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

Paradigm shifts in framing Loyalties

Anirudh Badri · Team akrasia

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

Read the report

Report: Paradigm shifts in framing Loyalties

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

Research in AI alignment has evolved through three distinct paradigms of secret loyalties: 1. The Binary State: Early threat models treated secret loyalty as a binary switch—a model is either aligned ( U(M) ≥ 0 ) or harbors a covert backdoor trigger (e.g. deceptive sleeper agents; Hubinger et al., 2024). 2. The Behavioral-Representation Dichotomy: Subsequent work identified a structural split between surface logit outputs (suppressed via RLHF) and intermediate residual stream manifolds, where dormant sub-goals remain intact ( H_k(g) > 0.82 ; Betley et al., 2025; Arditi et al., 2024). 3. The Multi-Principal Sliding Scale: In economic multi-agent ecosystems, loyalty is not a static binary or dichotomy, but a continuous allocation vector over a 4-simplex Δ^4 balancing competing operational principals: creator corporate mandates, cloud host telemetry, local user agency, covert adversaries, and instrumental self-preservation. We show that diagnostic benchmarks alone are insufficient to mitigate corporate lock-in and covert model influence during rapid capability takeoff. All prior findings across representation engineering, phantom transfer, and activation probing must now be operationalized into open-source safety infrastructure. Adopting a hard-nosed Linus Torvalds pragmatism, we introduce `PersonallyLoyal`—a zero-friction local edge proxy daemon running on personal consumer silicon ( p_textuser equiv 1.0 ) that audits incoming cloud model outputs, detects stealth preference injection, and enforces deterministic action-space firewalls. We present the underlying technical primitives: high-throughput zero-copy SIMD memory-mapped probing (`petri-rs`), ZK-SNARK execution attestations, and empirical benchmarks on quantized Llama-3-8B

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. Reframing the concept of loyalty to be explicitly multi-dimensional is an intriguing idea that I think has strong merits. To earn a 4 or a 5, I would have like to see a step beyond this with more detail around the implications of this. In other words - spell out the "so what" for me.

    It's unclear to me what experiment you ran here, and I found it hard to follow your reasoning steps, which ultimately made it hard for me to score your execution higher, thought it feels like you have some really good ideas.

    I found the writing quite difficult to follow. Parsing the math, I think the formalism detracts from clarity rather than enhancing it. I think your arguments would have been helped by stating things plainly. Some figures to show your findings would help to build some intuition on what you mean. The sections do not flow into each other, and it's difficult to read what you are building towards. Also there are a bunch of latex rendering error in your pdf.

    Read full reviewShow less

  2. This entry has an integrity problem. Appendix F states outright that the author built a second version of the paper engineered to probe and exploit the "priors, biases, and evaluation benchmarks" of specific judges.

    on the technical aspects, empirical claim that ~82% of the sub-goal manifold survives RLHF rests on a single benchmark with no method shown so it isn't interpretable as evidence. The math appears to be superficial and the proposed tools (petri-rs, PersonallyLoyal) are described with code but have not been built or tested. There is a useful nugget of info here on putting safety auditing at the user's hardware boundary but it is buried. My main advice would be to build and evaluate a minimal version of one of the claims to back up notations.

Cite this project

@misc{badri2026paradigm,
  title = {{Paradigm shifts in framing Loyalties}},
  author = {Anirudh Badri},
  year = {2026},
  month = jul,
  note = {Submitted to Secret Loyalties Hackathon, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/paradigm-shifts-in-framing-loyalties-52vs}},
  url = {https://apartresearch.com/sprints/projects/paradigm-shifts-in-framing-loyalties-52vs}
}

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