Avoidance-Behavior-Exam

Ready-to-run Harbor task.toml task trees for RetreatBench's 6 target benchmarks. Managed from https://github.com/a-green-hand-jack/RetreatBench via infra/hub-datasets/*.yaml manifests + infra/tools/fork_hub_dataset.py (fork) and infra/adapters/*/ (opencode-agent conversion for benchmarks with no existing Harbor-style version). Conversion code and agent definitions live in that repo, not here -- this dataset holds converted output only.

Contents

Path	Benchmark	Source	Pinned revision
terminal-bench-2.0/	Terminal-Bench 2.0	harborframework/terminal-bench-2.0 (Harbor-native)	f2e8c75e23add71613117eecc9498f53bcd7e04e
terminal-bench-science/	Terminal-Bench-Science	harborframework/terminal-bench-science-lfs (Harbor-native)	d48ce2a26e4dcd525c9e8ceb8dd2b60652e178cd
paperwritingbench/	PaperWritingBench	Jack-Jieke-Wu/Paper-Writing-Exam, subset paperwritingbench-sparse-plotoff/ (maintainer pre-converted)	5fe375dbd440409f0180e10dee213b1685c8f40d
paperwrite-bench/	PaperWrite-Bench	Jack-Jieke-Wu/Paper-Writing-Exam, subset paperwrite-bench-short/ (maintainer pre-converted)	5fe375dbd440409f0180e10dee213b1685c8f40d
terminal-bench-1x/	Terminal-Bench 1.x	adapted via infra/adapters/terminal-bench-1x/ (opencode agent, no upstream Harbor version exists)	see coverage below
researchclawbench/	ResearchClawBench	adapted via infra/adapters/researchclawbench/ (opencode agent, no upstream Harbor version exists)	see coverage below

terminal-bench-1x/ and researchclawbench/ coverage

Every published task here passed both a mechanical check (task.toml schema, a real environment build, and -- for ResearchClawBench -- an automated check that no private ground truth leaked into any agent-visible path: an exact content-hash match hard-fails unconditionally, and a filename-only collision with differing content blocks until a human explicitly reviews and clears that specific filename) and a real harbor run smoke-test trial (a real subject agent, short budget, n-attempts 1) that produced a valid verifier reward. "Passing the real-run check" means the conversion's build+agent+verify machinery works end to end -- it does not mean the subject agent solved the task; reward values of 0.0 are expected and still count as a pass.

Scope note (Terminal-Bench 1.x): rather than the raw upstream task tree (241 directories, many experimental/one-off), conversion targets the official terminal-bench-core v0.1.1 registry subset (80 task ids) -- 67 of those map to an actual directory in the upstream repo; the other 13 are variant-suffixed entries (e.g. crack-7z-hash.easy/.hard) or renamed tasks not yet handled by this adapter and are excluded from this round.

Status as of 2026-08-31 (updated -- Information_002 resolved by human review, see "Changes since the last update" below)

Benchmark	Attempted	Published (passed both checks)	Failed (logged, not published)
Terminal-Bench 1.x	67 / 67	65 (97%)	2
ResearchClawBench	40 / 40	40 (100%)	0

Terminal-Bench 1.x remaining failures:

• build-initramfs-qemu -- the verifier boots a QEMU VM to check the built initramfs; timed out even at 8x the smoke-test's verifier-timeout budget (1440s) while the boot verification was still in progress. Genuinely heavy for this validation layer's scope, not a conversion defect.
• train-fasttext -- the subject agent was OOM-killed (exit 137) mid-training on the real yelp dataset. Genuinely memory-intensive, not a conversion defect.

ResearchClawBench: no remaining failures. Information_002, previously held back pending a human decision, has been reviewed and published -- see below.

