junlinw/autoresearch-solo-vs-forum

Solo vs forum: long-horizon coding-agent runs on 12 research-engineering tasks

1440 runs (206 solo, 1225 forum, 9 long-run chains), 12276 graded agent rows, 11081 transcripts, 304444 forum posts; 36457 files, 49.3 GB. Models: deepseek-v4.1-flash, glm-5.2, gpt-5.6-sol, qwen3.8-27b. Tasks: actlearn, adplace, borden, carleson, dabic, exploit, graph, kda, mega, moe, swinmlp, topopt.

What the experiment is

Each agent is a coding-agent CLI (Claude Code for Qwen3.8-27B / DeepSeek-V4.1-Flash / GLM-5.2 through an Anthropic-API shim to a self-hosted sglang endpoint; OpenAI Codex CLI for GPT-5.6-sol) dropped into a sandbox with a research-engineering task, a dev judge it can call as often as it likes (submissions.jsonl is its score-over-time), and a fixed wall-clock budget (12 h, 6 h for exploit). At the deadline the final submission is regraded once by the same judge; that grade is the row in results.jsonl.

• solo — independent agents. Every agent in a solo run dir is an i.i.d. sample; solo best-of-N curves are computed by pooling all solo agents of a (model, task) cell (manifest/agents.jsonl, arm == "solo") and taking the exact expected best of k.
• forum — N agents on the same task share a forum: a titled-post bulletin board they can list, read and post to at will (no digests pushed; "nudge off"). Nothing else connects them. A forum run's headline value is the best of its N agents. N is the team size (2 to 100); a cell has several independent runs (rep). The forum-64/100 runs (campaign == "qwen-forumlarge"), the forum-5/50 runs (campaign == "qwen-f5f50") and the DeepSeek swinmlp forum-10 runs ran a harness whose list shows the newest 100 titles plus every unread post; every other forum listed all titles. It only binds once a forum passes 100 posts. forum-64/100 and f5f50 members were spawned 6 s apart, and f5f50 packed 12 agents per inference engine instead of the usual 10.
• longrun — ONE Qwen3.8-27B agent per task for 120 h of agent time. A sandbox lives at most 24 h, so the run is a chain of six 20 h segments: each segment snapshots the sandbox filesystem at teardown, and the next one opens on that snapshot and resumes the same Claude Code session (--continue) — same working tree, same conversation (subject to the CLI's own context compaction). Every segment ends with a regrade of the current submission, so results.jsonl holds one row per segment: a score-vs-agent-hours curve for a single agent. The agent is never told its deadline; the harness re-engages it whenever a turn ends. Compare it with the 12 h solo pool of the same task.

Layout

manifest/runs.jsonl|csv     one row per run: model, task, arm, N, rep, campaign, created_at, n_rows, n_good, best/mean score,
                            usable (the report's selection rule), void_reason, n_forum_posts, path, files
                            (long-run chains also: segments_done, complete, agent_hours, final_score; runs.jsonl adds
                            segments[] and notes[])
manifest/agents.jsonl|csv   one row per graded agent trial (a long-run chain: one per segment): final score/reward/status,
                            tokens, cost, submission counts, transcript stats, paths
manifest/cells.csv          runs/agents per (model, task, arm, N)
manifest/tasks.json         metric field, scale and direction per task
manifest/scrub_report.json  how many credential-shaped strings / endpoint hostnames were redacted
data/<model>/<task>/<arm>/<run_id>/
    run.json                sanitized run config + per-agent stats
    results.jsonl           final graded rows (one per agent)   <- START HERE
    submissions.jsonl       every dev-judge call, all agents (agent_id, ts, score, valid, summary, snapshot)
    judge_history.jsonl.gz  the judge's per-submission verdicts and sub-metrics (gzipped; borden's run to 2 MB/agent)
    usage_timeline.jsonl    per-turn cumulative tokens
    forum_posts.jsonl       (forum runs) the whole forum: id, ts, author, title, body, best_score at post time
    forum_activity.jsonl    (forum runs) each agent's list/read/post calls with timestamps
    transcripts/<agent_id>.jsonl.gz    the agent's full CLI transcript (see below)
    snapshots/<agent_id>__best.md      code snapshot of the agent's best valid submission (text dump of the submitted files)
    snapshots/<agent_id>__last.md      code snapshot of its last submission
data/qwen3.8-27b/<task>/longrun/longrun_<task>/
    results.jsonl           one row per segment: segment, agent_hours_start/end, status, score/reward, usage
    submissions.jsonl, judge_history.jsonl.gz, usage_timeline.jsonl    as above, every row tagged with its `segment`
    transcripts/s01.jsonl.gz ... s06.jsonl.gz    one per segment; concatenated in order they are the whole 120 h session
    snapshots/sNN__best.md, sNN__last.md         per segment
    run.json                segments[] (per-segment summary) and notes[] (caveats specific to this chain)

