Add public verification summary
Browse files# Verification dossier — public summary
This is the public part of RE-Lab's verification dossier for `binary-reasoning-v1`. The full buyer-ready version adds procurement, provenance tables, calibration plans, and delivery details. Contact: samueljanttila@gmail.com.
**Dataset:** https://huggingface.co/datasets/S3RL4/binary-reasoning-v1
**Self-serve sample kit:** `verification_kit/README.md` in this dataset repo
**Canary for contamination checks:** `RE-Lab-brv-2026-10-29-8c1e4f7a`
## What is verified now
### 1. Pack counts and grader split
The full pack contains `10,251` items:
- `10,172` deterministic-graded items.
- `79` rubric-judged open-ended items, disclosed rather than sold as deterministic.
The public sample includes 9 free items and a 100-item open sample in the Prime environment package. The sample kit skips rubric-judged items instead of pretending they are deterministic.
### 2. Grader sanity checks
On the 100-item open sample, the public sample kit reports:
- `gold.py`: deterministic gold replies pass `84/84`; `16` judge items skipped.
- `do_nothing.py`: empty answers score `0` on all `84` deterministic items.
- `random_wrong.py`: deterministic mutated wrong answers score `0` on all `84` deterministic items.
- `grader_tamper.py`: mutating the answer key makes the original gold reply fail on all `84` deterministic items.
- `leak_check.py`: `items=100 unique_ids=100 checked_rows=73 leaks=0`.
On the 9-item free sample:
- deterministic sanity checks pass `4/4`; `5` judge items skipped.
- `leak_check.py`: `items=9 unique_ids=9 checked_rows=3 leaks=0`.
### 3. Answer-leak audit
The existing full-pack answer-token audit reports `0/10,251` exact-token leaks. The public sample script is a smaller reproduction tool and reports checked-row coverage (`73/100` checkable rows on the open sample; `3/9` on the free sample). It does not replace the full-pack audit.
### 4. Agentic environments
The public agentic tier has `29` Harbor-format flag-hunt environments. The Docker reward chain was tested with gold answers and wrong-answer tests, and each binary was checked with 24 single-character mutations that scored zero. In the JSONL/verifiers path, the deterministic grader checks whether the stored flag appears in the reply as a substring; that is weaker than the Harbor oracle.
### 5. Hard-trial reproduction evidence
The 99-item hard trial pack has 19 family-level reproduction scripts in RE-Lab's pinned Ubuntu 24.04/WSL2 and Windows test environment. They reproduce `99/99` gold answers without reading the `answer` field. About `77` rows are command/binary re-execution (reproducibility, not an independent solve); about `22` rows use static ELF/PE/argv-derived reasoning. No human solved these tasks.
### 6. Calibration evidence so far
The public agentic tier is intentionally warm-up level for a frontier agent: Claude Opus 5.5 solved `29/29` environments and `145/145` attempts in a non-isolated host run with Bash/Read tools, graded by the binary oracle.
A separate local open-weight calibration run on the 99-item hard trial pack scored `14/99` at pass@8:
- model: `qwen3.8-flash-next-iq3_s`, 3-bit quantized;
- single-turn, no tools;
- temperature `0.7`, `max_tokens=600`, with a `reasoning_content` fallback;
- whole-reply normalized matching rather than each item's verifier rule.
That is headroom evidence, not a published band table or proof that each task sits in the RL mid-band. Per-buyer model calibration is part of the paid pilot path.
## Known limitations and caveats
- Current deterministic graders can accept some "shotgun" replies that contain the correct token among many distractors. A dedicated `shotgun.py` probe is a planned kit addition.
- Some hard-trial reproduction scripts are environment-bound: they depend on local paths, WSL/Windows tools, pinned package digests, or host-specific loader behavior. They are evidence, not yet a portable buyer-runnable solver suite.
- Some shipped artifact sets include grader-side metadata or unstripped twins that could reveal an answer if mounted to the model. A model-visible vs grader-only manifest is needed before selling sealed evaluation as fully isolated.
- The full-pack public exposure window was 2026-10-28 22:02 → 2026-10-29 17:08 UTC, then older public versions were deleted. Deletion reduces future access but does not prove no copies exist.
- The full pack uses production distro binaries with per-item provenance (`binary` path + `sha256_16`, plus address/size/architecture where applicable). A source-offer manifest lists `649` distinct third-party file names; upstream package/source mapping is a procurement follow-up for any redistribution-heavy order.
## What this does not claim
- It does not claim human or expert hand-checking of every task.
- It does not claim every task is model-solvable without tools.
- It does not claim the deterministic graders reject all low-effort answers.
- It does not claim the hard trial pack has a published pass-rate band for every task yet.
|
@@ -0,0 +1,62 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Verification dossier — public summary
|
| 2 |
+
|
| 3 |
+
This is the public part of RE-Lab's verification dossier for `binary-reasoning-v1`. The full buyer-ready version adds procurement, provenance tables, calibration plans, and delivery details. Contact: samueljanttila@gmail.com.
|
| 4 |
+
|
| 5 |
+
**Dataset:** https://huggingface.co/datasets/S3RL4/binary-reasoning-v1
|
| 6 |
+
**Self-serve sample kit:** `verification_kit/README.md` in this dataset repo
|
| 7 |
+
**Canary for contamination checks:** `RE-Lab-brv-2026-10-29-8c1e4f7a`
|
| 8 |
+
|
| 9 |
+
## What is verified now
|
| 10 |
+
|
| 11 |
+
### 1. Pack counts and grader split
|
| 12 |
+
The full pack contains `10,251` items:
|
| 13 |
+
- `10,172` deterministic-graded items.
