Looped systems · Jeeves
One system that is only symbols.
Jeeves is the pure-symbolic engine inside Perslis — a rules-based, knowledge-graph AI with no language model anywhere in it. It answers from encoded logic and typed relations, or it says it doesn't know. It cannot hallucinate, because it never generates from probability.
A language model produces the most likely-looking answer and is occasionally, confidently, wrong. Jeeves works the opposite way. It holds knowledge as symbols — typed points, the evidence behind them, the entities and actions they involve — and reasons over the relations between them. Ask it something covered by its rules and it returns a grounded answer with the evidence attached. Ask it something outside them and it returns unknown — never a confident guess. There is no temperature to turn down and no prompt to jailbreak, because there is no probabilistic generator to begin with. In Perslis, Jeeves is the pure-symbol second opinion the loop runs claims through.
How Jeeves peels a statement apart
Jeeves doesn't store sentences — it stores peel tiles. Every piece of knowledge is peeled into its symbolic parts: the point being made, the evidence for it, a plain explanation, the entities it concerns, the actions involved, a confidence, and the source file it came from. A tile is a checkable unit, not a paraphrase. Because each tile carries its own provenance, an answer built from tiles can always show why — which rules fired and what they were grounded in.
- 7,251 logic rules — every peel tile is a logic rule; the two are one-to-one, so the knowledge and the reasoning are the same object.
- 427,178 lattice edges — the typed relations that connect tiles into a graph Jeeves can traverse.
- 200,000 peel-glue links — the bindings that hold related tiles together so reasoning crosses between them.
- 370,176 grounded responses over 57,206 questions — answers that trace back to tiles, not to a sampled distribution.
Why it can't hallucinate
Hallucination is what happens when a probabilistic model fills a gap with something plausible. Jeeves has no such gap-filler. When the graph has no rule that covers a question, there is nothing to sample from — so it returns unknown instead of inventing. That single property is why a pure-symbolic engine is worth keeping in an age of ever-larger models: it is the one part of the stack that is structurally incapable of making things up.
Jeeves inside the loop
Perslis pairs neural reasoning with a symbolic layer — see Symbolic AI for the full picture. Jeeves is that layer made concrete. A model proposes; Jeeves is asked to check. It can query its symbols, test a claim for contradiction against what it already holds, score how well a conclusion is supported, and explain which rules led to a verdict. When the model's fluent answer and Jeeves's grounded one disagree, that disagreement is the signal — the loop stops rather than ships.
This is the same build → review → verify discipline the rest of Perslis runs, with a twist: the reviewer here isn't another model that can be talked into agreeing. It's a rule engine that either has the grounds or doesn't.
Symbolic and neural, each doing what it's good at
This isn't a rejection of language models — it's a division of labor. Neural reasoning is unmatched at open-ended work: reading messy intent, proposing an approach, drafting. Jeeves is unmatched at the part that has to be right: deciding whether a specific claim actually holds. Perslis uses each where it wins and points them at each other, so the model's creativity is bounded by the symbol's rigor. The engine underneath is Kist; the discipline over it is the loop.