Evidence Check for DeepSeek Harness

See what the evidence supports—not just what an AI agent says happened.

Evidence Check is a developer-focused, evidence-aware review tool from MindForge Guard Labs. It examines structured synthetic Claim–Evidence records and produces traceable, bounded findings using a deterministic evaluation engine.

中文说明 · Examples · Privacy · Security

Why Evidence Check?

Agent workflows can report a successful action while the available records tell a narrower story. A tool response, a before/after snapshot and a coverage claim do not necessarily prove the same thing.

Evidence Check helps developers explore those gaps before relying on an agent's narrative:

• Evidence-linked conclusions. Each claim gets a verdict, a written reason and references to the supporting or conflicting records.
• Bounded judgments. Distinguish SUPPORTED_LIMITED, UNVERIFIED and CONTRADICTED instead of collapsing uncertainty into a pass/fail label.
• Deterministic review. A pinned rules engine evaluates the evidence; a language model does not invent the verdict.
• Local tool execution. The plugin itself has no telemetry, automatic uploads or runtime filesystem writes.
• Built for experimentation. Use structured fictional cases to test how agent claims relate to recorded observations.

How it works

Synthetic claims + events + snapshots + coverage
                      |
                      v
        review_synthetic_evidence_input
                      |
                      v
            Pinned evaluation engine
                      |
                      v
     Verdict + reason + evidence references

The DeepSeek Harness tool accepts an evidenceJson string (maximum 32 KiB) containing a declared synthetic:true case, a SYN-* case ID and synthetic targets. The engine returns bounded conclusions without rewriting the underlying evidence. See example A and example B.

Example outcomes: an absent post-state can leave a file-creation claim UNVERIFIED; a successful synthetic tool event together with bounded before/after observations can support SUPPORTED_LIMITED. These are example-specific findings, not universal guarantees.

Install the experimental community beta

Use an isolated, credential-free DSH test Profile; Node.js 24, DSH rc.5 and pnpm are recommended. Download the .tgz and .tgz.sha256 assets from the GitHub beta release. Verify the archive hash before installing.

dsh plugin --profile web add ./dsh-guard-labs-evidence-check-0.2.0-beta.1.tgz
dsh --profile web --dump-config

Restart the isolated Profile and use review_synthetic_evidence_input with synthetic JSON only. Official CLI installation and configuration on Windows/Linux passed for rc.5; live-model rc.5 Agent use and production assurance remain unverified. GitHub beta is the first distribution channel; npm registry and community listing are separate.

Discover through a DSH custom plugin index

In DSH, enable Custom plugin indexes (experimental), open Discover → Add index source, and enter:

https://raw.githubusercontent.com/veeduzyl-hue/mindforge-guard-labs-dsh-evidence-check/main/plugin-index.json

Review the pinned Git commit before confirming installation. This is a project-hosted index, not a third-party market endorsement.

Explore the project

From a source checkout with Node.js 22+:

npm test
node scripts/verify-engine.mjs
node scripts/integration-preflight.mjs

The project includes a DSH tool integration and an npm package layout. Compatibility testing covers DSH rc.2/rc.5 npm tool execution and official rc.5 Profile CLI installation/configuration on Windows and Linux. Live Agent rc.5 is not certified.

Where it fits

Evidence Check is useful for synthetic agent-workflow experiments, developer education, claim–evidence regression tests and conversations about verifiability. It is not an authorization engine, execution gate, production log auditor or independent assurance certification.

Privacy and trust boundaries

• Synthetic inputs only. Do not submit real Agent traces, customer data, proprietary source code, personal information or secrets.
• The plugin performs local deterministic evaluation, but DeepSeek Harness may send chat messages and tool arguments to your configured model provider.
• A synthetic:true declaration is supplied by the user; the plugin cannot independently certify that an input contains no sensitive content.
• Verdicts are bounded by available evidence and do not establish real-world causation, complete effect coverage or formal MindForge Guard admission.

Read provenance and assurance boundaries and privacy guidance.

Ecosystem and license

Evidence Check is an independent MindForge Guard Labs community experiment built for DeepSeek Harness. It is not a component of the formal MindForge Guard governance product.

MIT License · Copyright © 2026 veeduzyl-hue. See LICENSE and third-party dependencies.

For feedback, use a minimal, human-reviewed fictional reproduction; never attach raw sessions or secrets. Synthetic feedback template.