README.md

---
configs:
- config_name: map_sep13_runs
  data_files: map_sep13/runs.jsonl
- config_name: map_sep13_tasks
  data_files: map_sep13/tasks.jsonl
- config_name: map_sep13_conversations
  data_files: map_sep13/conversations.jsonl
---

# web-benchmark · map_sep13 agent runs

Browser-agent runs on the Map Arena site, collected while building the `map_sep13` task set
(2026-09-14 to 2026-09-16). One row per run, with the task it attempted, the agent's answer, the
grader verdict, and the full step-by-step trajectory.

- Agent: browser-use, model `openai/gpt-5.6-sol` via OpenRouter, max 30 steps
- Runs: 311 (runs voided because the API ran out of credits, or because a task wording could not be graded, are excluded)
- Tasks referenced: 296
- Screenshots are **not** included.

## Files

| file | one row per | contents |
|---|---|---|
| `map_sep13/runs.jsonl` | run | task id, level, template, intent, answer, verdict, per-check results, timings, tokens, cost, and `trajectory`: every step's thinking, evaluation, memory, next goal, actions, results, URL, and site clock time |
| `map_sep13/conversations.jsonl` | run | `llm_conversation`: the full text sent to the model at each step (system prompt, agent state, browser state) |
| `map_sep13/tasks.jsonl` | task | intent, answer checks (`eval`), and task metadata |

Join the three on `task_id`; join runs and conversations on `run_id`.

## Levels

- **L1**: time-independent lookups, including Street View.
- **L2**: answer only after an event happens.
- **L3**: answer before a deadline.
- **L4**: answer within an interval after an event.

"Route A" tasks run the site clock 6x faster with whole-minute ETAs; "route B" tasks run at 1x with per-second ETAs.

| level | runs | passed | pass rate |
|---|---|---|---|
| L1 | 105 | 49 | 47% |
| L2 | 75 | 21 | 28% |
| L3 | 70 | 26 | 37% |
| L4 | 61 | 24 | 39% |

These are pass rates over the runs in this file, not over the final task set: pilots and reruns are included.

## Batches

| batch | runs |
|---|---|
| L2 pilot, first run | 14 |
| L2 pilot, clean baseline | 14 |
| L3 pilot (latest ETA) | 2 |
| L3/L4 pilot batch | 24 |
| map_sep13 half sample | 124 |
| map_sep13 remaining half | 111 |
| street-view fix rerun | 13 |
| street-view navigation-hint rerun | 9 |

## Verdicts

`passed` is the verdict recorded at run time. `passed_under_current_task_definition` re-grades the same answer
with the current task definition at the moment the agent submitted (`null` where the task was later deleted).
They differ for 3 runs:

- run 106 `map_l2_19150`: recorded fail, pass under the current definition
- run 226 `map_l1_20052`: recorded fail, pass under the current definition
- run 432 `map_l1_20017`: recorded fail, pass under the current definition

Run 106 was originally graded against a clock read after the browser closed; the other two belong to Street View
tasks whose answer checks were corrected after the run (see below).

## Known issues in the source tasks

- **Street View outside New York.** 514 of the 533 panoramas are genuine New York imagery. The rest are single
  photos taken elsewhere (Indonesia, Copenhagen, an Alpine forest) placed at US city coordinates. Tasks that required
  the city name were changed on 2026-09-16 to grade only the photo content, and tasks that said "a Chicago dot" were
  pinned to one panorama's coordinates.
- **Brief events.** Some route-B gap/lead tasks have a condition that holds for only a few seconds before reverting.
  They were kept deliberately (the page re-prices every second, so the condition is visible), but failures on them
  partly reflect chance. Their task metadata carries `brief_event`.
- **Tight L3 windows.** In several L3 "latest value" tasks the gap between the scheduled ETA change and the deadline
  is shorter than the agent's median read-then-submit time (about 17 s).
- Six passed L2 pilot tasks were deleted from the task set; their `tasks.jsonl` rows are recovered from the run
  conversations and have no answer checks.