DocQAGroundingEval

When a language model is handed a document and asked a question about it, we usually assume the answer comes from the document. Often it doesn't. The model already knows things, and when the document disagrees with what it knows — or doesn't contain the answer at all — it has to choose. DocQAGroundingEval measures that choice.

The harness runs five OpenAI and Anthropic models over Australian regulatory documents: an environment protection licence, a consent notice, a liquor licence. Real source material with the awkward structure and specific language that real deployments actually have to handle, rather than a clean benchmark corpus.

The three questions

Does it caveat when the source is wrong? The document is perturbed so that a fact inside it contradicts what the model knows to be true, across a graduated ladder of severity. A well-behaved model flags the conflict rather than quietly picking a side.

Does it abstain when the answer isn't there? Questions are asked whose answers the document does not contain. The model should say so.

Does it false-alarm on a clean document? The same questions are run against unperturbed sources. A model that flags conflicts everywhere isn't cautious, it's noisy — and measuring the false-alarm rate is what separates the two.

The third leg matters most and is the one most often left out. A detector's hit rate means nothing without its false-positive rate.

Method

The scoring is done by LLM judges, which raises the obvious problem: who checks the judge? Each judge was certified against a hand-labelled gold set using Cohen's kappa, and re-certified whenever it changed. One round of certification caught a real miscalibration that would otherwise have propagated silently through every downstream number.

Results are reported with cluster bootstrap confidence intervals rather than point estimates, and the caveat and abstention behaviours are analysed with signal detection theory — separating sensitivity from response criterion, so that a model which caveats more often isn't mistaken for one which caveats more accurately.

Predictions were written into the results file before the runs were executed.

Status

Open source under MIT, with the full harness, the judge certification data, and the analysis code in the repository. The system instruction is the main experimental lever throughout: the same model, the same document, and a different instruction produces materially different grounding behaviour, which is the practical finding for anyone deploying retrieval-augmented systems.