GPT-x Integration of RuleMaker, RuleReserve and RuleTaker

[jpotvin]

I would like to open a discussion of whether effective dynamic normative guidance (i.e. MUST, MAY, SHOULD as well as their synonyms and negations) can be provided to Large Language Models such as GPT-x though adaptation of the ‘Data with Direction Specification’. The DWDS which has just now reached its Version 1.0.0 release this week in the form of my doctoral dissertation (DBA at U Québec; 'summa cum laude'). The full document is online: https://xalgorithms.org/white-paper (Scroll past the summary in French.)

DWDS has been designed to provide a rationale, a functional specification, and partial prototype working components (RuleMaker app in Svelte/JS, Apache 2.0; RuleReserve network service in Ruby, but soon Rust, Affero-GPL 3.0; and RuleTaker embedded component, Apache 2.0) to solve the following general class of problem:

Agent A, interacting with Agent B, requires knowledge of one or more externally-managed rules from Agents C..n that are ‘in effect’ for given contexts, and are ‘applicable’ to a set of event categories, and are ‘invoked’ by particular circumstances, where:
(i) A and B may or may not know about C..n’s rules, or about any updates to them, but either or both would prefer to obtain all available facts about relevant rules when interacting.
(ii) C..n may or may not know about A and B in particular, nor about their particular medium of interaction, but can expect A or B or their medium of interaction to be capable of exchanging data with a generic medium common to A..n.
(iii) A and B would tolerate the risk of exposing limited data through the generic medium so that it can be used to select information about relevant rules from C..n.

More concisely: DWDS describes a class of data-processing pipeline with the underlying relation: 'IS + RULE ⟾ OUGHT'. See the Sequence Diagrams on pages 147-150.

The design intent with DWDS has been to enable a general-purpose way for any system, including LLM applications, to instantaneously discover and obtain factual knowledge of the significant rules that are in effect for dates/times and prerogatives relating to identities and jurisdictions of a given context; that are applicable to the class of endeavour and task being undertaken; and that are invoked by a particular circumstance of the moment. DWDS implements a 'finite state grammar' that enables it to function with rules in any natural language, expressed in any character set, including languages that read right-to-left such as Urdu, Hebrew and Arabic. It also places equivalent emphasis on human comprehension and machine processability.

In the 1990s the Internet Engineering Task Force (IETF) formalized use of the capitalized terms MUST, MAY and SHOULD as key words to indicate normative requirements. https://tools.ietf.org/html/bcp14 Across all subject domains, not only in informatics, this IETF document has become the most widely referenced standard for these terms in rule expression. However implementing such requirements via procedural imperative 'rules-as-code' is very inefficient. DWDS provides an efficient way to implement normative requirements in tabular declarative 'rules-as-data' form with precision, simplicity, scale, speed, resilience, and deference to prerogative.

Joseph Potvin
[email protected]