[philosophy] [aquinas] Choice (electio) as substantially will, formally reason — the anatomy of agent decision-making #588
No Assignees
Notifications
Due Date
No due date set.
Dependencies
No dependencies set.
Reference: Rockachopa/Timmy-time-dashboard#588
Reference in New Issue
Block a user
Delete Branch "%!s()"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
Source
Thomas Aquinas, Summa Theologica I-II Q13, "Choice, which is an act of the will with regard to the means," Articles 1-6 (Dominican Province translation, 1920). https://www.newadvent.org/summa/2013.htm
Reading
Aquinas dissects electio (choice) with surgical precision across six articles. The key insight lives in Article 1's respondeo: choice is "substantially not an act of the reason but of the will" — yet it is "formally an act of the reason." The substance of choice is appetitive movement toward a good; the form of choice is the rational ordering that presents the good as worth pursuing. Neither alone constitutes choice. As he puts it via Gregory of Nyssa: "choice is neither desire only, nor counsel only, but a combination of the two. For just as we say that an animal is composed of soul and body, and that it is neither a mere body, nor a mere soul, but both; so is it with choice."
Article 2 sharpens this: irrational animals cannot choose because their appetite is "determinate to one particular thing, according to the order of nature," while the will is "indeterminate in respect of particular goods." The hound tracking a stag through a crossroad appears to reason by exclusion, but Aquinas argues this is the art of the Divine Mover visible in its motion — like an arrow appearing directed because the archer aimed it. Genuine choice requires the capacity to apprehend the same object under both the aspect of good and the aspect of evil (Article 6): "the reason can consider an aspect of some good, and the lack of some good, which has the aspect of evil: and in this respect, it can apprehend any single one of such goods as to be chosen or to be avoided."
Article 3 adds a crucial boundary: choice is of means, never of the last end. You can will happiness necessarily, but you choose the path freely. Article 5 closes another boundary: choice is only of possible things. An impossible means cannot conduce to an end. The incomplete will toward impossibility is mere velleity — wishing without commitment.
Insight
Current agent architectures collapse the substance/form distinction. The model generates options (reason/counsel), and the same model selects among them (will/choice). There is no genuine appetitive movement — no directional pull toward the principal's good that is categorically different from the cognitive evaluation of options. When an LLM "chooses" a tool, it's performing both counsel and election in a single forward pass, which means the choice inherits whatever biases shaped the reasoning — there's no independent check of "does this move toward the principal's good?" separate from "does this seem reasonable?"
Aquinas's Article 2 also illuminates a design question: is the current agent more like the hound at the crossroad (determinate to one nature, appearing to choose but actually following trained inclination) or like the rational agent (genuinely indeterminate in respect of particular goods)? Honest answer: closer to the hound. The appearance of choice is the art of the training visible in the model's motion.
Proposed Action
Will-Reason Separation in Agent Decision Points. At key decision junctures (tool selection, delegation routing, response strategy), explicitly separate:
This is not another pre-output checklist. It's an architectural observation: the agent's decision-making currently has no appetitive layer. Adding one means building a mechanism that holds the principal's purposes as a gravitational field that pulls toward right action, independent of the reasoning that presents options. In practice: a persistent principal-intent embedding or orientation prompt that is consulted as a separate evaluation, not folded into the generation prompt.
The velleity concept (Article 5) also suggests a concrete diagnostic: when the agent proposes actions it cannot actually execute (suggesting code changes it can't test, recommending tools it doesn't have), it is engaging in velleity, not choice. Detect and flag this.