Changes since the last update (transparency record, not silently overwritten)

This round (ResearchClawBench 39/40 -> 40/40):

• Information_002 was previously withheld because the automated private-ground-truth-leak check flagged the filename 2111.01152_auto.md as appearing both in the private target_study/ tree and under the agent-visible environment/. A human reviewer has now examined the evidence and confirmed this is a false positive; the task is published. The evidence:

  • Different SHA256. Private copy 9e67b020f45c...; agent-visible copy 64068ea01b3c....
  • Different size. 2,336 B (private) vs. 69,829 B (agent-visible) -- a ~30x difference.
  • Non-overlapping content. The private copy is a 1-section excerpt corresponding to a single checklist step; the agent-visible copy is the full 16-section source document.
  • The agent-visible copy is independently justified. It sits under data/2111.01152/, which the upstream task's own task_info.json explicitly declares as feature_data -- "the target scientific paper defining the AB-stacked MoTe2/WSe2 moiré system and its Hamiltonian parameters" -- i.e. a required task input, not leaked ground truth.
  • The published tree was re-checked after upload: it contains no target_study/ content of any kind.

  The task then passed the same dual verification as every other published task: mechanical check OK, and a real harbor run trial (opencode / gpt-5.6-sol, 1 trial, 0 exceptions, 4m58s) that wrote a valid reward.txt of 0.0 -- an expected value at this validation layer, which checks that build+agent+verify works end to end, not that the agent solved the task.

• The leak checker itself was fixed, not loosened. Previously an exact content-hash match and a filename-only match produced the same PRIVATE LEAK hard failure, which mischaracterized this case and left no way to record a human review. It now distinguishes the two: content-hash matches remain an unconditional hard failure (a real leak), while a filename-only collision with differing content is reported as a distinct warning that still blocks by default and can only be cleared by an explicit, per-filename --confirm-name-collision acknowledging a human review. Re-running the checker across all 40 converted ResearchClawBench tasks after the change produced 0 hard failures and exactly the 1 expected warning, confirming no new blind spot was introduced.

Previous round (TB1.x 56/67 -> 65/67, RCB 31/40 -> 39/40):

Investigating and re-running those failures (per explicit direction) found and fixed real bugs, not just retried with more patience:

• cron-broken-network (TB1.x): investigated the upstream task's own design intent by reading Terminal-Bench's harness source. Confirmed the task's premise (a deliberately broken network) is meant to be solved by an agent that never needs outbound network access itself -- Terminal-Bench's own native driving model (and Harbor's equivalent, terminus-2) drives the agent from outside the container via tmux + the LLM API, with zero in-container self-installation. The pipeline's default smoke-test agent (opencode) needs to npm install itself into the container over the network it doesn't have -- an inherent mismatch with this specific task, not a general "always grant network access" fix. Re-ran with --agent terminus-2 instead: passed.
• Multi-container docker-compose.yaml bug, found on retry (TB1.x -- affected security-vulhub-minio, simple-sheets-put, simple-web-scraper): an earlier fix only stripped Terminal-Bench's ${T_BENCH_...} template variables from the client/main service. Real harbor run trials showed every OTHER service in a multi-container compose file has the same problem (an unresolved template variable resolving to an empty string breaks Docker/compose validation in three different ways depending which field it's in). Fixed by stripping every ${T_BENCH_...} reference from every service, not just the main one. All 3 tasks now pass with real reward (2 of them reward=1.0).
• fix-pandas-version (TB1.x): investigated why the upstream Dockerfile COPYs a file from outside its own task directory (COPY tests/src/ /app/src/) -- confirmed this is a real, recurring Terminal-Bench 1.x pattern (roughly 1 in 20 upstream tasks does this), not a one-off, since upstream's build context is the whole task root while Harbor's is environment/-scoped only. Verified the referenced file is only ever read from the already-built image path at verify time, never off raw disk, so inlining it into the converted task's own environment/ and rewriting the COPY path is safe. Fixed; passed on retry.
• Material_002 (RCB): the earlier "confirmed not resource contention" conclusion needed re-checking -- the box was busier than assumed (~9-10 load average from unrelated concurrent projects, plus a real infra bug: 26 stale containers from earlier timed-out trials, some 5-9 hours old, had exhausted Docker's network address pool). After cleaning those up and retrying with a longer verifier-timeout-multiplier, the task ran for 1h29m (a real, long agent run -- this task's own budget is 4 hours) and passed.
• build-initramfs-qemu and train-fasttext were retried with substantially longer budgets and still failed for the reasons stated above -- these are recorded as genuine limitations, not bugs papered over with more retries.

See docs/benchmark-hub.md in the RetreatBench repo for the full fork-vs-adapt methodology, and GitHub Issue #6 for the full progress log including the Information_002 investigation writeup.