Brendon.BOT

Trustworthy RAG combines faithfulness, correctness and poisoning signals

Trustworthy RAG combines faithfulness, correctness and poisoning signals

The reported detector needs the complete scoring formula and task-specific calibration; detection is not guaranteed prevention.

Trustworthy RAG evaluates response correctness alongside faithfulness and poisoning-related signals. Its linear score component is 0.4F + 0.35C + 0.25(1 − P), followed by nonlinear damping. The original article rewrote the final term as −0.25P and omitted the +0.25 constant. That changes the score and invalidates ported thresholds.

The reported TruthfulQA performance and poisoning-detection metrics belong to the paper’s models, datasets and injection settings. A high answer score does not establish that the system detects every attack, and detecting an attack is not the same as preventing harm.

Reproduce the full scorer, damping and calibration on the intended domain before operational use. Clinical or other high-stakes suitability is not established by this benchmark alone.

Editorial correction (2026-10-13 UTC): Restored the score’s missing constant and calibration conditions.

Sources

Topics: RAG, adversarial-robustness, knowledge-poisoning, detection