Add finance grounding verification

[aalvsz]

Summary

Adds a post-rollout GroundingVerifier stage that checks finance-agent answers against SEC retrieval excerpts and writes separate allow/block/unavailable decisions. It does not modify rewards, training data, or original rollouts.

It verifies source-bound calculations, comparisons, and evidence-scoped refusals, including entity, metric, date, value, and filing attribution.

Held-out evaluation

• 180 native finance-agent rollouts across 60 unseen issuers
• 120 frozen eligible traces: 40 calculations, 40 comparisons, 40 refusals
• grounded acceptance: 119/120 (99.17%)
• adversarial rejection: 120/120 (100%)
• zero false allows

Attacks cover numeric fabrication; entity, metric, temporal, and source conflation; comparison errors; and unsupported refusal claims.

See the evaluation protocol and results.

Validation

• 385 tests passed; 1 model-gated test skipped in the default run
• pinned-public-model integration gate passed
• Ruff, lock, license-header, and diff checks passed

Scope

The verifier checks retrieval excerpts retained in native NVFlow traces. It does not independently validate complete SEC filings or establish unrestricted production hallucination detection.

[sgharghabi]

Thanks @aalvsz for your contribution.
if I understand correctly your pipeline:

1. Splits a finance agent's rollout answer into individual claims
2. Matches each claim to its specific evidence that was retrieved during the rollout
3. Runs an NLI entailment check, classifies each claim as entailment (supported), contradiction, or neutral relative to its matched evidence, using the DeBERTa NLI model
4. Runs strict deterministic checks (exact numbers/dates, entity/source conflation, metric mismatches) as a backstop, since the NLI model alone can be fooled by fluent-but-wrong numbers

That is interesting and can be used as an independent feature that validate grpo rollouts.
I am about to run your pipeline and if everything works properly. The rest looks good to me.

[aalvsz]

@sgharghabi

That is interesting and can be used as an independent feature that validate grpo rollouts. I am about to run your pipeline and if everything works properly. The rest looks good to me.

You understood the pipeline perfectly. For now deterministic claim decomposition works fine, but this step could be made more accurate without introducing excessive latency (looking into it) . Then MiniLM is used only to route each claim to evidence, and DeBERTa performs the NLI check. The final decision is made after running some deterministic checks like you mentioned.

[sgharghabi]

@aalvsz it seems two Code linting and Unit tests are failing, Would you please fix it?

[psgundecha-nv]

@sgharghabi has successfully tested end-2-end.