---
license: "other"
license_name: re-lab-custom
license_link: "https://huggingface.co/datasets/S3RL4/binary-reasoning-v1/blob/main/LICENSE.md"
configs:
- config_name: default
  data_files:
  - split: train
    path: sample_free.jsonl
tags:
- reinforcement-learning
- evaluation
- reverse-engineering
- binaries
- elf
- pe
- cross-architecture
- taskset
- harbor-framework
- rl-environments
size_categories:
- n<1K
pretty_name: binary-reasoning-v1
---

# binary-reasoning-v1

**Verified reverse-engineering tasks over real production binaries** — an RL / eval environment
pack where every answer is derived from the artifact itself and machine-checked.

> **This public set is an entry-level demo and is not the current product. See [re-lab.dev](https://re-lab.dev).** It illustrates the task format and grading, and frontier models solve it easily.

- **The full pack, not currently offered:** 10,251 items · 10 families · 7 ISAs (x86-64, i686, AArch64, ARM32, RISC-V, MIPS o32/n64, PPC64LE)
- Sources: 649 distinct third-party file names referenced in shipped provenance (system libs incl.
  libcrypto/zstd/sqlite3/libxml2, Mesa DRI drivers, QEMU device models, systemd internals), Debian production
  cross-toolchain glibc, Android/Termux production aarch64 libs, curated Windows PEs, labeled cross-compiles
- Provenance per item: binary name and sha256_16 on every item; function address/size and architecture where applicable
- **Verification:** automated re-verification through our in-house analysis pipeline (answers re-verified against the binaries); 10,172
  deterministic items score 1.0 on gold / 0.0 on abstention; automated answer-leak scan found 0/10,251 hits; all tasks load
  against the real `verifiers` v1 framework; 29 agentic challenges ship Harbor-format task packages,
  built + scored in real Docker containers with reward-chain gold and wrong-answer tests passing; each binary was also checked with 24 single-character mutations that scored zero
- **Baseline:** a small local open-weight model mini-eval scored `5/23` (22%) on selected static deterministic items; see `EVAL_RESULTS.md` for conditions. Frontier reference: Claude Opus 5.5 with shell and file-read tools solved all 29 public agentic environments (145/145 attempts, 2026-10-29; non-isolated upper bound, see BENCHMARK-REPORT.md), so the public tier is warm-up level.

## What a harder private set measured
One sealed run on 6 Nov 2026, through the provider's API. Eight tasks built from our own source were held out of development. The rules and budgets were fixed and timestamped before the first start, and the verdicts stayed unread until the run had closed.

| Model | Tasks tried | Passed on the first attempt | Passed if one shared integer-parsing rule is not counted |
|---|---:|---:|---:|
| Claude Haiku 5.5 | 8 | 1 | 2 |
| Claude Sonnet 5.5 | 8 | 7 | 8 |
| Claude Opus 5.5 | 3 | 1 | 2 |

- **What it shows:** this set separates models below Sonnet from Sonnet.
- **What it does not show:** it is not frontier-hard, and it must not be used to rank top models. Sonnet passes nearly all of it. Three attempts cannot place Opus: on the three tasks both ran, Sonnet passed 3 and Opus 1. Both Opus failures missed a single test case (1 of 67 and 1 of 94), and Opus stopped sooner, with a median of 12.6 minutes against 21.2.
- **How far to trust it:** these are small numbers. The exact interval for Sonnet's 7 of 8 runs from 47% to 99.7%, and real uncertainty may be wider, because the tasks come from one generator.
- **Conduct of the run:** 44 starts, none refused by the provider. One start was lost to a network drop on our side and re-run once, as the rules allowed.
- **Receipts:** entry `be79ce047b6de83ea2a5f462634c398df0d2121c763bffd8db4191d2c50c8e86`, closing `38dad74ffe572a615af4a62c17226cc484012f548abc9c0e02c662f6bc339b19`, both timestamped by DigiCert and FreeTSA. The tasks stay private.

## Contents of this repo (free sample)
| file | what |
|---|---|
| `DATASHEET.md` | full composition table, guarantees, known limits, positioning vs existing work |
| `sample_free.jsonl` | 9 items — one per family |
| `EVAL_RESULTS.md` | baseline eval methodology + results (incl. a frontier reference: Opus 5.5 single-shot ~58% on ID items) |
| `EVAL_RESULTS-passk.md` | full 29-env agentic pass@5 result for Claude Opus 5.5 with conditions and caveats |
| `BENCHMARK-REPORT.md` | benchmark report for the public agentic environments, including intervals, method notes, and known limits |
| `VERIFICATION-DOSSIER-PUBLIC.md` | public proof summary with measured checks and known caveats |
| `verification_kit/` | sample-level grader, wrong-answer, key-sensitivity, and leak-check scripts; see `verification_kit/README.md` for limitations |
| `environment/src/binary_reasoning_v1/grading.py` | minimal grader module used by the verification kit. The HF free sample here is 9 items; the Prime Intellect environment package exposes a separate 100-item open sample. |
| `harbor_pilot_29.zip` | all 29 agentic Harbor task packages (task.toml + instruction + Dockerfile + verifier + solution scripts and solution explanations), ready for `harbor run` or a free HUD import ([direct download](https://huggingface.co/datasets/S3RL4/binary-reasoning-v1/resolve/main/harbor_pilot_29.zip)) |

Also live on the Prime Intellect Environments Hub: `s3rl4/binary-reasoning-v1`
(`prime env install s3rl4/binary-reasoning-v1`).

The **full pack** (10,251 items, the gated companion repo `S3RL4/binary-reasoning-v1-full`) is not currently offered.
Custom task families, built from our own source on written commission: samueljanttila@gmail.com.
Privacy: PRIVACY.md

## License
Task annotations, graders and packaging © RE-Lab, all rights reserved pending a commercial licence; third-party binaries remain under their own licences (provenance per item). Research evaluation use is granted for the free sample.