Rule-Gated Generation: A Harness That Filters, Repairs, Certifies and Re-Checks Machine-Generated Output and Improves Its Own Rules Under a Gate
Year: 2026
Type: article
Abstract: A harness wraps one or more generators (language models, other neural models, programs or people) with a symbolic layer: a rule set and, optionally, a knowledge store. A checker evaluates each candidate output against the rules and returns a verdict and the violated rules. A release gate lets the output, or part of it, leave only when it passes or under an exception policy, fixed in advance or chosen per request. That is the minimal core. Optional modules, used alone or in any combination, extend it. Rules may filter inputs, decoding, finished outputs, agent actions before execution, and outputs already delivered. A failing output may be repaired by a minimal symbolic change, or revised under a budget the generator cannot alter, with the passing part released under caveats. A passing output may carry a receipt tied to the state it was checked against, so that a later change re-checks affected outputs and supersedes a receipt only when its verdict changes (or, optionally, when the facts it records change). The rule set may improve from evidence, each change held inactive until it passes a gate that can refuse unapproved weakening and require independent approval. Further modules separate claims from outcomes, count correlated sources once, re-extract rendered prose, replay decisions, accept statements by cost, add human review, and turn the harness's verdicts, repairs and revisions into training signals. Outputs may also be read into typed records before checking, with size caps and a per-rule estimate of evaluation work, or a step counter, so that each check made by a rule whose work can be estimated or counted is bounded in advance, and a request may bring rules of its own, written by the generator or another source and admitted only when they cannot relax the rules in force. Knowledge and policy changes can then be made in rules instead of model weights, and any fine-tuning that remains needs less manual labelling.
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7222904794
FWCI:
Cites:
Cited by:
Related to: