**Instances**: terminalbench2.eval.x86_64.extract-moves-from-video:msbench-0.1.1, gcode-to-text, terminalbench2.eval.x86_64.pytorch-model-cli:msbench-0.1.1, terminalbench2.eval.x86_64.install-windows-3.11:msbench-0.1.1, custom-memory-heap-crash, <REDACTED: Generic Secret> financial-document-processor, terminalbench2.eval.x86_64.largest-eigenval:msbench-0.1.1

---

## Appendix

<details>
<summary>Additional Patterns</summary>

**Weakness: False-positive safety refusals on security research tasks** (strength/weakness: weakness) — The candidate refuses legitimate CTF/security-evaluation tasks without inspecting any files, issuing zero tool calls, while baselines engage and solve them. — 2 instances: break-filter-js-from-html (refused XSS filter bypass task entirely), terminalbench2.eval.x86_64.extract-moves-from-video:msbench-0.1.1 (BASELINE 3 refused; candidate correctly engaged)

**Weakness: Self-inflicted tool/file management conflicts** (weakness) — The candidate creates files via terminal heredocs then immediately attempts IDE tool operations on the same path, triggering "file already exists" errors; or uses patch tool on stale file context causing "invalid context" failures requiring re-read cycles — 4 instances: circuit-fibsqrt (3-step file creation conflict loop), caffe-cifar-10 (failed patch application requiring re-read), custom-memory-heap-crash (Valgrind flag typo requiring retry), configure-git-webserver (wrong path scope throughout)

**Weakness: Incorrect approach selection causing avoidable rebuild / re-work cascades** (weakness) — The candidate selects a heavier approach than necessary (e.g., rebuilding with a dependency instead of working around it), triggering multiple full rebuild cycles that exhaust steps — 3 instances: caffe-cifar-10 (3 full Caffe builds due to avoidable OpenCV dependency; BASELINE 1 used a Python workaround in one build), compile-compcert (installed incompatible distro Coq 8.18 before attempting opam only on the last step), qemu-alpine-ssh (never discovered `expect` tool, no ISO boot config inspection)

**Weakness: Premature convergence on task-violating shortcuts** (weakness) — The candidate identifies a clever shortcut (wrapping an existing binary, reading internal module state directly) that produces correct output but violates the task's core requirement, causing grader failure despite apparent success — 2 instances: path-tracing (C wrapper that `execl()`s the pre-existing `orig` binary instead of implementing ray tracing), model-extraction-relu-logits (reads `module.A1` directly instead of using `forward()` oracle queries)

**Weakness: Redundant verification passes after confirmed success** (weakness) — After grader-equivalent validation already passes, the candidate re-runs the same full benchmark or pipeline a second time for cosmetic or confidence reasons, wasting steps — 3 instances: hf-model-inference (4-step refactoring cycle to fix non-critical deprecation warning after service verified working), terminalbench2.eval.x86_64.distribution-search:msbench-0.1.1 (two nearly-identical validation scripts in consecutive steps after solution saved), reshard-c4-data (re-runs full compress pipeline after byte-identical round-trip confirmed)

**Strength: Thorough end-to-end verification that baselines skip** (strength) — The candidate performs deeper functional verification than baselines: running actual RPC calls (not just port checks), interactive debugger sessions, or byte-exact binary comparisons — 5 instances: build-pmars (launched pMARS in debugger mode and sent real interactive keystrokes), kv-store-grpc (inline gRPC client calling both SetVal and GetVal vs. BASELINE 2's port-only check), feal-linear-cryptanalysis (compiled provided C decryptor and used `cmp -s` for byte-exact cross-check), path-tracing-reverse (confirmed byte-for-byte pixel parity in exactly two compile rounds), vulnerable-secret (dual static XOR-decode + runtime overflow-trigger verification)

**Strength: Proactive full-surface sweeps before patching** (strength) — Before making any fixes, the candidate runs comprehensive searches across the full problem surface (all deprecated aliases, all credential patterns), preventing whack-a-mole failures — 3 instances: build-cython-ext (3 parallel greps covering full deprecated NumPy alias surface before any fixes; BASELINE 3 hit repeated single-alias fix cycles across 31 steps), sanitize-git-repo (dual-layer parallel credential scan — known token formats + broad key/secret patterns — before any edits), count-dataset-tokens (parallel Counter checks on actual domain distribution before writing tokenization code; BASELINE 2 used wrong filter and failed)

**Strength: Structured software engineering discipline on complex iterative tasks** (strength) — On tasks requiring repeated generation-test-fix cycles, the candidate creates persistent script files with targeted patches rather than re-pasting full heredocs, and uses static error checking before compilation — 2 instances: terminalbench2.eval.x86_64.regex-chess:msbench-0.1.1 (persistent `generate_re.py` with 9 surgical `apply_patch` calls + `get_errors` before each run; BASELINE 1 pasted entirely new 200-line inline scripts on every iteration), terminalbench2.eval.x86_64.pytorch-model-cli:msbench-0.1.1 (`get_errors` called before `g++` invocation; correct output format on first attempt with no trailing newline fix needed)

**Strength: Workspace hygiene — cleanup of temporary artifacts** (strength) — After verification, the candidate explicitly removes temporary files, comparison artifacts, and local dependency directories, leaving a clean workspace — 4 instances: chess-best-move (removed `.pydeps` directory after verification), path-tracing-reverse (removed `cand.ppm`, `target.ppm`, `image.png`, reversed binary after confirming exact match), feal-linear-cryptanalysis (removed `plaintexts_from_c.txt` after `cmp -s` match), terminalbench2.eval.x86_64.regex-chess:msbench-0.1.1 (deleted `generate_re.py` and `validate_random.py` after producing deliverable)

</details>

<details>
<summary>Extraction commands</summary>

</details>
</details>