File size: 3,030 Bytes
53274d6
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
# IMD oracle protocol compliance: explicit format and checker feedback

Qwen3.8-27B passed 102/102 construction and 102/102 repair cases on its first attempt. No corrections were needed. The original run remains available for comparison.

This condition fixes an ambiguity in the original prompt: the input contains a recipe encoded as text, while the submission requires a JSON object. It supplies an explicit outer JSON schema, a synthetic example and instructions to parse the embedded recipe. The final grader and all 204 cases are unchanged.

This is a **development evaluation on reused public cases**. The prompt was revised after examining v1 failures, so improvements are not an independent estimate of generalization. It measures formatting and controlled repair of supplied data, not factual oracle answers or full IMD node performance. Construction and repair share 102 source questions; report the two tracks separately.

Each case gets one initial attempt and at most two corrections. Feedback comes from the pinned official checker and comparison with fields already visible in the input. Hidden reference labels never enter the model conversation. No output is automatically repaired by the harness. A transport failure ends the task and remains in the denominator.

Report first-attempt and final pass rates, recovered tasks, full task time, all-attempt tokens and tokens per successful task. Missing usage remains unknown. The prompt and settings were frozen before this rerun; no changes are made during it.

Settings: temperature 0, 4,096 output tokens per attempt, 120-second API timeout per attempt, concurrency 2. Qwen uses ModelScope with thinking disabled. The schema is supplied in the prompt; constrained decoding is not enabled. The outer schema does not enumerate all recipe-specific rules: the official checker and exact preservation of the supplied recipe are authoritative.

## Reproduce

Extract the reproduction archive into an empty directory. It includes the same v1 task and label files, the frozen v2 protocol, grader and runner. Python 3.9+ and Node 22+ are required.

```sh
python3 scripts/fetch_node_helpers.py
python3 -m unittest discover -s tests -p 'test_protocol_v2.py'
# Set QWEN_API_KEY in the environment or a local .env file.
python3 scripts/run_protocol_v2.py --model qwen-27b --run-tag reproduce
```

The helper fetcher requires the exact pinned upstream hashes. If the upstream helper changes, it refuses to continue. The runner checks the frozen task, label, prompt and harness hashes and refuses to overwrite a run directory.

The task data is still configuration `oracle_protocol_v1` on Hugging Face. This is a new evaluation condition, not an additional 204 independent tasks. Send only the task input and frozen prompt; reference artifacts are evaluator data.

License: unknown. Collected from the publicly accessible IMD API.

[Results and frozen prompt](../metadata/protocol-benchmark-v2.json) · [All attempts](results.jsonl) · [Reproduction archive](reproduce.tar.gz)