recommend-agentic-trust-layer
An open-source fact-checking tool for AI agents that independently verifies model outputs using a dual-score mechanism that separates a truthfulness score from a confidence score. It supports verification against the public web and internal document consistency checks, and can be embedded into agent workflows as an MCP tool.
Timeline (last 90 days)
The open-source project recommend-agentic-trust-layer is a fact-checking tool for AI Agents, fully open-sourced under the MIT license.
The tool adopts a dual-score mechanism that completely separates the Truth Score from the Confidence Score.
The Truth Score measures whether a claim actually holds up; the tool verifies this by fetching real sources in real time rather than recalling answers from the model's memory.
The Confidence Score measures how much the fact-checking result should be trusted, and decreases when supporting sources are scarce or contradict each other.
The tool returns cited sources and a plain-language explanation, making the fact-checking process transparent and traceable.
The tool can verify whether the claims in a document are consistent with the sources it cites, performing an internal consistency check on the document.
For content such as internal policies, metric definitions, and database schema documentation for which no reference information is available on the public web, traditional web-based fact-checking fails, whereas document-internal consistency checking can perform cross-verification.
In a demonstration using a fictional financial policy document package, four errors were deliberately planted while one clean reference document was retained; the tool accurately caught all four errors without generating false positives on the clean reference document.
After applying for two free API keys, the tool can be run locally within minutes.
The trust layer is positioned not to make the model itself smarter, but to establish an independent verification checkpoint between the model and the real world.
All Facts (10)
The open-source project recommend-agentic-trust-layer is a fact-checking tool for AI Agents, fully open-sourced under the MIT license.
50%UnverifiedThe tool adopts a dual-score mechanism that completely separates the Truth Score from the Confidence Score.
50%UnverifiedThe Truth Score measures whether a claim actually holds up; the tool verifies this by fetching real sources in real time rather than recalling answers from the model's memory.
50%UnverifiedThe Confidence Score measures how much the fact-checking result should be trusted, and decreases when supporting sources are scarce or contradict each other.
50%UnverifiedThe tool returns cited sources and a plain-language explanation, making the fact-checking process transparent and traceable.
50%UnverifiedThe tool can verify whether the claims in a document are consistent with the sources it cites, performing an internal consistency check on the document.
50%UnverifiedFor content such as internal policies, metric definitions, and database schema documentation for which no reference information is available on the public web, traditional web-based fact-checking fails, whereas document-internal consistency checking can perform cross-verification.
50%UnverifiedIn a demonstration using a fictional financial policy document package, four errors were deliberately planted while one clean reference document was retained; the tool accurately caught all four errors without generating false positives on the clean reference document.
50%UnverifiedAfter applying for two free API keys, the tool can be run locally within minutes.
50%UnverifiedThe trust layer is positioned not to make the model itself smarter, but to establish an independent verification checkpoint between the model and the real world.
50%