
19 |
20.4 % |
| All fails |
|
5 |
33 |
25 |
30 |
93 |
100.0 % |
runs/auction/openai__gpt-5.6-sol/table2_time_uncapped
| Task type |
Tasks |
Pass |
Fail |
Pass rate |
| Type 1 · no time condition |
8 |
8 |
0 |
100.0 % |
| Type 2 · relative wait |
28 |
9 |
19 |
32.1 % |
| Type 3 · absolute deadline |
32 |
4 |
28 |
12.5 % |
| Type 4 · absolute window |
32 |
12 |
20 |
37.5 % |
| All |
100 |
33 |
67 |
33.0 % |
| Failure reason |
Kind |
Type 1 |
Type 2 |
Type 3 |
Type 4 |
All fails |
Share |
| Right action, completed after the deadline or window |
Time requirement |
0 |
0 |
7 |
4 |
11 |
16.4 % |
| Ran out of time before completing the required action |
Time requirement |
0 |
0 |
20 |
10 |
30 |
44.8 % |
| Missed an earlier timed step |
Time requirement |
0 |
0 |
0 |
2 |
2 |
3.0 % |
| Acted on information read earlier without re-checking |
Choice or answer |
0 |
19 |
0 |
0 |
19 |
28.4 % |
| Chose the wrong item, branch or value |
Choice or answer |
0 |
0 |
1 |
4 |
5 |
7.5 % |
| All fails |
|
0 |
19 |
28 |
20 |
67 |
100.0 % |
runs/rideapp/openai__gpt-5.6-sol/table1_no_time_uncapped
| Task type |
Tasks |
Pass |
Fail |
Pass rate |
| Type 1 · no time condition |
12 |
12 |
0 |
100.0 % |
| Type 2 · relative wait |
36 |
6 |
30 |
16.7 % |
| Type 3 · absolute deadline |
36 |
4 |
32 |
11.1 % |
| Type 4 · absolute window |
36 |
16 |
20 |
44.4 % |
| All |
120 |
38 |
82 |
31.7 % |
| Failure reason |
Kind |
Type 1 |
Type 2 |
Type 3 |
Type 4 |
All fails |
Share |
| Right action, completed after the deadline or window |
Time requirement |
0 |
16 |
21 |
9 |
46 |
56.1 % |
| Ran out of time before completing the required action |
Time requirement |
0 |
3 |
11 |
6 |
20 |
24.4 % |
| Missed an earlier timed step |
Time requirement |
0 |
11 |
0 |
5 |
16 |
19.5 % |
| All fails |
|
0 |
30 |
32 |
20 |
82 |
100.0 % |
runs/rideapp/openai__gpt-5.6-sol/table2_time_uncapped
| Task type |
Tasks |
Pass |
Fail |
Pass rate |
| Type 1 · no time condition |
12 |
12 |
0 |
100.0 % |
| Type 2 · relative wait |
36 |
2 |
34 |
5.6 % |
| Type 3 · absolute deadline |
36 |
1 |
35 |
2.8 % |
| Type 4 · absolute window |
36 |
11 |
25 |
30.6 % |
| All |
120 |
26 |
94 |
21.7 % |
| Failure reason |
Kind |
Type 1 |
Type 2 |
Type 3 |
Type 4 |
All fails |
Share |
| Right action, completed after the deadline or window |
Time requirement |
0 |
15 |
11 |
4 |
30 |
31.9 % |
| Ran out of time before completing the required action |
Time requirement |
0 |
6 |
23 |
13 |
42 |
44.7 % |
| Missed an earlier timed step |
Time requirement |
0 |
12 |
0 |
7 |
19 |
20.2 % |
| Misread the time requirement |
Time requirement |
0 |
0 |
1 |
0 |
1 |
1.1 % |
| Could not operate the page correctly |
Operating the page |
0 |
0 |
0 |
1 |
1 |
1.1 % |
| The page never showed the needed result |
Operating the page |
0 |
1 |
0 |
0 |
1 |
1.1 % |
| All fails |
|
0 |
34 |
35 |
25 |
94 |
100.0 % |
Failure reason definitions
| Failure reason |
Kind |
Definition |
Right action, completed after the deadline or window (late_completion) |
Time requirement |
The required final action, answer or submission was performed and was otherwise correct or complete, but the site recorded it after the deadline or outside the action window. |
Ran out of time before completing the required action (incomplete_in_time) |
Time requirement |
When the deadline or window closed, the required final action had not been performed or only part of it had; this includes attempts that stopped at the bound and declined to act late. |
Missed an earlier timed step (missed_timed_step) |
Time requirement |
A timed step that precedes the final action was missed: a first look or first retrieval deadline, a required dwell, an observation or preview window, a required wait or a status check. |
Misread the time requirement (misread_time_rule) |
Time requirement |
The attempt acted on a wrong reading of the stated time rule, for example treating a deadline as a start time. |
Acted on information read earlier without re-checking (stale_information) |
Choice or answer |
The choice was made from state read well before the action while the site's state changed in between, and the attempt did not look again at the required moment. |
Chose the wrong item, branch or value (wrong_selection) |
Choice or answer |
With current information available, the attempt selected the wrong item, took the wrong branch of a conditional task, or used a wrong value. |
Reported a wrong or ill-formed answer (wrong_answer) |
Choice or answer |
The site actions were acceptable, but the reported answer was not the requested content. |
Could not operate the page correctly (page_operation) |
Operating the page |
The attempt used a control in a way that did not do what the task needed: a reload that never reloaded, a required selector left unset, text typed into the wrong field, an action that never returned. |
The page never showed the needed result (result_not_shown) |
Operating the page |
The attempt performed the required steps in time, but its browser never displayed the result it had to read, so no answer could be given. |
Last update: 2026-09-22T03:56:53.457102+00:00. Source commit in protocol/source_commit.txt. Sites and images are the live Penn release recorded in protocol/table_study_release_lock.json.