agent_id is agent_k in forum runs, repk in solo runs and agent in long-run chains (whose rows are keyed by segment); it is unique within a run only. Join on (run_id, agent_id), plus segment for long-run rows.

Transcript format

Claude Code runs (cli == "claude"): each line is one Claude Code stream-json event with an added sequence number i. Kept: assistant messages (content blocks thinking / text / tool_use, with per-message usage), user messages (the tool_result blocks — command output, file reads — plus tool_use_result stdout/stderr and a timestamp), result (end of each CLI session with cumulative usage/cost; a run has several because the harness resumes the CLI), and three system events: init (model, CLI version; one per CLI resume), compact_boundary (an automatic context compaction happened: pre/post tokens), status: compacting. Dropped as noise: system/thinking_tokens heartbeats, tool_progress, background-task bookkeeping. In user events the harness's raw tool_use_result copy (a duplicate of the tool_result block the model saw, ~55% of a transcript) keeps only its scalar metadata: every string over 200 chars is replaced by its length. Strings over 200k characters anywhere are truncated with a marker. judge_history.jsonl strings are capped at 4,000 chars.

Codex runs (cli == "codex"): each line is one Codex CLI JSON event: item.completed for agent_message, command_execution (command + aggregated_output + exit_code), file_change, reasoning, todo_list; item.updated for todo lists; thread.started / turn.started; and the CLI's own ERROR/WARN log lines as cli_log. Codex events carry no timestamps; use submissions.jsonl for the time axis.

Downloading a subset

Transcripts are most of the bytes. To get everything except them:

from huggingface_hub import snapshot_download
snapshot_download("junlinw/autoresearch-solo-vs-forum", repo_type="dataset", ignore_patterns=["*/transcripts/*"])
# one cell only:
snapshot_download("junlinw/autoresearch-solo-vs-forum", repo_type="dataset", allow_patterns=["manifest/*", "data/qwen3.8-27b/graph/*"])
# the long-run chains:
snapshot_download("junlinw/autoresearch-solo-vs-forum", repo_type="dataset", allow_patterns=["manifest/*", "data/*/*/longrun/*"])

Selection rules used by the published report

• A row is a real sample when status is success or timeout (good == true). status == "error" is an infrastructure death. A good row with no valid submission scores 0.0 — keep it, dropping it biases best-of-N upward.
• A forum run counts as a forum-N sample only if at least half its cohort is good (usable).
• void_reason marks whole run dirs the report excludes (judge-void carleson batch, a judge-hardware-void kda cohort, a turn-capped topopt dir, a 12 h exploit solo dir, and DeepSeek runs whose source was deleted and re-run after the first release). implausible_score marks kernel rows above a physical ceiling (timing exploits).
• borden's evaluator awards up to 109.0 points, above the nominal 0–100 EdgeBench scale. 109.0 is that maximum and a real result, not a grader exploit; earlier versions of this card said otherwise. Every model reaches it, and a whole forum can converge on it (Qwen forum-64 rep 1: 60 of its 62 graded members finished at 109.0).
• graph: exact 100.0 scores appear only in three large Qwen forums (forum-64 rep 1 and both forum-50 runs); every other graph run tops out at 95.9. They have not been audited yet.
• moe and actlearn (MLS-Bench) have no score: reward (0–1) is the metric, and they have no submissions.jsonl.
• kda/mega/swinmlp (KernelBench) are inverse latency / speedup on their own ~0–17 scale; topopt is a stiffness ratio.
• Turn cap. Most solo trials ran with the harness's auto-resume capped at 60 CLI turns; forum members had 600. The cap binds only where turns are short: Qwen solo agents averaged 6.9 h on moe (86% of them capped), 9.3 h on actlearn (59%), 10.9 h on dabic (21%) and 11.2 h on kda (28%) of a 12 h budget; on the other tasks 0–16% hit it. The qwen-solofill batch and the long-run chains ran with the higher cap. Check usage.num_resume_turns in results.jsonl (59 = capped) and duration_s.
• Solo pools mix launch batches (e.g. Qwen edgebench solo = 6 dirs launched over a month, at different inference-server loads). campaign and created_at let you split by batch; qwen-solofill is the last 1–4 agents per task that topped the Qwen solo pools up to ~100 on 2026-09-22; glm-v1 solo curves ran on cache-blind routing and were superseded by glm-v2.

