A product answering questions over governed data, where a confident wrong answer is worse than no answer and a raw model over the corpus will fabricate a plausible figure. On governed or regulated data, provenance and the ability to abstain are requirements rather than features. Outputs are acted on, so the system has to be crash-resumable with a human checkpoint before anything happens.
The felt need was a better or bigger model. The real problem is that trust cannot come from the model at all. It comes from the system around it: what is retrieved, what is checked against evidence before it is shown, and what happens when evidence is absent. The model is a component. The surrounding system is the product.
The reflex, and the fix.
Chase a bigger model
It is the lever everyone reaches for, and every release makes it look more reasonable.
A larger model fabricates more fluently. Nothing about scale converts an ungrounded claim into a checkable one, and on governed data the fluency is the hazard.
Put the model behind a gate and an abstain path
A faithfulness judge, an explicit abstain path, prompt versioning, a regression harness, and a graph runtime with a human checkpoint.
Trust as a property of the system rather than a hope about the model.
Make trust structural, because a model cannot be argued into being trustworthy.
Retrieval over the governed corpus
Vector search scoped to the data the answer is allowed to come from, which is what makes provenance possible at all.
The model as one component, deliberately not the authority
It generates. It does not decide whether what it generated is true, and it is not the last thing in the chain.
A faithfulness judge that gates the response
Every generated claim is scored against its cited evidence before anything is shown. Low-scoring output is blocked rather than shipped, which is precisely the failure a plain pipeline lets through.
An explicit abstain path
When grounding is absent the system returns insufficient data. Abstention is a correct output, not a failure to paper over.
Prompt versioning and a regression harness
Prior traces re-run on every model or prompt change, so drift is caught before it ships rather than after a customer finds it.
A crash-resumable graph runtime
The flow survives interruption and holds a human-in-the-loop checkpoint before any action is taken.
A single headline number hides where a system fails. This work was scored on the dimensions that actually decide whether it holds in production, measured on real, held-out cases rather than the demo path.
38% token-cost reduction and 45% accuracy lift are measured. The zero-fabrication claim stays qualitative pending a provenance artifact.
A fluent system that answers everything, cites nothing checkable, and fabricates a plausible figure on exactly the questions where a wrong answer is most expensive.
A retrieval system where every claim is checked against evidence, insufficient data is a valid answer, every prompt change is gated by a regression evaluation, and the flow is crash-resumable with a human checkpoint before any action. Once the harness existed, cost work became safe to do, because a change that saved tokens and lost accuracy showed up immediately: a 38% token-cost reduction alongside a 45% accuracy lift.
What it owns, and what it hands to a person.
The cost and accuracy figures are measured. The zero-fabrication claim is stated qualitatively, deliberately, until a provenance artifact is attached to back it. A judge measures what it has evidence for, the corpus bounds what can be answered at all, and the regression set has to be maintained as the data and the questions change.
The reliability problem at the top of the stack is not model quality. It is that a system with no gate, no abstain path, and no regression suite has no way to know it has stopped being right. That is ordinary production engineering discipline applied to a layer where it is frequently absent.