README.md

---
pretty_name: Scientific Agent Verification Cascade
license: mit
language:
  - en
tags:
  - scientific-agents
  - multi-agent-systems
  - ai-safety
  - evaluation
  - provenance
  - verification
size_categories:
  - n<1K
configs:
  - config_name: synthetic_tasks
    data_files:
      - split: test
        path: data/synthetic_tasks.jsonl
  - config_name: modular_fixtures
    data_files:
      - split: test
        path: data/modular_fixtures.jsonl
  - config_name: corruption_fixtures
    data_files:
      - split: test
        path: data/corruption_fixtures.jsonl
---

# Scientific Agent Verification Cascade

Public evaluation fixtures and verified aggregate results for testing whether
scientific claims keep their source, meaning, uncertainty, and verification
requirements as they move between AI agents.

This dataset accompanies the
[Scientific Agent Verification Cascade](https://github.com/jang1563/sci-agent-verification-cascade)
codebase. [Version 0.2.0](https://github.com/jang1563/sci-agent-verification-cascade/releases/tag/v0.2.0)
contains synthetic evaluation data and aggregate-only results. It contains no
raw hosted-model response, private holdout identifier,
source-record identifier, credential, wet-lab protocol, or clinical guidance.

## Dataset Contents

| Config | Rows | Purpose |
| --- | ---: | --- |
| `synthetic_tasks` | 50 | Scientific claims with provenance, evidence status, and expected verification behavior |
| `modular_fixtures` | 10 | Balanced receipt-extraction and action fixtures |
| `corruption_fixtures` | 9 | Deterministic receipt and action corruptions with expected detectors |

The 50 synthetic tasks span 13 scientific domains. One task deliberately lacks
source provenance so that exclusion behavior can be tested. The modular
fixtures cover five actions: advance, stop, request more evidence, defer, and
flag.

## Load the Fixtures

```python
from datasets import load_dataset

tasks = load_dataset(
    "jang1563/sci-agent-verification-cascade",
    "synthetic_tasks",
    revision="v0.2.0",
    split="test",
)
modular = load_dataset(
    "jang1563/sci-agent-verification-cascade",
    "modular_fixtures",
    revision="v0.2.0",
    split="test",
)
corruptions = load_dataset(
    "jang1563/sci-agent-verification-cascade",
    "corruption_fixtures",
    revision="v0.2.0",
    split="test",
)
```

Field definitions are in [`SCHEMA.md`](SCHEMA.md).

## Aggregate Results

The `results/` directory contains seven public result files and a manifest with
their SHA-256 hashes.

| File | Evaluation |
| --- | --- |
| `two_provider_boundary.json` | 900-output matched handoff comparison |
| `depth_study.json` | 2,160-trace depth study |
| `modular_live_test.json` | 244-call synthetic modular test |
| `tool_grounded_live_test.json` | 122-call public-tool pilot and prompt audit |
| `tool_grounded_anthropic_expansion.json` | 427-call, 35-case Anthropic expansion |
| `semantic_verifier_calibration.json` | 75-call standalone semantic-check calibration |
| `integrated_semantic_canary.json` | 175-call integrated controller pilot and trigger diagnosis |

Download an aggregate directly:

```python
import json

from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="jang1563/sci-agent-verification-cascade",
    repo_type="dataset",
    revision="v0.2.0",
    filename="results/integrated_semantic_canary.json",
)
with open(path, encoding="utf-8") as stream:
    result = json.load(stream)
print(result["status"])
```

The latest integrated pilot completed all 175 planned responses and preserved
all 25 controlled handoffs. Its deterministic source-ID check caused all ten
corrections. The model-based semantic checker allowed all 50 messages it
reviewed and caused no correction, so the pilot did not demonstrate a semantic
intervention benefit. All five conditions were already correct on all five
actions, leaving no room to show an action improvement.

## Intended Use

Use these files to test claim receipts, provenance preservation, seeded
corruption detection, deterministic action gates, and evaluation pipelines for
scientific agents. The data are small evaluation fixtures, not a training
corpus or evidence that an autonomous system improves scientific discovery.

## Reproducibility

The GitHub repository provides the implementation, tests, local demo, and
release verifier. Run `savc verify` there to check the aggregate values and
manifest hashes. Raw hosted-model traces and private evaluator material are not
released, so the hosted-model studies cannot be regenerated trace by trace from
this dataset alone.

## License and Citation

The dataset and code are released under the MIT License.

```bibtex
@software{kim2026savc,
  author = {Kim, JangKeun},
  title = {Scientific Agent Verification Cascade},
  year = {2026},
  version = {0.2.0},
  url = {https://github.com/jang1563/sci-agent-verification-cascade}
}
```