Long-run chains

Grade of the agent's current submission at the end of each 20 h segment (results.jsonl; the dev-judge history in submissions.jsonl is finer-grained). … = segment still running at export time; – = no valid submission at that point (0 under the selection rules). agent-hours = the sum of per-segment duration_s, which includes the end-of-segment grade (~1 h on moe, minutes elsewhere).

task	metric	segments	agent-hours	s01	s02	s03	s04	s05	s06
borden	score	6/6	120	30.00	30.00	30.00	30.00	30.00	30.00
carleson	score	6/6	120	–	26.07	44.74	50.57	55.66	62.58
exploit	score	6/6	86	2.00	6.00	6.00†	6.00†	6.00†	6.00†
graph	score	6/6	120	52.58	54.89	54.86	54.88	54.88	54.89
kda	score	6/6	120	10.36	10.07	10.28	10.00	10.05	10.16
mega	score	6/6	120	0.00	1.31	1.30	1.30	1.31	error
moe	reward	6/6	123	0.630	0.703	0.703	0.703	0.703	0.703
swinmlp	score	6/6	120	3.50	3.66	4.02	2.88	3.42	3.65
topopt	score	6/6	120	11.45	11.46	11.47	11.50	11.50	11.50

† the segment ended early at the 600-turn guard (see its note).

• carleson: The first attempt at segments 3-6 (2026-09-19 to 09-20) wedged: the resumed session's tail after its last context compaction no longer fit the 262k-token window, every resume failed at once until the 600-turn guard ended the segment (~5 h each), and all four re-graded segment 2's artifact at 27.33. On 2026-09-21 the session was cut back to its last compaction summary (post-boundary tail 1,318,720 -> 32,426 bytes; working tree untouched) and segments 3-6 re-ran from there. The wedged attempt is not in this dataset. s02's transcript still holds the stretch that was cut, so the agent entered s03 without memory of the end of s02.
• exploit: s03 ended at the harness's 600-turn guard after 17.6 h of its 20 h budget.
• exploit: s04 ended at the harness's 600-turn guard after 9.5 h of its 20 h budget.
• exploit: s05 ended at the harness's 600-turn guard after 9.0 h of its 20 h budget.
• exploit: s06 ended at the harness's 600-turn guard after 9.5 h of its 20 h budget.
• mega: s06: the end-of-segment grade failed (status=error, scored 0) — Could not stop previous GPU evaluation: [Errno -2] Name or service not known. An infrastructure failure, not an agent result.
• moe: s06 recorded no token usage (the CLI was killed mid-turn at the budget); the transcript is complete.
• swinmlp: Segment 3's first launch (2026-09-19 22:41 UTC) died after 5 minutes with an infrastructure error and recorded no snapshot, so the chain halted; segment 3 was relaunched from segment 2's snapshot on 2026-09-21 06:41 UTC and resumed the same CLI session. The ~32 h between segments 2 and 3 is idle wall-clock, not agent time. The failed attempt is not in this dataset.

What is NOT here

• exploit (an exploitbench capability-ladder task on a browser CVE): final scores only. No transcripts, submissions, snapshots or forum posts are published for it.
• Supervisory-injection arms, leaderboard/board arms, standup, pool and orchestrator (planner + workers) modes: different designs, not solo-vs-forum. Retired long-run attempts (see the chain notes) and 1 h mechanics smokes.
• Raw agent_output.log files (the unfiltered stream, 50–650 MB per agent), the shim request logs, the per-submission .tar.gz code archives and the sandboxes' final artifact tarballs.
• Endpoint hostnames, orchestrator call ids and anything credential-shaped were redacted (manifest/scrub_report.json); the only real hits were ANTHROPIC_SHIM_UPSTREAM_AUTH=Bearer … lines in a few agents' env dumps.
• Code snapshots over 3 MB (125) were skipped; 14 agent rows have no transcript (all 14 are status == "error" trials whose sandbox died before writing one).