|
| 14 |
+
- `79` rubric-judged open-ended items, disclosed rather than sold as deterministic.
|
| 15 |
+
|
| 16 |
+
The public sample includes 9 free items and a 100-item open sample in the Prime environment package. The sample kit skips rubric-judged items instead of pretending they are deterministic.
|
| 17 |
+
|
| 18 |
+
### 2. Grader sanity checks
|
| 19 |
+
On the 100-item open sample, the public sample kit reports:
|
| 20 |
+
- `gold.py`: deterministic gold replies pass `84/84`; `16` judge items skipped.
|
| 21 |
+
- `do_nothing.py`: empty answers score `0` on all `84` deterministic items.
|
| 22 |
+
- `random_wrong.py`: deterministic mutated wrong answers score `0` on all `84` deterministic items.
|
| 23 |
+
- `grader_tamper.py`: mutating the answer key makes the original gold reply fail on all `84` deterministic items.
|
| 24 |
+
- `leak_check.py`: `items=100 unique_ids=100 checked_rows=73 leaks=0`.
|
| 25 |
+
|
| 26 |
+
On the 9-item free sample:
|
| 27 |
+
- deterministic sanity checks pass `4/4`; `5` judge items skipped.
|
| 28 |
+
- `leak_check.py`: `items=9 unique_ids=9 checked_rows=3 leaks=0`.
|
| 29 |
+
|
| 30 |
+
### 3. Answer-leak audit
|
| 31 |
+
The existing full-pack answer-token audit reports `0/10,251` exact-token leaks. The public sample script is a smaller reproduction tool and reports checked-row coverage (`73/100` checkable rows on the open sample; `3/9` on the free sample). It does not replace the full-pack audit.
|
| 32 |
+
|
| 33 |
+
### 4. Agentic environments
|
| 34 |
+
The public agentic tier has `29` Harbor-format flag-hunt environments. The Docker reward chain was tested with gold answers and wrong-answer tests, and each binary was checked with 24 single-character mutations that scored zero. In the JSONL/verifiers path, the deterministic grader checks whether the stored flag appears in the reply as a substring; that is weaker than the Harbor oracle.
|
| 35 |
+
|
| 36 |
+
### 5. Hard-trial reproduction evidence
|
| 37 |
+
The 99-item hard trial pack has 19 family-level reproduction scripts in RE-Lab's pinned Ubuntu 24.04/WSL2 and Windows test environment. They reproduce `99/99` gold answers without reading the `answer` field. About `77` rows are command/binary re-execution (reproducibility, not an independent solve); about `22` rows use static ELF/PE/argv-derived reasoning. No human solved these tasks.
|
| 38 |
+
|
| 39 |
+
### 6. Calibration evidence so far
|
| 40 |
+
The public agentic tier is intentionally warm-up level for a frontier agent: Claude Opus 5.5 solved `29/29` environments and `145/145` attempts in a non-isolated host run with Bash/Read tools, graded by the binary oracle.
|
| 41 |
+
|
| 42 |
+
A separate local open-weight calibration run on the 99-item hard trial pack scored `14/99` at pass@8:
|
| 43 |
+
- model: `qwen3.8-flash-next-iq3_s`, 3-bit quantized;
|
| 44 |
+
- single-turn, no tools;
|
| 45 |
+
- temperature `0.7`, `max_tokens=600`, with a `reasoning_content` fallback;
|
| 46 |
+
- whole-reply normalized matching rather than each item's verifier rule.
|
| 47 |
+
|
| 48 |
+
That is headroom evidence, not a published band table or proof that each task sits in the RL mid-band. Per-buyer model calibration is part of the paid pilot path.
|
| 49 |
+
|
| 50 |
+
## Known limitations and caveats
|
| 51 |
+
- Current deterministic graders can accept some "shotgun" replies that contain the correct token among many distractors. A dedicated `shotgun.py` probe is a planned kit addition.
|
| 52 |
+
- Some hard-trial reproduction scripts are environment-bound: they depend on local paths, WSL/Windows tools, pinned package digests, or host-specific loader behavior. They are evidence, not yet a portable buyer-runnable solver suite.
|
| 53 |
+
- Some shipped artifact sets include grader-side metadata or unstripped twins that could reveal an answer if mounted to the model. A model-visible vs grader-only manifest is needed before selling sealed evaluation as fully isolated.
|
| 54 |
+
- The full-pack public exposure window was 2026-10-28 22:02 → 2026-10-29 17:08 UTC, then older public versions were deleted. Deletion reduces future access but does not prove no copies exist.
|
| 55 |
+
- The full pack uses production distro binaries with per-item provenance (`binary` path + `sha256_16`, plus address/size/architecture where applicable). A source-offer manifest lists `649` distinct third-party file names; upstream package/source mapping is a procurement follow-up for any redistribution-heavy order.
|
| 56 |
+
|
| 57 |
+
## What this does not claim
|
| 58 |
+
- It does not claim human or expert hand-checking of every task.
|
| 59 |
+
- It does not claim every task is model-solvable without tools.
|
| 60 |
+
- It does not claim the deterministic graders reject all low-effort answers.
|
| 61 |
+
- It does not claim the hard trial pack has a published pass-rate band for every task yet.
|
| 62 |
+
Add public verification summary
|