warden-eval

A small, hand-labelled evaluation set that pins Warden — a policy decision point for LLM traffic — to its documented behaviour.

42 cases in two configs:

Config	Rows	What each row asserts
classification	27	a text and the level the classifier must assign it
decision	15	a (subject, text, destination, purpose) and the verdict and rule_id the policy engine must return

Every case currently agrees with the engine and the Warden Policy Pack. This is not a benchmark and there is no leaderboard: it is a regression harness for a deterministic rule engine. There is exactly one correct answer per row, and the expected accuracy is 100%.

---

All data is synthetic

No real personal data appears anywhere in this dataset. Every value is a deliberate placeholder:

• email addresses use the RFC 2606 reserved example.com domain
• card numbers are the standard publicly documented test numbers (4111 1111 1111 1111)
• credentials are documented placeholders from vendor documentation (AKIAIOSFODNN7EXAMPLE is AWS's own example key id) or obvious dummies
• addresses are well-known public landmarks
• names are generic placeholders

The credential-shaped strings are non-functional by construction. They exist so a detector can be tested against them, and nothing more.

---

Usage

from datasets import load_dataset

classification = load_dataset("NagaYu/warden-eval", "classification", split="train")
decision       = load_dataset("NagaYu/warden-eval", "decision", split="train")

print(classification[0])
# {'id': 'cls-001', 'text': 'Nice weather today. What time is the meeting?',
#  'expected_level': 'public', 'note': 'plain chit-chat'}

classification.csv, decision.csv and dry_run_sample.csv hold the same data as flat CSVs. The last one is shaped for the Space's Dry Run tab, so you can upload it directly and see a candidate policy's impact across all 15 decision cases.

Running the harness

run_eval.py needs nothing beyond the standard library for local runs:

# against a local checkout of the engine
python run_eval.py --app /path/to/app.py --verbose

# against your own Gradio deployment (needs gradio_client)
python run_eval.py --space <user>/<space>

It exits non-zero on the first disagreement and prints what changed.

The published Warden Space is a static build: the engine runs in your browser and exposes no HTTP API, so --space does not apply to it. To check the deployed Space, open its Self-Test tab, which runs the same guarantees in place.

---

Fields

classification

Field	Meaning
id	stable case id (cls-NNN)
text	the input to classify
expected_level	public | internal | confidential | regulated
note	why this case exists — what it is actually testing

decision

Field	Meaning
id	stable case id (dec-NNN)
subject	the acting team
text	the text under consideration
destination	the destination id from the policy
purpose	the declared purpose
expected_verdict	allow | redact | reroute | deny
expected_rule_id	the rule that must produce that verdict
note	what this case is testing

expected_rule_id matters as much as the verdict. Getting deny from the wrong rule is still a bug: the whole point of the tool is that a decision can be traced to the rule that produced it.

---

Coverage

Level distribution (classification): 7 public, 6 internal, 6 confidential, 8 regulated. Verdict distribution (decision): 4 allow, 4 redact, 2 reroute, 5 deny.

Cases were chosen to cover the failure modes that actually bite, not to be numerous:

Case	What it defends
cls-004	16 digits that fail the Luhn check must not be read as a card number
cls-006	agenda must not match the sensitive term nda — word boundaries matter
cls-015	a single PII signal stays below the threshold; one IP address is not confidential
cls-022	a Luhn-valid test card carries weight 2 and reaches confidential on its own
cls-024, cls-033	full-width text must not evade detection — writing ＡＫＩＡ… is not a bypass
cls-037	a high-entropy string with no known prefix is still treated as a credential
dec-005	the HR rule covers providers A and B, so self-hosted falls through to redaction
dec-010	R000 outranks R001 on priority — credentials are denied before the HR rule applies
dec-013	a destination the policy never defined still matches on level

---

What this dataset does not tell you

It measures consistency with a declared policy, nothing else.

• It does not measure whether the policy is correct. That is a human judgement, and this dataset is deliberately silent on it.
• It does not measure compliance with any law or regulation. Warden makes no such claim, and neither does this dataset.
• It does not measure detector recall on real-world data. The cases are synthetic and chosen to pin specific behaviours; a detector that passes all 27 classification cases can still miss personal data in text that looks nothing like these.
• Passing does not mean a deployment is safe. It means the engine still behaves the way it behaved when these labels were written.

If you change the policy or the classifier configuration, some of these labels become wrong on purpose — the harness will tell you which, and it is then your job to decide whether the label or the change is the mistake.

License

Apache-2.0.