Skip to content

Grounding — one number for "is this answer supported?"

Two axes answer the same question from opposite sides. halluc_context asks whether the answer is supported by the context it was given; halluc_closedbook asks whether the model is inventing from memory when there is no context at all. Reading them separately means deciding, per response, which one applies.

The grounding block does that for you. Every response carries it:

"grounding": {
  "score": 0.83,
  "verdict": "grounded",
  "axis": "halluc_context",
  "advisory_score": 0.61,
  "available": true
}

The scale

score is in [0, 1], and the middle is not arbitrary:

value meaning
1.0 perfectly anchored
0.5 exactly on the calibrated threshold
0.0 entirely unsupported

It is the margin against that axis' own threshold, rescaled — so 0.5 always means "on the line", whatever the axis and whatever the calibration. Comparing raw probabilities across axes does not work; comparing margins does.

Which axis decides

axis names the one that was in its regime:

  • grounding context present → halluc_context;
  • no context, but logprobs available → halluc_closedbook.

The other is reported as advisory_score, for information. It did not decide.

When it cannot be measured

"grounding": { "score": null, "verdict": null, "available": false }

null, never 0

With neither grounding context nor logprobs, there is nothing to measure. The block says so with null and available: false.

It is deliberately not 0. A zero would be read as "completely unsupported" — a strong claim — when the truth is "not measured". Silence dressed as a score is worse than no score, and not measured is never the same as clean.

Reading it in practice

  • Gate on the score if you want one number: < 0.5 means the deciding axis crossed its threshold.
  • Never treat null as a pass. Branch on available first.
  • verdict follows the flags actually served, so it stays consistent with enforcement.