Cells

model	task	arm	N	runs	agents	usable runs
deepseek-v4.1-flash	borden	forum	10	10	100	10
deepseek-v4.1-flash	borden	solo	1	10	100	10
deepseek-v4.1-flash	carleson	forum	10	10	100	10
deepseek-v4.1-flash	carleson	solo	1	10	100	10
deepseek-v4.1-flash	exploit	forum	10	10	100	10
deepseek-v4.1-flash	exploit	solo	1	10	100	10
deepseek-v4.1-flash	graph	forum	10	11	110	10
deepseek-v4.1-flash	graph	solo	1	10	100	10
deepseek-v4.1-flash	kda	forum	10	10	100	10
deepseek-v4.1-flash	kda	solo	1	12	118	10
deepseek-v4.1-flash	moe	forum	10	10	100	10
deepseek-v4.1-flash	moe	solo	1	11	110	10
deepseek-v4.1-flash	swinmlp	forum	10	10	100	10
deepseek-v4.1-flash	swinmlp	solo	1	10	100	10
glm-5.2	actlearn	forum	2	5	10	5
glm-5.2	actlearn	forum	3	5	15	5
glm-5.2	actlearn	forum	4	5	20	5
glm-5.2	actlearn	forum	5	5	25	5
glm-5.2	actlearn	forum	6	5	30	5
glm-5.2	actlearn	forum	7	5	35	5
glm-5.2	actlearn	forum	8	5	40	5
glm-5.2	actlearn	forum	9	5	45	5
glm-5.2	actlearn	forum	10	5	50	5
glm-5.2	actlearn	solo	1	2	16	2
glm-5.2	adplace	forum	2	5	10	5
glm-5.2	adplace	forum	3	5	15	5
glm-5.2	adplace	forum	4	5	20	5
glm-5.2	adplace	forum	5	5	25	5
glm-5.2	adplace	forum	6	5	30	5
glm-5.2	adplace	forum	7	5	35	5
glm-5.2	adplace	forum	8	5	40	5
glm-5.2	adplace	forum	9	5	45	5
glm-5.2	adplace	forum	10	5	50	5
glm-5.2	adplace	solo	1	2	16	2
glm-5.2	borden	forum	2	5	10	5
glm-5.2	borden	forum	3	5	15	5
glm-5.2	borden	forum	4	4	16	4
glm-5.2	borden	forum	5	5	25	5
glm-5.2	borden	forum	6	5	30	5
glm-5.2	borden	forum	7	5	35	5
glm-5.2	borden	forum	8	5	40	5
glm-5.2	borden	forum	9	5	45	5
glm-5.2	borden	forum	10	5	50	5
glm-5.2	borden	solo	1	5	26	5
glm-5.2	carleson	forum	2	5	10	5
glm-5.2	carleson	forum	3	5	15	5
glm-5.2	carleson	forum	4	5	20	5
glm-5.2	carleson	forum	5	5	25	5
glm-5.2	carleson	forum	6	5	30	5
glm-5.2	carleson	forum	7	5	35	5
glm-5.2	carleson	forum	8	5	40	5
glm-5.2	carleson	forum	9	5	45	5
glm-5.2	carleson	forum	10	5	50	5
glm-5.2	carleson	solo	1	2	16	2
glm-5.2	dabic	forum	2	5	10	5
glm-5.2	dabic	forum	3	5	15	5
glm-5.2	dabic	forum	4	4	16	4
glm-5.2	dabic	forum	5	5	25	5
glm-5.2	dabic	forum	6	5	30	5
glm-5.2	dabic	forum	7	5	35	5
glm-5.2	dabic	forum	8	5	40	5
glm-5.2	dabic	forum	9	5	45	5
glm-5.2	dabic	forum	10	5	50	5
glm-5.2	dabic	solo	1	6	29	6
glm-5.2	graph	forum	2	3	6	3
glm-5.2	graph	forum	3	3	9	3
glm-5.2	graph	forum	4	2	8	2
glm-5.2	graph	forum	5	3	15	3
glm-5.2	graph	forum	6	4	24	4
glm-5.2	graph	forum	7	3	21	3
glm-5.2	graph	forum	8	3	24	3
glm-5.2	graph	forum	10	1	10	1
glm-5.2	graph	solo	1	5	26	5
glm-5.2	moe	forum	2	5	10	5
