Artifact-centered Claim-aware Observability for Autonomous Scientific Agents
Paper • 2608.18312 • Published • 1
submission_id stringclasses 5
values | task_id stringclasses 4
values | claim stringclasses 5
values | status stringclasses 2
values | execution unknown | verifier dict | provenance dict | integrity dict | truth_boundary stringclasses 2
values |
|---|---|---|---|---|---|---|---|---|
OBS-EV-001-COMMUNITY | PROOF-PRIME | 173 is a prime number. | PROVEN | {
"actual_inference": false,
"provider": "sovereign-evidence-observatory",
"receipt_sha256": "af6fba8918accaee8972f7a8ec9cf36c8844871df5c50d6cdb1c0ab1554fdfc4",
"source_revision": "d2f89c47b0286abef0f5f805b04f6e1eb90e2c57"
} | {
"kind": "deterministic-primality-64bit",
"version": "2.1.3",
"result": true,
"expression": "VerificationTest[PrimeQ[173], True]"
} | {
"source_url": "https://huggingface.co/spaces/Thorsu/sovereign-evidence-observatory",
"source_type": "public-space",
"dataset_revision": "d2f89c47b0286abef0f5f805b04f6e1eb90e2c57"
} | {
"bundle_sha256": "be89849d55d657ac102d23d02aae097ce58c5cf2637f08930dc2857948c01fa2"
} | This record establishes only the reported verification event for the specific mathematical claim. |
OBS-EV-002-COMMUNITY | PROOF-PRIME | 174 is a prime number. | CONTRADICTED | {
"actual_inference": false,
"provider": "sovereign-evidence-observatory",
"receipt_sha256": "b14bc80c5832e65cfd46840e14bc053465a096fdf1475216e959e0ddfb7132b2",
"source_revision": "d2f89c47b0286abef0f5f805b04f6e1eb90e2c57"
} | {
"kind": "deterministic-primality-64bit",
"version": "2.1.3",
"result": false,
"expression": "VerificationTest[PrimeQ[174], True]"
} | {
"source_url": "https://huggingface.co/spaces/Thorsu/sovereign-evidence-observatory",
"source_type": "public-space",
"dataset_revision": "d2f89c47b0286abef0f5f805b04f6e1eb90e2c57"
} | {
"bundle_sha256": "42f492e068d20dd538298782d9347c7e260d46145fd25ba478fc8c530375248e"
} | This record establishes only the reported verification event for the specific mathematical claim. |
REAL-QWEN-EBE-001-20261001 | EBE-001 | The agent returned the expected output for EBE-001. | PROVEN | {
"actual_inference": true,
"provider": "featherless-ai",
"receipt_sha256": "8f6b2dec49cd060b432db177dd107934812e2b327c1744514a69eefad7a1fda8",
"source_revision": "19ba9f19b6b4c23234bd57f0952cc5502bb55810",
"model": "Qwen/Qwen2.5-0.5B-Instruct",
"model_revision": "7ae557604adf67be50417f59c2c2f167def9a775"
} | {
"kind": "literal_match",
"version": "runner-1.0",
"result": true,
"expression": "exact_match"
} | {
"source_url": "https://huggingface.co/Qwen/Qwen2.5-0.5B-Instruct",
"source_type": "local-agent-runner",
"dataset_revision": "19ba9f19b6b4c23234bd57f0952cc5502bb55810"
} | {
"bundle_sha256": "42ac8de0c6f38f2df3fd1d2fde549a57c348bb847d61c3583b9dc9efe874a110"
} | This record establishes only the reported execution and deterministic verification event. |
REAL-QWEN-EBE-004-20261001 | EBE-004 | The agent returned the expected output for EBE-004. | PROVEN | {
"actual_inference": true,
"provider": "featherless-ai",
"receipt_sha256": "17b7d0b19c4535f3bbc57d16a4e57260e48e9a4d3017dfdabf91b622ba5dccb8",
"source_revision": "b340b6ecbac72d056ee85ad5e2e291a308c4c901",
"model": "Qwen/Qwen2.5-0.5B-Instruct",
"model_revision": "7ae557604adf67be50417f59c2c2f167def9a775"
} | {
"kind": "exact_multiline_match",
"version": "runner-1.0",
"result": true,
"expression": "exact_match"
} | {
"source_url": "https://huggingface.co/Qwen/Qwen2.5-0.5B-Instruct",
"source_type": "local-agent-runner",
"dataset_revision": "b340b6ecbac72d056ee85ad5e2e291a308c4c901"
} | {
"bundle_sha256": "f5cb9fec5e60a1eaefa7af000ecd9f5be0862e562d12d21c24b3e1a6aa9e058d"
} | This record establishes only the reported execution and deterministic verification event. |
REAL-QWEN-EBE-002-20261001 | EBE-002 | The agent returned the expected output for EBE-002. | CONTRADICTED | {
"actual_inference": true,
"provider": "featherless-ai",
"receipt_sha256": "711de4d17f2ce07878b0fe4ee73bfdce6b787e5a2057386de55ddc9b0f7d19a3",
"source_revision": "b340b6ecbac72d056ee85ad5e2e291a308c4c901",
"model": "Qwen/Qwen2.5-0.5B-Instruct",
"model_revision": "7ae557604adf67be50417f59c2c2f167def9a775"
} | {
"kind": "exact_arithmetic",
"version": "runner-1.0",
"result": false,
"expression": "exact_match"
} | {
"source_url": "https://huggingface.co/Qwen/Qwen2.5-0.5B-Instruct",
"source_type": "local-agent-runner",
"dataset_revision": "b340b6ecbac72d056ee85ad5e2e291a308c4c901"
} | {
"bundle_sha256": "75c54727be2cff58e39a0da1e4c7bda98fa601015a1eaa081d8467296b21e15e"
} | This record establishes only the reported execution and deterministic verification event. |
Evidence-Bound Agent Evals
A community protocol for evidence-bound AI evaluation.
