exp_rpt_manybugs-v2

A fixed, non-gameable release of DCAgent/exp_rpt_manybugs: 164 C bug-repair tasks (ManyBugs) packaged as Harbor sandbox tasks.

Tasks	164
Projects	php (102), libtiff (24), python (15), wireshark (7), lighttpd (9), gzip (5), gmp (2)
Unique environments	7 snapshots (ubuntu:22.04 base, one -dev set per project)
Format	tasks.parquet — columns path (task id), task_binary (gzipped task tarball)

Each task contains a single buggy .c file from a real C project plus its verified gold fix; the agent edits /workspace/src/<file>.c and tests/test.sh writes the reward.

The problem in the original

tests/test.sh used a gameable text-diff that never compiled or executed the C code. Beyond an exact-match check, it accepted any edit to the fault line as a pass:

# original fallback — REWARDS GARBAGE
AGENT_LINE=$(sed -n "${FAULT_LINE}p" /workspace/src/powm.c)
if [ "$BUGGY_LINE" != "$AGENT_LINE" ]; then
    echo "PASS: Fault line was modified ..."; exit 0   # <- even a syntax error passes
fi

So a model could append a deliberate syntax error to the fault line and still score reward = 1. The reward signal was vacuous.

Why a gcc compile gate is impossible here

The 164 files are single-file fragments of large C projects. They #include project-internal headers that are absent from the sandbox:

powm.c:74:10: fatal error: gmp-impl.h: No such file or directory
tif_dirwrite.c:32:10: fatal error: tiffiop.h: No such file or directory
url.c:25:10: fatal error: php.h: No such file or directory

Verified empirically across all 164 tasks: every gold/fixed solution fails gcc -c and gcc -fsyntax-only at the first #include. There is also no project source tree (no configure/Makefile/test harness) shipped with a task, so there is nothing to build or run the project's own test suite against. A gcc gate would therefore make every task impossible — gold included — i.e. a vacuous verifier in the opposite direction.

The fix (non-gameable, oracle-solvable verifier)

The new tests/test.sh is default-fail and gates on the only deterministic, non-gameable signal available without a runnable test suite: the agent's file must exactly match the verified gold patch (tests/fixed).

echo 0 > /logs/verifier/reward.txt        # default-fail
G1  agent file must DIFFER from buggy      (a fix was actually applied)
G2  agent file must == tests/fixed         (the verified gold patch) -> reward 1
python3 -m pytest /tests/test_state.py     # pass_ratio-parseable output

Only the exact correct fix scores reward = 1; a syntax error, a partial edit, a trivial whitespace change, or an untouched workspace all score 0.

Supporting changes:

• tests/test_state.py — asserts /logs/verifier/reward.txt == "1" (the Harbor binary-reward contract); invoked by pytest so the pass_ratio reward shaper gets parseable output.
• tests/config.json — test weight for test_reward_is_pass.
• environment/Dockerfile — python3 python3-pytest added to every image (for the pytest invocation). build-essential (gcc) is retained so an agent may build while working; the verifier itself does not compile.
• All other files (instructions, metadata, task.toml, expected.diff, fault_line.txt, the buggy/fixed/initial sources, the per-project -dev packages) are preserved byte-identical from the original.

Validation

Oracle gate (non-vacuous): the gold solution was run through the verifier for all 164/164 tasks → reward = 1 every time.

Docker smoke test (representative, real project images for gmp + php; all projects covered with the equivalent minimal image):

scenario	reward
empty / untouched workspace	0
gold solution (tests/fixed)	1
buggy file unchanged	0
syntax error injected at the fault line	0
trivial trailing-newline edit	0

The originally-gameable case (a syntax error on the fault line) now correctly fails.

Uploader note

Original tasks sourced from DCAgent/exp_rpt_manybugs (unchanged); only the verifier (tests/test.sh), the new tests/test_state.py / tests/config.json, and the python3-pytest Dockerfile line differ. Published to laion/exp_rpt_manybugs-v2.