FORGE AGENTS

Paper

Refuse or Cite

A source-grounded assistant for front-desk procedures, and the test apparatus that keeps it honest. Dustin Aldridge, Forge Agents LLC. Version 3.0, October 5, 2026.

Read the paper (PDF, 13 pages, 179 KB). Archived on Zenodo: DOI 10.5281/zenodo.23156948.

Abstract

A hosted assistant answers front-desk questions only from an office's own procedure documents: each answer quotes a verified passage, states that the answer is not documented and routes the question to a person, or asks one clarifying question. Most earlier failures were correct answers rejected on wording. A paraphrase sweep widened 83 tests whose required wording was too narrow. Written wrong answers exposed 14 tests that accepted an incorrect answer, then three safety and four veterinary tests; all were tightened. The grader reads negations through Markdown emphasis and ends them at contrast words, and both serving paths share one grader. Two veterinary answers that version 2.0 counted as a model error, later called a grader miss, are confirmed as errors: they added timing advice absent from the clinic's after-hours procedure. The test now rejects that advice; answers stopped giving it only once the procedure explicitly forbade it. In the last live run of each of the nine public suites, at five trials per test, the serving model passed 310 of 310 tests, or 309 without the negation referral, which version 2.0's headline excluded. Two outside judge models, rating 4,865 answers, agreed with the grader on 96.9 and 83.9 percent.

What is in it

The tests, checked in both directions.

The paper sets out how the test definitions behind the public scores were checked from both sides: rewordings of correct answers found tests too narrow to pass a correct answer worded differently, and written wrong answers found tests too loose to fail an incorrect one. It covers two corrections to how the grader reads negations, one grader shared by the console and the connector, and a rule for loosening a safety test. It then reports the last live run of every public suite at five trials, re-checked by two outside judge models, and corrects one earlier verdict on a veterinary after-hours answer. The backup model, capacity, retrieval, entailment, injection and failover results are carried from version 2.0.

Records

The scores are public; the records are on request.

The public scores, with every failing test named and a note on each, are on the proof page, which is rebuilt hourly. The stored answers, test definitions, written wrong answers, rewordings, judge ratings and grader versions behind the paper are available on request. The founder's account of why the product is built this way is on the founder page.

To cite: Aldridge, D. (2026). Refuse or Cite: A Source-Grounded Assistant for Front-Desk Procedures and the Test Apparatus That Keeps It Honest. Version 3.0. Forge Agents LLC, Saint Paul, Minnesota. https://doi.org/10.5281/zenodo.23156948

Version 2.0 of September 26, 2026 remains available as a PDF and on Zenodo as 10.5281/zenodo.22985559. Version 1.0 of September 20, 2026 remains available as a PDF and on Zenodo as 10.5281/zenodo.22866053. The DOI for all versions is 10.5281/zenodo.22866052; it always resolves to the newest version.