DutyGate

Catch the legal obligations hiding in ordinary support messages before your chatbot replies.

"pls stop texting me" is an opt-out. "I want everything you have on me" is a privacy access request. "My lawyer will hear about this" is a legal threat. Topic triage files these under billing or angry customer, keyword lists miss the way people actually write, and an LLM bot that answers "Done, you're unsubscribed!" when nothing happened creates the liability itself.

DutyGate is a small, chatbot-agnostic gate placed in front of the reply step. It reads every inbound message, asks TypeSafe Jev a batch of yes/no questions in one call, applies versioned rules from a YAML policy pack, and tells your code what to do:

inbound message ──► DUTYGATE ──► continue ──► bot replies normally
                        ├── route  ──► holding reply + case in the right queue
                        └── review ──► bot may reply; a human takes a look

• Flags and routes, never acts. It does not unsubscribe, delete, refund or reply on its own. Your people and workflows do that.
• Recall-first. A miss can cost a statutory deadline; a false alarm costs a reviewer a few seconds. Low-confidence signals go to review, not to continue.
• Fails safe. If TypeSafe is down, slow, or returns something malformed, you get a review decision with an error code, never an exception or a dropped message.
• Auditable. Rules are versioned YAML, testable offline, with a published evaluation harness and a cross-language conformance suite.

Not legal advice. Which categories matter, their deadlines, and how they are routed depend on your jurisdiction and business. Have counsel review the packs and your holding replies before production use.

Quickstart

Offline, with no API key. The reference packs, their sample datasets and keyword lists ship inside the package, so they can be used by name. The keyword backend here is the naive baseline, which exists only for evaluation and demos.

pip install 'dutygate[server]'

dutygate validate legal-triggers outbound-claims
dutygate run legal-triggers --backend keyword --state "pls stop texting me"
dutygate eval legal-triggers --backend keyword

With TypeSafe Jev, the real backend:

export TYPESAFE_API_KEY=...        # from typesafe.ai
export DUTYGATE_JEV_MODEL=jev-1.13.0   # pin the model in production
dutygate run legal-triggers --state "I want everything you have on me"
dutygate eval legal-triggers    # scores Jev on the bundled sample dataset

Make it yours

Copy a reference pack, with its sample dataset and keyword list, into your project and edit it. No code is involved: questions, thresholds, queues and new categories are all YAML.

dutygate init legal-triggers      # writes legal-triggers.yaml, .dataset.jsonl, .keywords.yaml
# edit legal-triggers.yaml (and give it your own `name`), add rows to the dataset
dutygate validate legal-triggers.yaml
dutygate eval legal-triggers.yaml legal-triggers.dataset.jsonl --save-answers answers.jsonl
dutygate eval legal-triggers.yaml legal-triggers.dataset.jsonl --replay answers.jsonl --sweep

Then point your bot at your file: Gate.from_pack("legal-triggers.yaml"). To measure it on real traffic, add your own labeled messages to the dataset; see docs/labeling.md. The format is in SPEC.md.

Use it from your bot

Python

from dutygate import Gate, default_holding_reply

gate = Gate.from_pack("packs/legal-triggers.yaml")  # TypeSafe Jev, configured from env

decision = await gate.check_async(message, conversation_id=cid, channel="sms")
if decision.action == "route":
    await create_case(decision.flags, message)  # your ticketing / queue
    reply = default_holding_reply(decision)  # neutral; no bot improvisation
elif decision.action == "review":
    await create_review_task(decision.flags, message)
    if decision.error:  # the gate could not judge it
        await enqueue_rescan(message)
    reply = await bot.reply(message)
else:
    reply = await bot.reply(message)

GatedHandler (in dutygate.adapters.webhook) wraps exactly this logic for any web framework, with_legal_gate does the same for LangChain runnables, and gate_node / outbound_node add DutyGate to a LangGraph graph. See docs/adapters.md.

Any language: the HTTP sidecar

DUTYGATE_SIDECAR_KEYS=change-me TYPESAFE_API_KEY=... \
  dutygate serve packs/legal-triggers.yaml packs/outbound-claims.yaml --host 0.0.0.0

curl -H 'Authorization: Bearer change-me' -H 'content-type: application/json' \
  -d '{"message": "pls stop texting me", "channel": "sms"}' localhost:8080/v1/gate

Or use the Docker image, the TypeScript client (dutygate-client), or any no-code tool (n8n / Zapier recipe). See docs/sidecar.md.

The decision

{
  "id": "dec_4f1c2a9e8b7d4c3fa1e2b3c4d5e6f708",
  "action": "route",
  "primary": "privacy_request",
  "flags": [
    {"category": "privacy_request", "rule_id": "privacy-request", "queue": "privacy",
     "priority": "high", "confidence": 0.93, "level": "confident"},
    {"category": "opt_out", "rule_id": "opt-out", "queue": "compliance",
     "priority": "high", "confidence": 0.31, "level": "grey"}
  ],
  "error": null,
  "policy": {"name": "legal-triggers", "version": "0.1.0"},
  "backend": {"name": "typesafe-jev", "model": "jev-1.13.0"},
  "conversation_id": "c_123",
  "latency_ms": 212
}

primary is the first confident flag in the pack's rule order (or the first grey one). Error codes and every field are specified in SPEC.md.

What it detects

One message can raise several flags ("erase my data and stop the emails"). Queues and priorities are examples: map them to your own systems. Packs are plain YAML, so you can add categories or tighten questions without touching code (SPEC.md).

Surfaces

Every surface must reproduce conformance/cases.json.

Evaluation

The repo ships synthetic labeled datasets, a deliberately naive keyword baseline, and a harness that scores any backend, saves raw answers, and sweeps thresholds offline.

On the bundled legal-triggers set, the keyword baseline catches 48% of triggers (0% of non-English ones) and false-alarms on 13% of clean messages. The datasets were written by the same author as the baseline, so they are a sanity check, not a benchmark: measure on your own traffic before you rely on this. See evals/README.md.

If Jev does not beat a keyword list on your data, do not pay for it.

Status

Docs

• SPEC.md: pack format, evaluation algorithm, decision contract (normative)
• docs/integration.md: host contract, holding replies, failure behavior, latency patterns
• docs/sidecar.md: HTTP API, auth, metrics, audit log, Docker
• docs/adapters.md: GatedHandler, LangChain, LangGraph
• docs/privacy.md: what leaves your system, redaction, prompt injection
• docs/labeling.md: building a real evaluation set
• CONTRIBUTING.md · SECURITY.md · CHANGELOG.md

License

MIT. See LICENSE.