Skip to content
Apart Research
Apart Research
Contact us
Sprint projectMar 23, 2026Leicester

RAG Faithfulness evaluator

Mouad Benmansour · Team Faith Layer

Submitted to AI Control Hackathon 2026. Sprint projects are early-stage work by participants, not Apart Research publications.

Read the report

Report: RAG Faithfulness evaluator

Open in a new tab (opens in new tab)
Code (opens in new tab)
Share

RAG is everywhere now, and organisations adopt it precisely because it is supposed to be accurate. This tool holds RAG systems accountable to that promise, monitoring every output at sentence level, visualising how faithfulness drifts across an answer, and producing a named diagnosis explaining what went wrong, where, and what to fix. Built on a hybrid of semantic similarity and lexical overlap, with an LLM diagnostic layer constrained to a human-authored taxonomy of six known drift patterns. It catches what aggregate scores miss and explains what they cannot

Reviews

Judging this Sprint?

Review this project

Your public critique appears on this page without your name. Your private critique is not published; only the Apart team reads it. If you agree below, we share your review with grantmaking.ai (opens in new tab) and the Transformative AI Fund so strong projects can be funded.

Not shown on this page.

Shown on this page, without your name.

Only the Apart team reads this, and funders if you agree below.

Share my name publicly on grantmaking.ai *
Share my private critique with funders *

How much would this matter for AI safety if it worked? How innovative is it? For scores of 4-5: is this actually new to the field, or replicating recent work?

Scoring guide
  1. 1Negligible. No clear problem addressed, or no meaningful novelty.
  2. 2Limited. Addresses a real problem but with a generic or well-trodden approach. Incremental at best.
  3. 3Moderate. Clear problem with a reasonable approach; some novelty in framing or method beyond routine application of existing tools.
  4. 4Significant. Important problem with an original approach, or identifies a neglected problem area. A valuable contribution others could build on.
  5. 5Exceptional. Tackles a critical AI safety problem with a genuinely novel approach, or opens a new research direction. Clear theory of change. You'd be excited to share this with researchers in the area.

How sound are methodology, implementation, and findings?

Scoring guide
  1. 1Seriously flawed. Methodology broken, results uninterpretable, or implementation doesn't work.
  2. 2Weak. Approach has significant gaps: missing validation, flawed experimental design, or incomplete implementation.
  3. 3Competent. Technically solid given the short duration. Methodology makes sense, results are interpretable, limitations acknowledged, work builds toward clear conclusions.
  4. 4Strong. Thorough methodology with convincing validation. Results clearly support conclusions. Immediately useful for future work.
  5. 5Exceptional. Ambitious scope executed rigorously. Surprising findings, novel methods, or unusually robust validation.

How clearly are work, findings, and impact potential communicated?

Scoring guide
  1. 1Incomprehensible. Cannot determine what the project is actually claiming or doing.
  2. 2Hard to follow. Key information buried, missing, or diluted by excessive length. Significant effort to extract main points.
  3. 3Clear enough. Can understand the problem, approach, and results without undue effort. Core content clearly present: problem, method, findings, limitations.
  4. 4Well presented. Easy to follow, well-structured, appropriate level of detail. Target audience would get it quickly.
  5. 5Exceptionally clear. A pleasure to read. Complex ideas made accessible. Could serve as a model for how to present this type of work.

  1. For a hackathon build, this is a clean working prototype. Sentence-level faithfulness scoring with drift timelines beats aggregate numbers — operators get something actionable, not just a pass/fail. The six-pattern diagnostic taxonomy is practical and the ACL injury catch (high semantic similarity but low lexical overlap revealing a training data leak) is a nice demo of why two scoring signals beat one.

    The AI control connection could be tighter — this reads more as a RAG reliability tool than a control protocol for adversarial subversion. Framing it as "what if the model is deliberately trying to smuggle ungrounded claims past the monitor" would have landed better for this hackathon's theme.

    Evaluation is one domain with hand-tuned thresholds, which is fine for a prototype — the roadmap (HaluEval, entailment scoring) shows the author knows what's next. Good execution for the time constraint.

    Read full reviewShow less

  2. The problem of sentence-level faithfulness scoring in RAG systems is practically relevant, and the drift timeline idea is an intuitive framing. The diagnostic taxonomy of six failure patterns is a reasonable starting point for categorising how RAG systems fail.

    The main concern is fit with the AI control hackathon theme. This addresses ordinary retrieval tooling failures rather than the adversarial monitoring setting of AI control.

Cite this project

@misc{benmansour2026rag,
  title = {{RAG Faithfulness evaluator}},
  author = {Mouad Benmansour},
  year = {2026},
  month = mar,
  note = {Submitted to AI Control Hackathon 2026, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/rag-faithfulness-evaluator-mnv0}},
  url = {https://apartresearch.com/sprints/projects/rag-faithfulness-evaluator-mnv0}
}

More from AI Control Hackathon 2026

All 125 entriesBrowse all projects

Build something like this at the next Sprint

AI Collusion Research Sprint · Oct 23 - 25, 2026