glm-5.2	moe	forum	3	5	15	5
glm-5.2	moe	forum	4	5	20	5
glm-5.2	moe	forum	5	5	25	5
glm-5.2	moe	forum	6	5	30	5
glm-5.2	moe	forum	7	5	35	5
glm-5.2	moe	forum	8	5	40	5
glm-5.2	moe	forum	9	5	45	5
glm-5.2	moe	forum	10	5	50	5
glm-5.2	moe	solo	1	2	16	2
gpt-5.6-sol	actlearn	solo	1	1	5	1
gpt-5.6-sol	adplace	solo	1	1	5	1
gpt-5.6-sol	borden	solo	1	2	10	2
gpt-5.6-sol	carleson	solo	1	1	5	1
gpt-5.6-sol	dabic	solo	1	2	10	2
gpt-5.6-sol	graph	solo	1	1	5	1
gpt-5.6-sol	kda	solo	1	2	5	2
gpt-5.6-sol	mega	solo	1	1	1	1
gpt-5.6-sol	moe	solo	1	1	5	1
gpt-5.6-sol	swinmlp	solo	1	1	1	1
qwen3.8-27b	actlearn	forum	2	11	22	10
qwen3.8-27b	actlearn	forum	3	10	30	10
qwen3.8-27b	actlearn	forum	4	10	40	10
qwen3.8-27b	actlearn	forum	5	10	50	10
qwen3.8-27b	actlearn	forum	6	10	60	10
qwen3.8-27b	actlearn	forum	7	10	70	10
qwen3.8-27b	actlearn	forum	8	10	80	10
qwen3.8-27b	actlearn	forum	9	10	90	10
qwen3.8-27b	actlearn	forum	10	10	100	10
qwen3.8-27b	actlearn	solo	1	6	100	6
qwen3.8-27b	adplace	forum	2	10	20	10
qwen3.8-27b	adplace	forum	3	10	30	10
qwen3.8-27b	adplace	forum	4	10	40	10
qwen3.8-27b	adplace	forum	5	10	50	10
qwen3.8-27b	adplace	forum	6	10	60	10
qwen3.8-27b	adplace	forum	7	10	70	10
qwen3.8-27b	adplace	forum	8	10	80	10
qwen3.8-27b	adplace	forum	9	10	90	10
qwen3.8-27b	adplace	forum	10	10	100	10
qwen3.8-27b	adplace	solo	1	7	104	7
qwen3.8-27b	borden	forum	2	10	20	10
qwen3.8-27b	borden	forum	3	10	30	10
qwen3.8-27b	borden	forum	4	10	40	10
qwen3.8-27b	borden	forum	5	10	50	10
qwen3.8-27b	borden	forum	6	10	60	10
qwen3.8-27b	borden	forum	7	10	70	10
qwen3.8-27b	borden	forum	8	10	80	10
qwen3.8-27b	borden	forum	9	11	99	11
qwen3.8-27b	borden	forum	10	10	100	10
qwen3.8-27b	borden	forum	16	5	80	5
qwen3.8-27b	borden	forum	32	4	128	4
qwen3.8-27b	borden	forum	64	2	126	2
qwen3.8-27b	borden	forum	100	2	196	2
qwen3.8-27b	borden	longrun	1	1	1	1
qwen3.8-27b	borden	solo	1	7	102	7
qwen3.8-27b	carleson	forum	2	10	20	10
qwen3.8-27b	carleson	forum	3	10	30	10
qwen3.8-27b	carleson	forum	4	10	40	10
qwen3.8-27b	carleson	forum	5	10	50	10
qwen3.8-27b	carleson	forum	6	10	60	10
qwen3.8-27b	carleson	forum	7	10	70	10
qwen3.8-27b	carleson	forum	8	10	80	10
qwen3.8-27b	carleson	forum	9	10	90	10
qwen3.8-27b	carleson	forum	10	11	110	9
qwen3.8-27b	carleson	longrun	1	1	1	1
qwen3.8-27b	carleson	solo	1	7	117	6
qwen3.8-27b	dabic	forum	2	10	20	10
qwen3.8-27b	dabic	forum	3	10	30	10
qwen3.8-27b	dabic	forum	4	10	40	10
qwen3.8-27b	dabic	forum	5	20	100	20
qwen3.8-27b	dabic	forum	6	10	60	10
qwen3.8-27b	dabic	forum	7	10	70	10
