English · 中文
An evidence-constrained agent workflow for turning project materials into auditable deliverables.
Evidence Runtime is built for work that starts with source material and ends with a proposal. It manages a project library, checks whether a fixed document contract is supported, prepares constrained writing tasks, receives writer responses, runs a claim-audit gate, and renders DOCX output.
It is neither a personal knowledge base nor an embedded language model. It can be used as a standalone Evidence Agent: the repository supplies the workflow, templates, state machine, and deterministic checks; a host agent or a human performs semantic writing and auditing.
| Stage | Runtime | Agent or human |
|---|---|---|
| Intake and library | Builds a ProjectSpec; stores project-scoped PDF, DOCX, MD, and TXT files | Confirms task parameters and data permissions |
| Evidence readiness | Retrieves, locates verbatim spans, applies rules, and marks slots filled / weak / missing | Resolves missing material and business choices |
| Generation | Emits constrained section tasks; validates and receives responses | Writes the requested sections |
| Claim Audit | Splits sentence-level units; checks coverage; recomputes reference, content, and jurisdiction checks | Submits a claim for each factual unit and assigns its support judgment |
| Delivery | Renders drafts or gated DOCX through one public entry point and verifies the document | Performs final business review and approval |
The package currently ships a fixed evidence contract and proposal templates for Chinese public-sector and enterprise solution documents. Contracts, crosswalks, and section specs are usable as-is and extensible through the same protocols.
The system cannot stop a writer from producing an incorrect sentence. It prevents unaudited content from being admitted as a deliverable:
- every non-placeholder factual unit must be covered by an audit claim;
- references must exist, with consistent evidence content and jurisdiction;
- an audit must match the current draft, so an old report cannot approve edited text;
- cached booleans such as
acceptedordelivery.admittedcannot bypass recomputation; - missing evidence remains an explicit
【待补充】placeholder.
The default integrity mode detects stale reports, accidental edits, and incomplete audits. It does not
stop a principal with write access to the entire workspace from coordinating all artifacts and self-authorizing.
Strong authorization requires an independent auditor identity and signing infrastructure, which is not yet
implemented. See Security and data boundaries.
Requirements: Python 3.11 or 3.12 and ripgrep.
python -m pip install -e ".[test]"
python -m pytest -q
bash examples/demo-proposal/run_chain.shThe synthetic demo runs the real intake, library, gap analysis, two-pass generation, Claim Audit, and DOCX
rendering path. It intentionally lacks some evidence and therefore produces a marked SHAREABLE_DRAFT.
demo-proposal-deliverable shows the fully supported
DELIVERABLE path. These demos verify workflow behavior, not retrieval accuracy.
The package installs six commands:
evidence-intake
evidence-library
evidence-gap-check
evidence-generate
evidence-audit
evidence-render
Their normal order is:
Task → ProjectSpec → Project library → Gap Report / GenerationBrief
→ Writer response → Claim Audit → Draft or Deliverable DOCX
See the usage guide for complete commands, outputs, and the two-pass handoff.
solution-drafter can act as a specialized Writer Agent: it consumes section tasks from Evidence Runtime and returns structured responses. It is not a runtime dependency; Claude Code, Codex, another agent, or a human can fill the same role.
Both projects remain independently useful today. A future unified product should compose them through the stable task/response protocol instead of reducing Evidence Runtime to an internal library.
Version 0.1.0 is Alpha. The main workflow and delivery gate are implemented and covered end to end. The project
is suitable for evaluation and controlled pilots, but it is not yet a general enterprise content platform.
Current limitations include a fixed document contract, no OCR, no embedded semantic model, no Web/API service,
and no signed authorization mode. See status and limitations.
- Documentation index
- Usage
- Architecture and boundaries
- Security and data boundaries
- Evaluation
- Design decisions
- Changelog
- Third-party notices
Do not commit customer material, project libraries, runtime output, internal acceptance records, or development
diaries. .library/, .local/, .internal/, and out/ are ignored by default. The public repository contains
synthetic demos, reproducible tests, license-reviewed benchmark records, and user-facing documentation.
If an older clone ever contained private material, deleting it from the current tree is insufficient: review and clean the reachable Git history before publication.
The code and project documentation use the MIT License. Third-party datasets retain their own terms; see THIRD_PARTY_NOTICES.md.