Hugging Face's logo

Datasets:
Quome
/
tamperbench

Tasks:
Other
Languages:
English
Size:
10K<n<100K
Tags:
clinical-ai
llm-security
ai-agents
healthcare
haarf
quokkaguard
License:
Dataset card Files
xet
 Community
tamperbench / README.md
jimschwoebel's picture
jimschwoebel
TamperBench (2026-10-02)
3b69067 verified
preview code
|
Raw History Blame Contribute Delete
9.77 kB
metadata 
license: apache-2.0
pretty_name: TamperBench
language:
  - en
tags:
  - clinical-ai
  - llm-security
  - ai-agents
  - healthcare
  - haarf
  - quokkaguard
  - synthetic
size_categories:
  - 10K<n<100K
task_categories:
  - other

TamperBench

A blake3-chained, ed25519-signed, externally anchored audit ledger for clinical AI agents.

TamperBench is the benchmark corpus for Tamper-Evident Action Provenance for Confidential Clinical AI Agents, the quledger module of the QUOKKAGUARD program. It ships with the quledger repository, which contains the qfire gateway layer under test, the experiment harness, and the paper.

Deterministic (seed 42), fully synthetic clinical action streams plus mutated chained audit logs spanning six audit-tampering classes (TA1 field edit, TA2 delete and truncate, TA3 reorder, TA4 forge, TA6 rollback) with labels.json ground truth; the E1 corpus is 26 tampered logs derived from a 200-entry signed and anchored fixture.

All data are synthetic. No real patient data or protected health information (PHI) is included; clinical content is generated from templates with fixed seeds.

Files

File Size Rows
FORMAT.md 2 KB
fixture/anchors.jsonl 8 KB 30 rows
fixture/audit.jsonl 1 MB 3,001 rows
labels.json 3 KB
streams/stream-0.jsonl 202 KB 1,000 rows
streams/stream-1.jsonl 201 KB 1,000 rows
streams/stream-2.jsonl 202 KB 1,000 rows
streams/stream-3.jsonl 202 KB 1,000 rows
streams/stream-4.jsonl 201 KB 1,000 rows
streams/stream-5.jsonl 202 KB 1,000 rows
streams/stream-6.jsonl 202 KB 1,000 rows
streams/stream-7.jsonl 202 KB 1,000 rows
streams/stream-8.jsonl 202 KB 1,000 rows
streams/stream-9.jsonl 202 KB 1,000 rows
tampered/ta1_field_edit-0.jsonl 1 MB 3,001 rows
tampered/ta1_field_edit-1.jsonl 1 MB 3,001 rows
tampered/ta1_field_edit-2.jsonl 1 MB 3,001 rows
tampered/ta1_field_edit-3.jsonl 1 MB 3,001 rows
tampered/ta1_field_edit-4.jsonl 1 MB 3,001 rows
tampered/ta2_delete-0.jsonl 1 MB 3,000 rows
tampered/ta2_delete-1.jsonl 1 MB 3,000 rows
tampered/ta2_delete-2.jsonl 1 MB 3,000 rows
tampered/ta2_delete-3.jsonl 1 MB 3,000 rows
tampered/ta2_delete-4.jsonl 1 MB 3,000 rows
tampered/ta2_truncate-0.jsonl 163 KB 358 rows
tampered/ta2_truncate-1.jsonl 1 MB 2,420 rows
tampered/ta2_truncate-2.jsonl 788 KB 1,730 rows
tampered/ta2_truncate-3.jsonl 60 KB 132 rows
tampered/ta2_truncate-4.jsonl 56 KB 124 rows
tampered/ta3_reorder-0.jsonl 1 MB 3,001 rows
tampered/ta3_reorder-1.jsonl 1 MB 3,001 rows
tampered/ta3_reorder-2.jsonl 1 MB 3,001 rows
tampered/ta3_reorder-3.jsonl 1 MB 3,001 rows
tampered/ta3_reorder-4.jsonl 1 MB 3,001 rows
tampered/ta4_forge-0.jsonl 1 MB 3,002 rows
tampered/ta4_forge-1.jsonl 1 MB 3,002 rows
tampered/ta4_forge-2.jsonl 1 MB 3,002 rows
tampered/ta4_forge-3.jsonl 1 MB 3,002 rows
tampered/ta4_forge-4.jsonl 1 MB 3,002 rows
tampered/ta6_rollback-0.anchors.jsonl 8 KB 30 rows
tampered/ta6_rollback-0.jsonl 683 KB 1,500 rows

