Neuro-symbolic root-cause analysis
A language model can generate plausible explanations without knowing they are right. A symbolic checker can test explicit rules against available facts, but it is only as complete as its model and inputs. This reference pattern combines them to make evidence and uncertainty visible.
The problem with asking a model "why"
When an operation breaks, the first question is always the same: why. An order is going to ship late. A wing of the hotel is running cold. A warehouse station has stalled. Ask a capable language model and it will give you a fluent, plausible explanation almost instantly. That is the problem. It is fluent and plausible whether or not it is correct, and in operations a confidently wrong root cause is worse than no answer, because it sends a real team to fix the wrong thing.
The failure mode here is well documented. Large language models are excellent at producing hypotheses that sound like reasoning but are not grounded in the actual state of the system. They hallucinate causes the way they hallucinate citations: convincingly. For a system that is about to take an action in the physical world, "convincingly" is not good enough.
What symbolic systems get right, and wrong
The classical alternative is symbolic: a model of the system, a set of rules, and a deterministic engine that infers a supported explanation from observed state. It can show its work, but its conclusion is conditional on the completeness and correctness of the facts, model, and rules. That makes the chain auditable, not infallible.
What symbolic systems are bad at is the open world. They are brittle at the edges, they require someone to encode the rules in advance, and they fall silent the moment reality presents a situation the rules did not anticipate. The very rigidity that makes them trustworthy makes them narrow.
The division of labor
Fibric's approach follows a line of work, sometimes called neuro-symbolic AI, that pairs a neural model's flexibility with a symbolic engine's rigor. The two play distinct roles, and keeping the roles distinct is the whole point.
The neural side does what it is good at: generate hypotheses. Given the unified operational picture, the model proposes a ranked set of candidate root causes. It is allowed to be creative, even speculative, because nothing it proposes is trusted yet.
The symbolic side does what it is good at: check against the available facts. Each candidate cause is tested against governed data and the encoded structure of the system. A hypothesis can be marked supported, contradicted, or unverifiable; that is not the same as proving causality.
What comes out the other side is a chain, not a sentence: the cause Fibric believes, the evidence that supports it, and the candidates it ruled out and why. That chain is what makes the conclusion auditable. An operator does not have to take the answer on faith. They can read the reasoning the way they would read a colleague's note, and disagree with a step if they know something the data does not.
Why this matters before an action
On Fibric, root-cause analysis sits directly upstream of action. The platform's job is to do something about a problem, not only to explain it. That makes the grounding non-negotiable. An action is only as trustworthy as the reasoning that proposed it, and a hallucinated cause that drives a real action is precisely the failure the whole trust spine exists to prevent.
In the reference pattern, a supported explanation can feed a governed action proposal, while deterministic policy may allow, hold, or veto it. On supported paths, the explanation and action record can be linked so reviewers can inspect which evidence informed the proposal. That record documents the basis for a decision; it does not prove the proposed cause was the only cause.
Where we are honest about the limits
This is a direction, not a solved problem. Grounding a hypothesis requires that the relevant facts are actually in the operational picture; a cause that depends on a signal no connector provides cannot be confirmed or denied, only flagged as unverifiable. We would rather surface "we cannot confirm this" than launder a guess into a certainty. The honest "unverifiable" is itself a useful output, because it tells you where your sensing has a gap.
Grounding
- Neuro-symbolic AI: the research program of combining neural learning with symbolic reasoning to get both flexibility and verifiable inference.
- Model-based diagnosis: the classical, symbolic tradition of inferring root cause from a model of system structure and observed state.
- Hypothesis-and-test: treating an LLM as a generator of candidate explanations that an external, deterministic checker must confirm before they are trusted.