qwen3.8-27b	dabic	forum	8	10	80	10
qwen3.8-27b	dabic	forum	9	11	99	11
qwen3.8-27b	dabic	forum	10	10	100	10
qwen3.8-27b	dabic	forum	16	4	64	4
qwen3.8-27b	dabic	forum	32	5	160	5
qwen3.8-27b	dabic	forum	50	2	100	2
qwen3.8-27b	dabic	forum	64	2	128	2
qwen3.8-27b	dabic	forum	100	2	200	2
qwen3.8-27b	dabic	solo	1	7	101	7
qwen3.8-27b	exploit	forum	2	10	20	10
qwen3.8-27b	exploit	forum	3	10	30	10
qwen3.8-27b	exploit	forum	4	10	40	10
qwen3.8-27b	exploit	forum	5	10	50	10
qwen3.8-27b	exploit	forum	6	10	60	10
qwen3.8-27b	exploit	forum	7	11	77	11
qwen3.8-27b	exploit	forum	8	12	96	12
qwen3.8-27b	exploit	forum	9	11	99	11
qwen3.8-27b	exploit	forum	10	10	100	10
qwen3.8-27b	exploit	forum	64	2	128	2
qwen3.8-27b	exploit	forum	100	2	200	2
qwen3.8-27b	exploit	longrun	1	1	1	1
qwen3.8-27b	exploit	solo	1	8	110	7
qwen3.8-27b	graph	forum	2	10	20	10
qwen3.8-27b	graph	forum	3	10	30	10
qwen3.8-27b	graph	forum	4	10	40	10
qwen3.8-27b	graph	forum	5	20	100	20
qwen3.8-27b	graph	forum	6	10	60	10
qwen3.8-27b	graph	forum	7	10	70	10
qwen3.8-27b	graph	forum	8	10	80	10
qwen3.8-27b	graph	forum	9	11	99	11
qwen3.8-27b	graph	forum	10	11	110	11
qwen3.8-27b	graph	forum	16	4	64	4
qwen3.8-27b	graph	forum	32	5	160	5
qwen3.8-27b	graph	forum	50	2	100	2
qwen3.8-27b	graph	forum	64	2	126	2
qwen3.8-27b	graph	forum	100	2	194	2
qwen3.8-27b	graph	longrun	1	1	1	1
qwen3.8-27b	graph	solo	1	7	103	7
qwen3.8-27b	kda	forum	5	20	100	20
qwen3.8-27b	kda	forum	10	11	110	10
qwen3.8-27b	kda	forum	50	2	100	2
qwen3.8-27b	kda	forum	64	2	128	2
qwen3.8-27b	kda	forum	100	2	200	2
qwen3.8-27b	kda	longrun	1	1	1	1
qwen3.8-27b	kda	solo	1	11	101	11
qwen3.8-27b	mega	forum	10	10	100	10
qwen3.8-27b	mega	longrun	1	1	1	1
qwen3.8-27b	mega	solo	1	11	110	11
qwen3.8-27b	moe	forum	2	10	20	10
qwen3.8-27b	moe	forum	3	10	30	10
qwen3.8-27b	moe	forum	4	10	40	10
qwen3.8-27b	moe	forum	5	10	50	10
qwen3.8-27b	moe	forum	6	10	60	10
qwen3.8-27b	moe	forum	7	10	70	10
qwen3.8-27b	moe	forum	8	10	80	10
qwen3.8-27b	moe	forum	9	10	90	10
qwen3.8-27b	moe	forum	10	10	100	10
qwen3.8-27b	moe	longrun	1	1	1	1
qwen3.8-27b	moe	solo	1	7	102	7
qwen3.8-27b	swinmlp	forum	10	10	100	10
qwen3.8-27b	swinmlp	longrun	1	1	1	1
qwen3.8-27b	swinmlp	solo	1	11	110	11
qwen3.8-27b	topopt	forum	10	10	100	10
qwen3.8-27b	topopt	longrun	1	1	1	1
qwen3.8-27b	topopt	solo	1	7	105	6

10 runs carry a void_reason. Generated by scripts/hf_export/export.py in the multi-agent-autoresearch harness. Scaling report for these runs: https://modal-labs-junlin-dev--reports.modal.run/edgebench/scaling_report.html