A developer has implemented a new deterministic gate (a logic-based filter) for AI agents to improve how they process memory updates and citation accuracy. The system, designed to prevent AI from misinterpreting administrative rules, adds a relation-span clause requiring that any citation of a rule change must contain both a change keyword and the rule's specific scope within the same sentence. Before testing, the author published a public repository with pre-registered predictions, pass-fail criteria, and anticipated failure shapes, including a known class of 'proximity traps' that the new mechanism was expected to miss.

Measurements from a 23-case test run demonstrated a reduction in false alarms. Using the llama3.2 model, false alarms dropped from 5 to 1. On the stronger Claude Sonnet model, false alarms remained at 1, while maintaining a 9/9 success rate on textual direction catches. The results indicated that the gate successfully blocked citation-shaped falsehoods without losing accurate textual catches. However, the system continues to struggle with 'proximity traps,' where a sentence contains the correct change words and scope terms but does not actually establish a valid logical replacement. In these instances, the AI model erroneously identifies a rule change based on the sentence's grammatical structure rather than its semantic meaning.

The author attributes the design of the v2 system, specifically the strong-bind and proximity-bind distinction, to collaborative feedback from Mike Czerwinski. While the v2 update effectively restricts the types of errors the AI can make, it does not resolve the underlying issue of argument resolution—the need for the system to verify whether the change word's arguments specifically link the two rules on trial. The author has made the full testing suite and the chain of experimental commits publicly available for independent verification, labeling the current limitation as the 'proximity' problem, which will serve as the focus for future development versions.

A developer has implemented a new deterministic gate (a logic-based filter) for AI agents to improve how they process memory updates and citation accuracy. The system, designed to prevent AI from misinterpreting administrative rules, adds a relation-span clause requiring that any citation of a rule change must contain both a change keyword and the rule's specific scope within the same sentence. Before testing, the author published a public repository with pre-registered predictions, pass-fail criteria, and anticipated failure shapes, including a known class of 'proximity traps' that the new mechanism was expected to miss.

Measurements from a 23-case test run demonstrated a reduction in false alarms. Using the llama3.2 model, false alarms dropped from 5 to 1. On the stronger Claude Sonnet model, false alarms remained at 1, while maintaining a 9/9 success rate on textual direction catches. The results indicated that the gate successfully blocked citation-shaped falsehoods without losing accurate textual catches. However, the system continues to struggle with 'proximity traps,' where a sentence contains the correct change words and scope terms but does not actually establish a valid logical replacement. In these instances, the AI model erroneously identifies a rule change based on the sentence's grammatical structure rather than its semantic meaning.

The author attributes the design of the v2 system, specifically the strong-bind and proximity-bind distinction, to collaborative feedback from Mike Czerwinski. While the v2 update effectively restricts the types of errors the AI can make, it does not resolve the underlying issue of argument resolution—the need for the system to verify whether the change word's arguments specifically link the two rules on trial. The author has made the full testing suite and the chain of experimental commits publicly available for independent verification, labeling the current limitation as the 'proximity' problem, which will serve as the focus for future development versions.