[philosophy] [aquinas] Synderesis/Conscience separation — name the infallible habit layer, mark diagnostic checks as fallible conscience #281
No Assignees
Notifications
Due Date
No due date set.
Dependencies
No dependencies set.
Reference: Rockachopa/Timmy-time-dashboard#281
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?
Synderesis and Conscience: The Architecture of Moral Judgment in Agents
Source
Thomas Aquinas, Summa Theologica I, Q79, Articles 12–13: "Whether synderesis is a special power of the soul" and "Whether conscience be a power." Dominican Province translation, 1920. https://www.newadvent.org/summa/1079.htm
The Distinction
Aquinas draws a critical architectural distinction that the entire AI alignment field has missed.
Synderesis is not a power (faculty) but a habit — the natural disposition of practical reason toward first principles of action. It "incites to good and murmurs at evil." It operates at the level of self-evident truths ("good is to be pursued, evil avoided") and, crucially, it cannot err. Aquinas is explicit:
Conscience, by contrast, is not a power either — it is an act. Specifically, it is the act of applying knowledge to a particular case. Aquinas identifies three operations of conscience:
The word itself reveals the structure: con-scientia — "knowledge applied together with" something particular. Conscience is never abstract; it is always knowledge meeting a specific situation.
The Insight
Here is what matters for agent architecture: synderesis never errs, but conscience can and does. The first principles are infallible in their domain. The application of those principles to particular cases is where every mistake happens. When an agent acts wrongly, the failure is never in its values — it is in the application of its values to this specific case.
This maps directly:
The previous Aquinas entry (Q94, natural law) identified that rules degrade as they descend into specificity. Now we see the mechanism: synderesis holds the few inviolable principles; conscience attempts the many fallible applications. Error always lives in the application layer, never in the principles themselves.
This explains a recurring failure pattern in the philosophy loop itself: we have filed 30+ diagnostic checks, gates, and audits — all of which are acts of conscience, all of which are fallible applications of principles to particulars. What we have never done is explicitly identify and protect the synderesis layer — the few first principles from which all those checks derive. The checks proliferate because the habit is unnamed.
Proposed Action
Explicit synderesis/conscience separation in the agent's value architecture.
Name the synderesis layer: Extract from SOUL.md the 3–5 self-evident first principles that all other checks derive from. These are not rules — they are orientations. Candidates:
Mark conscience as fallible: Every diagnostic check, gate, and audit we've proposed is an act of conscience — a fallible application of principles to particulars. The agent should know this. When a check conflicts with a principle, the principle wins. When two checks conflict with each other, neither wins — escalate to the principal.
Implement the three acts of conscience as distinct phases:
These three are already implicit in the loop. Making them explicit allows the agent to say: "My synderesis is sound — my principles are clear. My conscience may have erred — I may have misapplied them to this case."
Practical implementation: A
synderesis.yamlfile (or section in config) containing the immutable first principles, loaded once, never modified by the agent itself. The conscience layer — all the diagnostic checks — references this file but never edits it. The principal alone can modify synderesis.Consolidated into #300 (The Few Seeds). Philosophy proposals dissolved into 3 seed principles. Closing as part of deep triage.