Self-hosted groundedness verifier

prism’s citation path checks each claim against its retrieved source. The final, LLM-level step of that check — the groundedness lens — can be backed by a model you host instead of a hosted API. It is opt-in: set one environment variable and prism routes the citation check to your endpoint, falling back to your hosted verifiers if it is unreachable.

Why you’d want it:

• Zero per-call cost on the highest-frequency check. Citation groundedness is the check that runs most often; moving it to a local model takes it off your hosted-API bill.
• Evidence stays on your machine. The claim and the retrieved source never leave your host.
• A model tuned for the job. You can serve a model fine-tuned specifically for grounded entailment rather than a general-purpose one.

Where it sits

The citation path is a layered floor — each stage refuses what it can prove, else defers:

1. Existence floor — live arXiv/Crossref retrieval drops a fabricated identifier.
2. Numeric/unit floor — a percentage swap or comparison-direction falsehood is caught arithmetically.
3. Groundedness lens — the family-different LLM check against the retrieved abstract. This is the step the local verifier serves.
4. NLI floor (opt-in, PRISM_NLI_FLOOR) — an orthogonal encoder-NLI veto on a supported.

The local verifier registers as a distinct model family, local-verifier, separate from the local (mistral) family — so Lock 1 (family-different) still holds, and mistral remains a failover target rather than being displaced.

Enable it

Point prism at an OpenAI-compatible chat endpoint:

export PRISM_LOCAL_VERIFIER_ENDPOINT=http://127.0.0.1:8092   # base URL of your served model
export PRISM_LOCAL_VERIFIER_MODEL=qwen3-14b-groundedness     # optional label (this is the default)

When PRISM_LOCAL_VERIFIER_ENDPOINT is set, prism injects the specialist as the primary citation verifier for every caller, failing over to that caller’s existing cross-family verifiers (your hosted families, then mistral local) when its circuit opens. When the variable is unset, routing and behavior are exactly as they were — this changes nothing until you opt in.

Verdict mapping

The model’s three-way verdict maps onto prism’s citation outcomes:

model verdict	prism outcome	action
supported	supported	accept
unsupported	contradicted	revise — fix the claim to match the source
abstain	not_addressed	escalate — retrieve the full text

Fail-open, by design

If the served model is unreachable, returns a bad body, or produces no readable verdict, the provider raises an error rather than returning a default. prism’s circuit-breaker then fails over to the caller’s hosted (or mistral) verifiers. A verifier outage is a loud failover — never a silent “escalate everything.”

Because the breaker fails over on errors, not on a confident-but-wrong supported, pair the local verifier with the orthogonal NLI floor so a mechanistically-different model can veto a false confirm:

export PRISM_NLI_FLOOR=1     # recommended alongside the local verifier

Serve your own

The provider talks plain OpenAI-compatible chat: it POSTs to <endpoint>/v1/chat/completions and health-checks <endpoint>/health. You do not need your model to emit JSON or follow prism’s prompt format. The provider injects the groundedness system prompt, re-templates the source/claim into the EVIDENCE: … / CLAIM: … shape the model expects, strips a leading reasoning (<think>) block, reads the bare supported | unsupported | abstain verdict, and emits the JSON prism parses. So all you have to serve is a chat model that reasons over evidence and a claim and answers with one of those three words.

Any OpenAI-compatible server works (llama.cpp’s server, vLLM, and others). The reference setup serves a fine-tuned Qwen3-14B groundedness adapter under llama.cpp:

# llama.cpp server (CUDA), base model + the groundedness LoRA adapter
llama-server -m Qwen3-14B-Q4_K_M.gguf \
  --lora-init-without-apply --lora groundedness-adapter.gguf \
  --host 0.0.0.0 --port 8092

# the adapter is an rsLoRA adapter — apply it at scale 4
curl -s http://127.0.0.1:8092/lora-adapters \
  -H 'Content-Type: application/json' \
  -d '[{"id":0,"scale":4.0}]'

Then export PRISM_LOCAL_VERIFIER_ENDPOINT=http://127.0.0.1:8092 and run prism verify --type citations … as usual.

To build the training data for your own groundedness model from prism’s own verifications, see the groundedness harvest sink.