# InfoFlow-SWE — SI25 specification-level task suite Software-engineering agent tasks published at six graded **specification levels**, so the same task can be run while varying only how much of the specification the agent is told up front. Three benchmarks x 25 tasks x 6 levels = 450 task packages. Release id `si25-rebuilt-v1`, certified 2026-08-11. | directory | benchmark | upstream runner | tasks | |---|---|---|---| | `swebench_pro_si25/` | SWE-bench-Pro (SI25 subset) | Harbor | 25 | | `deepswe_si25/` | DeepSWE (SI25 subset) | Pier | 25 | | `tb_top25/` | Terminal-Bench 2 (top-25 subset) | Harbor | 25 | ## Layout ``` active.json per-benchmark task counts, certification flags, digests copy_manifest.json source lineage and logical-tree fingerprints / task_set.json the fixed 25 task ids for this benchmark original/ S0/ S1/ S2/ S3/ S4/ / one complete, runnable task package instruction.md the only file that changes across levels task.toml timeouts, network policy, resources environment/ image build context tests/ verifier scripts and hidden tests solution/ reference solution axis_data// evaluator-side metadata (never show to an agent) atoms.jsonl the task's atom inventory specification_levels.json which atoms each level discloses or withholds verification_targets.json per-test natural-language behaviour cards provenance.json extraction pipeline record _artifacts/ manifest.json certification record and subtree digests public// atoms, specifications, audit, source, provenance reviews/ per-task review verdicts ``` ## What each folder is for **`original/`** — the upstream instruction exactly as the source benchmark ships it. Use this as the unmodified baseline. **`S0/` … `S4/`** — the graded ladder. `S0` gives only workspace access ("Inspect and edit the provided repository in /app."); each step discloses more of the specification; `S4` is the fullest. Within one task, `environment/`, `tests/`, `solution/` and `task.toml` are byte-identical across all six levels — **only `instruction.md` differs** — so any outcome difference between levels is attributable to the instruction alone. **`axis_data/`** — evaluator-side. Keep it out of anything the agent can read. - `atoms.jsonl` — the task decomposed into atomic facts, one JSON object per line: `{id, text, category, default_specification_level, synthetic, provenance_quote, source}`. `source` carries a span anchor: `{source_id, path, field, quote_start, quote_end}`. Categories: `core_task`, `public_interface`, `behavioral_requirement`, `guidance_hint`, `debugging_context`, `workspace_access`. - `specification_levels.json` — the definition of the ladder. `levels.` gives `selected_atom_ids`, `withheld_atom_ids`, `content_text`, `instruction_text` and `renderer`. - `verification_targets.json` — maps each test node id to a natural-language description of the behaviour it checks (`cards`), plus `misses`. - `provenance.json` — how the inventory was produced: extraction and auditor models, repair rounds, code commit. **`_artifacts/public/`** — the same `atoms.jsonl` (byte-identical to `axis_data/`) alongside `specifications.json` (same schema as `specification_levels.json`), `source.json` (the original upstream instruction, which files were read, section decomposition), and `audit.json` — an independent-model audit reporting `duplicate_atom_groups`, `non_atomic_atom_ids`, `misclassified_atom_ids`, `unsupported_atom_ids`, and per-level `leaked_withheld_ids`. **`_artifacts/reviews/`** — `specifications//S1..S4.json` and `atoms/.json`, each with `verdict`, `reviewer` and `reviewed_at`. S0 is not reviewed; it is a fixed stub. ## Task package contents by benchmark | | swebench_pro_si25 | deepswe_si25 | tb_top25 | |---|---|---|---| | build | `environment/Dockerfile` | `environment/Dockerfile`, `pre_artifacts.sh` | `environment/Dockerfile`, `environment/protected.tar.gz.enc` | | tests | `config.json`, `parser.py`, `run_script.sh`, `test.sh` | `Dockerfile`, `config.json`, `grader.py`, `test.patch`, `test.sh` | `test.sh`, `test_outputs.py` | | solution | `solve.sh` | `solution.patch`, `solve.sh` | `solve.sh` | ## How to use **Run one condition.** Pick a level directory and run the package with the benchmark's own runner — Harbor for `swebench_pro_si25` and `tb_top25`, Pier for `deepswe_si25`. The agent reads `instruction.md`; `tests/` decides pass or fail. To compare regimes, run the same `` at two levels and hold everything else fixed. **Packaging caveat.** `deepswe_si25` and `tb_top25` packages carry `schema_version` and a `[task]` block (150/150 each). The `swebench_pro_si25` packages do **not** (0/150) — they are in upstream Harbor directory form, where the task name comes from the directory. Add a `[task]` name, `schema_version`, and any artifact-collection hooks your harness expects before running those through a registry-based launcher. **Ask / clarification experiments.** `specification_levels.json` gives `withheld_atom_ids` per level — exactly the facts an agent does *not* have at that level. Use them as the pool a simulated user may disclose on request, and score disclosure against atom ids. **Verification experiments.** `verification_targets.json` `cards` describe what each hidden test checks in words, so you can present a behavioural target without revealing the test. **Network policy.** `task.toml` controls agent and build network access separately, and the three namespaces do not agree: | namespace | agent network | set by | |---|---|---| | `swebench_pro_si25` | denied | `[agent] network_mode = "allowlist"` with `allowed_hosts = []` | | `deepswe_si25` | denied | `[environment] allow_internet = false` (Pier schema) | | `tb_top25` | **allowed** | `[environment] network_mode = "public"`, no `[agent]` override | Terminal-Bench tasks run with network access. Some of those tasks need it, but it also means an agent can reach the upstream repository, and these tasks are pinned to public commits whose fixes are findable online. If you are measuring what an agent derives from the specification alone, restrict egress for `tb_top25` yourself.