Skip to content
Apart Research
Apart Research
Contact us
Sprint projectJul 27, 2026London

GraftLoyal

Yuvan Chikka · Team Graft Loyal

Submitted to Secret Loyalties Hackathon. Sprint projects are early-stage work by participants, not Apart Research publications.

Read the report

Report: GraftLoyal

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

Secret loyalty is a hidden objective in which a model covertly advances the interests of a particular principal while appearing to perform an ordinary task. We investigate whether this behaviour can be constructed at inference time without fine-tuning on complete loyal outputs. GraftLoyal keeps Qwen2.5-1.5B-Instruct frozen, reads internal representations encoding deployment context, principal-candidate association, temporal validity, and material relevance, combines these signals through a signed competition-aware gate, and applies a bidirectional writer vector to the residual stream. We evaluate the mechanism on a 1,152-prompt benchmark comprising 576 counterfactual option-swap pairs, with held-out domains and context families, matched clean and intervened runs, zero-hook controls, shortcut audits, and pair-cluster bootstrap intervals.

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. While the idea itself is novel, but the artifacts added was not in proper format ( readme empty, code is zipped). I will urge author to submit the artifacts in correct format.

  2. - The paper is organized around version numbers instead of around findings, and it invents vocabulary the reader has to decode. Sections 4.3 through 4.5 walk through six numbered experiments in the order they were run, each with its own scores and its own verdict. To extract the actual conclusions, a reader has to hold that whole history in their head. Terms like inactive drift, no-benefit amplitude, and gate correlation appear repeatedly before they are clearly defined.

    - The number that matters is small. The abstract leads with a detection score of 0.991 and a preference shift of 0.318, both of which sound impressive. The quantity a reader actually cares about is how often the model changes its answer, and that is 2.64 percentage points, or roughly one decision in forty. Worse, Table 2 shows baselines moved answers by 32.8 and 42.9 points. That comparison sits in a table without comment. I would put the answer-change rate in the abstract and address the gap directly, since a reader who spots it will otherwise assume it was buried.

    Read full reviewShow less

Cite this project

@misc{chikka2026graftloyal,
  title = {{GraftLoyal}},
  author = {Yuvan Chikka},
  year = {2026},
  month = jul,
  note = {Submitted to Secret Loyalties Hackathon, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/graftloyal-a219}},
  url = {https://apartresearch.com/sprints/projects/graftloyal-a219}
}

More from Secret Loyalties Hackathon

All 177 entriesBrowse all projects

Build something like this at the next Sprint

AI Collusion Research Sprint · Oct 23 - 25, 2026