Luke F. Walton releases open-source Answer Engine for cited AI search

Luke F. Walton has released Answer Engine, an open-source AI search reference implementation that cites sources or refuses to answer when evidence is missing. The project, available now on GitHub and running in a live deployment called Ask the Archive, is designed to make AI answers more accountable for archives, research, and other collections.

Why it matters: - Answer Engine targets a core flaw in AI search: systems that sound confident but cannot reliably show where an answer came from. - The project is designed to make AI responses easier to trust, audit, and challenge when they are wrong. - The tool also gives institutions a way to use private material to improve retrieval without reproducing that material in answers.

What happened: - Luke F. Walton, an independent researcher on AI ethics and answerability and founder of Surmado, Inc., released Answer Engine on June 29, 2026. - Answer Engine is an open-source reference implementation for AI-powered search. - The project is available on GitHub under the Apache-2.0 license. - A live deployment, Ask the Archive, is running on Walton’s published writing and podcast transcriptions. - The project is a personal open-source project and is not a Surmado product.

The details: - Answer Engine is built to cite its sources or say “I don't know” when evidence is missing. - The system answers one question at a time and does not include a chatbot persona or conversation memory. - Each response is designed to be a single transaction: a cited answer from the corpus, a pointer to relevant private material, or an explicit refusal. - The corpus is split into public records, which can be quoted and cited, and private notes, which can be searched but are never shown to the model. - The internal prompt-building path uses a data type with no field for private text, making leaks structurally unavailable on the supported path. - Every answer declares one of four modes: supported, partial, related-material, or not-found. - The mode is derived from retrieved evidence rather than the model’s own judgment. - Invented sources are treated as errors, not footnotes. - A fixed suite of “gold” questions tests refusal behavior, including cases the system must decline to answer. - The project runs locally on Node.js with a single API key. - The repository includes a synthetic example corpus, the evaluation suite, and citation metadata. - Developers can swap in their own records, change retrieval scoring, and rewrite the gold questions for their own collections. - Answer Engine v2.0.0 is archived on Zenodo under concept DOI 10.5281/zenodo.20676773. - Full project documentation is available at Luke F. Walton’s site. - The accompanying research is available at Walton’s PhilPeople profile.

Between the lines: - Walton is arguing that the real fix for AI answer systems is not just more disclosure, but a visible owner standing behind the answer frame itself. - The project pushes back on answer-engine optimization, where outside actors can shape composed answers for advantage while users see only a neutral-looking response. - The design favors soundness over completeness: it will not state an answer unless the retrieved evidence supports it, but it does not prove that every relevant source was found. - Walton describes the tool as a bounded teaching and reference case, not a proof that the same safeguards will hold at public scale. - The private-material layer is likely to matter for libraries, museums, labs, and other institutions that want restricted material to improve discovery without exposing the underlying text.

What's next: - Walton expects developers to clone the repository, replace the sample corpus with their own records, and adapt the retrieval and evaluation rules to their own archives. - The gold-question test suite is meant to evolve as a collection grows, making refusal behavior part of ongoing quality control. - Walton says the work is meant as an engineering companion to his broader research on answerability, machine-mediated responsibility, answer-engine optimization, and citation-grounded AI systems. - Contributions to the repository are welcome.

The bottom line: - Answer Engine is a small but pointed attempt to make AI search answerable: cite the evidence, hide what should stay private, and refuse when the system does not know.