Granite-4.0-350M-VERIFIER — commercial-safe on-device FAITH judge

A 350M verifier fine-tuned to judge one meeting-notes bullet against transcript evidence with a single-word verdict: SUPPORTED / UNSUPPORTED / CONTRADICTED (the FAITH protocol of the agentic-summarizer pipeline). It replaces the 20B gpt-oss judge in the pipeline's in-stream verification and final VERIFY sweep, making the whole summarization pipeline on-device.

• Base: ibm-granite/granite-4.0-350m (Apache-2.0 — commercial use OK, unlike the earlier LFM2.5-350M-based verifier whose base license restricts commercial use).
• Data: 2,644 judged (bullet, evidence, verdict) triples harvested from the pipeline's own judged T1 runs (gpt-oss-20b verdicts, 3x majority), class-balanced to equal thirds — the balance removes the SUPPORTED/UNSUPPORTED collapse bias measured on unadapted verifiers (5% agreement) and on a 270M base (70% ceiling).
• Training: full fine-tune, LR 2e-5, 4 epochs, DDP on 2 GPUs.
• Measured: 97% agreement with gpt-oss-20b on 200 held-out triples (the LFM2.5-350M verifier: 96%; a Gemma-3-270M fine-tune: 70% — 270M is too small for this task).

Usage

llama.cpp server (greedy):

llama-server -m granite-4.0-350m-verifier.Q4_K_M.gguf --n-gpu-layers 999 --ctx-size 4096 \
  --parallel 1 --flash-attn on --jinja --temp 0

Client: the harness's in-stream verification or final sweep (eval/run_arms.py --verify-url <server> / --sweep-judge local:<port>/... in the agentic-summarizer repo). The system prompt and the EVIDENCE/BULLET prompt format must match the training distribution byte-for-byte (eval/judge.py _FAITH_SYS / faith_prompt).

Caveats

• Agreement is measured on the pipeline's own evidence-retrieval distribution.
• It is a verdict classifier, not a general judge: use only inside the pipeline.