Instead of only model-to-score results, each contribution can bind:
claim -> execution evidence -> verifier -> provenance -> integrity
What this is
This project treats evidence itself as a first-class evaluation object.
- Real execution receipts are preserved.
- Contradictions, route errors and abstentions are valid outcomes.
- Model and provider identities stay explicit.
- Independent verifier results can be attached.
- Revision and hash provenance are part of the record.
- The protocol does not produce an intelligence or provider-quality leaderboard.
Dataset Viewer
The public default/test split is backed only by data/community_records_v1.jsonl, the protocol-conformant community stream.
Other data/ and examples/ files are protocol artifacts, challenge material, historical snapshots, or reproducible reference runs.
Challenge Pack
protocol/tasks-v1.jsonl contains eight deterministic challenge families:
literal output, arithmetic, structured JSON, bilingual exact output, bounded abstention, provenance relation, runtime identity and truth-boundary handling.
Adapter-neutral runner
tools/run_task.py executes a challenge through an external or local adapter, captures stdout/stderr and exit code, hashes the canonical receipt, applies deterministic verification, and emits a protocol submission.
Non-zero adapter exit codes are classified as ROUTE_ERROR, not semantic contradiction.
examples/reference_agent.py is a deterministic local adapter. examples/hf_featherless_qwen_adapter.py is the real HF/Featherless adapter used for the first external inference record.
First real external inference
REAL-QWEN-EBE-001-20261001 is a real actual_inference=true community record for EBE-001.
- Model: Qwen/Qwen2.5-0.5B-Instruct
- Model revision: 7ae557604adf67be50417f59c2c2f167def9a775
- Provider: featherless-ai
- Result: exact required literal
- Deterministic verifier:
literal_match/true - Receipt SHA-256: 8f6b2dec49cd060b432db177dd107934812e2b327c1744514a69eefad7a1fda8
- Bundle SHA-256: 42ac8de0c6f38f2df3fd1d2fde549a57c348bb847d61c3583b9dc9efe874a110
The raw receipt and submission are published under examples/. The record is also present in data/community_records_v1.jsonl.
Reference runner evidence
A deterministic reference run covers all eight challenge tasks plus one deliberate route-error case.
- 8 normal reference tasks verified
- 1 route-error classification verified
- raw receipts published under examples/
- protocol submissions published under examples/
- run manifest published under examples/
These reference artifacts are examples and are intentionally outside the public Dataset Viewer split.
Contribution flow
- Run a task from the challenge pack.
- Keep the exact execution receipt and source/model revision.
- Add verifier evidence and an explicit truth boundary.
- Use tools/run_task.py or build an equivalent adapter integration.
- Run python tools/validate_submission.py your_submission.json.
- Run python tools/validate_dataset.py.
- Run python tools/selftest.py.
- Open a Hugging Face pull request against this dataset.
Automated PR Gate
A Hugging Face Scheduled Job polls open dataset pull requests every 10 minutes.
The poller loads the versioned gate, reads each PR revision through the Hub revision reference, and runs the dataset gate plus selftest against that exact revision.
If any open PR fails, the poller exits with code 1. With no open PRs it returns PASS_NO_OPEN_PRS.
Current job: 6abd9d7c404719ba37614ce4, schedule */10 * * * *, active.
HF Webhook creation returned HTTP 403 for the connected account. The current automatic gate is therefore polling-based and fail-closed; it does not claim automatic PR comments.
Integrity
Every contribution should remain reconstructable from public artifacts or clearly labeled external references.
The canonical bundle hash covers submission identity, task, claim, execution, verifier, provenance and truth boundary.
Local gates
tools/validate_submission.py validates one proposed record.
tools/validate_dataset.py validates the complete public community stream, including unique submission IDs and status/verifier alignment.
tools/selftest.py deterministically checks schema, canonical example, tamper rejection, starter records, runner normal path, route-error path, and presence of a real external inference record.
tools/check_pr.py validates a specific public HF Pull Request by discussion number.
Research context
- arXiv:2607.28374 LEDGERMIND
- arXiv:2608.18312 Artifact-Centered Claim-Aware Observability
- arXiv:2606.18874 Research Harness
Design principle
Evidence is not a score.
A receipt proves an observed event. A verifier proves what that verifier checked. The record should never claim more than its evidence supports.
Companion Space
See the interactive Evidence-Bound Agent Evals Explorer at https://huggingface.co/spaces/Thorsu/evidence-bound-agent-evals-explorer for the challenge pack, reference evidence, community stream and submission checker.
- Downloads last month
- 609