S3RL4 commited on
Commit
d6e4a89
·
verified ·
1 Parent(s): f6f6b68

Add verification kit README

Browse files

# Verification kit — sample-level checks

This small kit lets a buyer's researcher sanity-check the grader and answer hygiene without trusting RE-Lab. It runs on the public sample JSONL files; it does not require the paid full pack. The grader scripts import `binary_reasoning_v1.grading` from `environment/src/binary_reasoning_v1/grading.py` (`leak_check.py` does not), so include that file (or the Prime environment package) when downloading just the free sample.

## Run it

From the dataset root:

```bash
python verification_kit/gold.py sample_free.jsonl
python verification_kit/do_nothing.py sample_free.jsonl
python verification_kit/random_wrong.py sample_free.jsonl
python verification_kit/leak_check.py sample_free.jsonl
python verification_kit/grader_tamper.py sample_free.jsonl
```

If the 100-item open sample is present locally, point the same scripts at it:

```bash
python verification_kit/gold.py environment/src/binary_reasoning_v1/items.jsonl
python verification_kit/do_nothing.py environment/src/binary_reasoning_v1/items.jsonl
python verification_kit/random_wrong.py environment/src/binary_reasoning_v1/items.jsonl
python verification_kit/leak_check.py environment/src/binary_reasoning_v1/items.jsonl
python verification_kit/grader_tamper.py environment/src/binary_reasoning_v1/items.jsonl
```

## What each script checks

| Script | What it does | Expected result |
|---|---|---|
| `gold.py` | Builds a reply from the stored answer and grades it. | Deterministic items score `1.0`; rubric-judged items are skipped, not silently graded. |
| `do_nothing.py` | Grades an empty reply. | Deterministic items score `0`. |
| `random_wrong.py` | Builds a deterministic mutated wrong reply and grades it. | Deterministic items score `0`. It is not stochastic random guessing. |
| `leak_check.py` | Extracts exact hidden answer tokens where possible and searches the model-facing `input` field. | Reports checked-row coverage; no exact-token leaks on the public samples. |
| `grader_tamper.py` | Mutates the stored answer key and grades the original gold reply against it. | The original gold reply no longer scores, showing that grading depends on the answer key. |

## Measured results on this workspace (2026-10-30)

On `sample_free.jsonl`:
- `gold.py`, `do_nothing.py`, `random_wrong.py`, and `grader_tamper.py`: `4/4` deterministic items pass; `5` rubric-judged items are skipped.
- `leak_check.py`: `items=9 unique_ids=9 checked_rows=3 leaks=0`.

On the 100-item open sample at `environment/src/binary_reasoning_v1/items.jsonl`:
- `gold.py`, `do_nothing.py`, `random_wrong.py`, and `grader_tamper.py`: `84/84` deterministic items pass; `16` rubric-judged items are skipped.
- `leak_check.py`: `items=100 unique_ids=100 checked_rows=73 leaks=0`.

## Known limitations

- The current deterministic graders can accept some "shotgun" replies that contain the correct token among many distractors. This kit does not yet include a dedicated `shotgun.py` probe.
- `leak_check.py` scans the raw JSONL `input` field, not every rendered prompt variant. Rows with no extractable exact answer token are reported as unchecked.
- The full-pack answer-leak figure (`0/10,251`) comes from the existing full-pack audit, not from this public sample script.
- The `solvers/` directory contains RE-Lab-run reproduction scripts for internal hard-trial families. Several depend on local paths, WSL/Windows tools, pinned package digests, or host-specific loader behavior, so they are evidence rather than a portable buyer-runnable kit yet.

verification_kit/README.md/verification_kit/README.md ADDED
@@ -0,0 +1,53 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Verification kit — sample-level checks
2
+
3
+ This small kit lets a buyer's researcher sanity-check the grader and answer hygiene without trusting RE-Lab. It runs on the public sample JSONL files; it does not require the paid full pack. The grader scripts import `binary_reasoning_v1.grading` from `environment/src/binary_reasoning_v1/grading.py` (`leak_check.py` does not), so include that file (or the Prime environment package) when downloading just the free sample.
4
+
5
+ ## Run it
6
+
7
+ From the dataset root:
8
+
9
+ ```bash
10
+ python verification_kit/gold.py sample_free.jsonl
11
+ python verification_kit/do_nothing.py sample_free.jsonl
12
+ python verification_kit/random_wrong.py sample_free.jsonl
13
+ python verification_kit/leak_check.py sample_free.jsonl
14
+ python verification_kit/grader_tamper.py sample_free.jsonl
15
+ ```
16
+
17
+ If the 100-item open sample is present locally, point the same scripts at it:
18
+
19
+ ```bash
20
+ python verification_kit/gold.py environment/src/binary_reasoning_v1/items.jsonl
21
+ python verification_kit/do_nothing.py environment/src/binary_reasoning_v1/items.jsonl
22
+ python verification_kit/random_wrong.py environment/src/binary_reasoning_v1/items.jsonl
23
+ python verification_kit/leak_check.py environment/src/binary_reasoning_v1/items.jsonl
24
+ python verification_kit/grader_tamper.py environment/src/binary_reasoning_v1/items.jsonl
25
+ ```
26
+
27
+ ## What each script checks
28
+
29
+ | Script | What it does | Expected result |
30
+ |---|---|---|
31
+ | `gold.py` | Builds a reply from the stored answer and grades it. | Deterministic items score `1.0`; rubric-judged items are skipped, not silently graded. |
32
+ | `do_nothing.py` | Grades an empty reply. | Deterministic items score `0`. |
33
+ | `random_wrong.py` | Builds a deterministic mutated wrong reply and grades it. | Deterministic items score `0`. It is not stochastic random guessing. |
34
+ | `leak_check.py` | Extracts exact hidden answer tokens where possible and searches the model-facing `input` field. | Reports checked-row coverage; no exact-token leaks on the public samples. |
35
+ | `grader_tamper.py` | Mutates the stored answer key and grades the original gold reply against it. | The original gold reply no longer scores, showing that grading depends on the answer key. |
36
+
37
+ ## Measured results on this workspace (2026-10-30)
38
+
39
+ On `sample_free.jsonl`:
40
+ - `gold.py`, `do_nothing.py`, `random_wrong.py`, and `grader_tamper.py`: `4/4` deterministic items pass; `5` rubric-judged items are skipped.
41
+ - `leak_check.py`: `items=9 unique_ids=9 checked_rows=3 leaks=0`.
42
+
43
+ On the 100-item open sample at `environment/src/binary_reasoning_v1/items.jsonl`:
44
+ - `gold.py`, `do_nothing.py`, `random_wrong.py`, and `grader_tamper.py`: `84/84` deterministic items pass; `16` rubric-judged items are skipped.
45
+ - `leak_check.py`: `items=100 unique_ids=100 checked_rows=73 leaks=0`.
46
+
47
+ ## Known limitations
48
+
49
+ - The current deterministic graders can accept some "shotgun" replies that contain the correct token among many distractors. This kit does not yet include a dedicated `shotgun.py` probe.
50
+ - `leak_check.py` scans the raw JSONL `input` field, not every rendered prompt variant. Rows with no extractable exact answer token are reported as unchecked.
51
+ - The full-pack answer-leak figure (`0/10,251`) comes from the existing full-pack audit, not from this public sample script.
52
+ - The `solvers/` directory contains RE-Lab-run reproduction scripts for internal hard-trial families. Several depend on local paths, WSL/Windows tools, pinned package digests, or host-specific loader behavior, so they are evidence rather than a portable buyer-runnable kit yet.
53
+ Add verification kit README