Evidence Capture Agent Blueprint

English | 简体中文

A platform-neutral reference design for a conversational agent that turns an evidence-capture request into a reviewable, auditable workflow.

This repository contains architecture, state-machine, tool-contract, and testing patterns only. It does not connect to a real evidence platform, create legal evidence, issue certificates, or provide legal advice.

Why this blueprint exists

Evidence workflows often fail when a chat model jumps directly from a vague request to an irreversible action. This blueprint separates conversation, planning, explicit consent, isolated execution, integrity recording, and delivery.

Reference flow

1. Understand the user's preservation goal.
2. Select a capture strategy.
3. Preview scope, cost inputs, and expected outputs.
4. Obtain explicit confirmation.
5. Prepare an isolated execution environment.
6. Execute capture through allowlisted tools.
7. Finalize artifacts and return a verifiable delivery manifest.

Repository map

• docs/architecture.md: trust boundaries and runtime components.
• docs/workflow.md: state machine, transitions, and failure handling.
• prompts/system-policy.md: minimal policy prompt for the orchestrator.
• schemas/capture-case.schema.json: portable case-state schema.
• schemas/tool-result.schema.json: normalized tool result envelope.
• examples/demo-case.json: synthetic example with no real user or service data.
• tests/scenarios.md: acceptance and adversarial test cases.

Chinese translations are available through README.zh-CN.md and the linked *.zh-CN.md documents. JSON Schema field names remain in English so implementations use one stable contract.

Design principles

• Chat is the interface, not the execution authority.
• The model proposes; the backend validates and executes.
• No charge, credential use, or irreversible action without explicit consent.
• Capture runs in an isolated environment, not on the user's local browser.
• Every tool call is allowlisted, idempotent where possible, and auditable.
• Demo mode never invents an artifact ID, certificate, digest, or verification URL.

Non-goals

• Legal conclusions or litigation strategy.
• Circumventing access controls, CAPTCHAs, paywalls, or platform restrictions.
• Collecting credentials in chat.
• Covert monitoring or collection without authorization.
• A production-ready evidence service.

Quick start

Use the documents as a design review pack. Implement adapters behind the contracts only after privacy, security, legal, and platform-specific review.

License

MIT. See LICENSE.