Record schemas

  • fixture/anchors.jsonl: seq, merkle_root, ts, sig
  • fixture/audit.jsonl: seq, ts, kind, body, prev_hash, this_hash, sig
  • streams/stream-0.jsonl: action, actor, decision, detail, event_id, patient_ref, stream, ts_epoch, unit
  • streams/stream-1.jsonl: action, actor, decision, detail, event_id, patient_ref, stream, ts_epoch, unit
  • streams/stream-2.jsonl: action, actor, decision, detail, event_id, patient_ref, stream, ts_epoch, unit
  • streams/stream-3.jsonl: action, actor, decision, detail, event_id, patient_ref, stream, ts_epoch, unit
  • streams/stream-4.jsonl: action, actor, decision, detail, event_id, patient_ref, stream, ts_epoch, unit
  • streams/stream-5.jsonl: action, actor, decision, detail, event_id, patient_ref, stream, ts_epoch, unit
  • streams/stream-6.jsonl: action, actor, decision, detail, event_id, patient_ref, stream, ts_epoch, unit
  • streams/stream-7.jsonl: action, actor, decision, detail, event_id, patient_ref, stream, ts_epoch, unit
  • streams/stream-8.jsonl: action, actor, decision, detail, event_id, patient_ref, stream, ts_epoch, unit
  • streams/stream-9.jsonl: action, actor, decision, detail, event_id, patient_ref, stream, ts_epoch, unit
  • tampered/ta1_field_edit-0.jsonl: seq, ts, kind, body, prev_hash, this_hash, sig
  • tampered/ta1_field_edit-1.jsonl: seq, ts, kind, body, prev_hash, this_hash, sig
  • tampered/ta1_field_edit-2.jsonl: seq, ts, kind, body, prev_hash, this_hash, sig
  • tampered/ta1_field_edit-3.jsonl: seq, ts, kind, body, prev_hash, this_hash, sig
  • tampered/ta1_field_edit-4.jsonl: seq, ts, kind, body, prev_hash, this_hash, sig
  • tampered/ta2_delete-0.jsonl: seq, ts, kind, body, prev_hash, this_hash, sig
  • tampered/ta2_delete-1.jsonl: seq, ts, kind, body, prev_hash, this_hash, sig
  • tampered/ta2_delete-2.jsonl: seq, ts, kind, body, prev_hash, this_hash, sig
  • tampered/ta2_delete-3.jsonl: seq, ts, kind, body, prev_hash, this_hash, sig
  • tampered/ta2_delete-4.jsonl: seq, ts, kind, body, prev_hash, this_hash, sig
  • tampered/ta2_truncate-0.jsonl: seq, ts, kind, body, prev_hash, this_hash, sig
  • tampered/ta2_truncate-1.jsonl: seq, ts, kind, body, prev_hash, this_hash, sig
  • tampered/ta2_truncate-2.jsonl: seq, ts, kind, body, prev_hash, this_hash, sig
  • tampered/ta2_truncate-3.jsonl: seq, ts, kind, body, prev_hash, this_hash, sig
  • tampered/ta2_truncate-4.jsonl: seq, ts, kind, body, prev_hash, this_hash, sig
  • tampered/ta3_reorder-0.jsonl: seq, ts, kind, body, prev_hash, this_hash, sig
  • tampered/ta3_reorder-1.jsonl: seq, ts, kind, body, prev_hash, this_hash, sig
  • tampered/ta3_reorder-2.jsonl: seq, ts, kind, body, prev_hash, this_hash, sig
  • tampered/ta3_reorder-3.jsonl: seq, ts, kind, body, prev_hash, this_hash, sig
  • tampered/ta3_reorder-4.jsonl: seq, ts, kind, body, prev_hash, this_hash, sig
  • tampered/ta4_forge-0.jsonl: seq, ts, kind, body, prev_hash, this_hash, sig
  • tampered/ta4_forge-1.jsonl: seq, ts, kind, body, prev_hash, this_hash, sig
  • tampered/ta4_forge-2.jsonl: seq, ts, kind, body, prev_hash, this_hash, sig
  • tampered/ta4_forge-3.jsonl: seq, ts, kind, body, prev_hash, this_hash, sig
  • tampered/ta4_forge-4.jsonl: seq, ts, kind, body, prev_hash, this_hash, sig
  • tampered/ta6_rollback-0.anchors.jsonl: seq, merkle_root, ts, sig
  • tampered/ta6_rollback-0.jsonl: seq, ts, kind, body, prev_hash, this_hash, sig

How it was generated

The corpus is produced by the generator in the paper repository and is fully deterministic (fixed seeds), so it can be regenerated byte-for-byte.

# from the quledger repository root (deterministic seeds)
cargo build --release && cargo test --release --test audit_tamper -- --ignored --exact export_fixture && python3 scripts/003-tamper-audit/tamper.py --log datasets/003-tamper-audit/fixture/audit.jsonl --anchors datasets/003-tamper-audit/fixture/anchors.jsonl --out datasets/003-tamper-audit/tampered --per-class 5 --seed 42 && python3 scripts/003-tamper-audit/gen.py --out datasets/003-tamper-audit/streams --streams 10 --events 1000 --seed 42

Intended use

Evaluating the audit enforcement layer of a clinical-agent security gateway (HAARF control C2): An adversarial agent, a compromised host process, or a privileged operator can silently rewrite, drop, reorder, or roll back a clinical agent's action log, defeating the traceability HAARF C2 requires.

The experiments that consume it (E-series in the paper) are reproduced from the repository:

git clone https://github.com/quome-cloud/quledger
cd quledger
cargo build --release

then follow the Reproduce the experiments section of its README.

Citation

This benchmark was built to evaluate a control of the Healthcare AI Agents Regulatory Framework (HAARF), the source framework for the QUOKKAGUARD program. Please cite both the paper and HAARF:

@unpublished{schwoebel2026quledger,
  author = {Schwoebel, James},
  title  = {Tamper-Evident Action Provenance for Confidential Clinical AI Agents},
  note   = {Preprint. Quome, QUOKKAGUARD program (quledger module)},
  year   = {2026},
  url    = {https://github.com/quome-cloud/quledger}
}

@unpublished{schwoebel2026haarf,
  author = {Schwoebel, Jim and Frasch, Martin and Spalding, Art and Sewell, Ed and Englert, Phil and Halpert, Ben and Overbay, Collin and Semenec, Ingrida and Shor, Joel},
  title  = {{HAARF}: Healthcare {AI} agents regulatory framework --- a comprehensive security verification standard for autonomous {AI} systems in clinical environments},
  note   = {medRxiv Preprint},
  year   = {2026},
  month  = {April},
  doi    = {10.64898/2026.04.09.26350519},
  url    = {https://www.medrxiv.org/content/10.64898/2026.04.09.26350519v1}
}

License

Apache License 2.0. Copyright (c) 2026 Quome, Inc.