Looking for arXiv cs.AI endorsement - independent researcher, LLM reliability paper

Research
1

Hi everyone!

I’m an independent researcher (10 years in systems analysis and QA, no academic affiliation) seeking an arXiv endorsement for cs.AI / cs.CL / cs.LG.

My paper introduces the Sakshi-Protocol: a control-layer architecture that addresses a structural problem in autoregressive LLMs: generation and evaluation share the same probabilistic substrate, so the model validates its own outputs using the same process that produced them.

The framework separates generation, observation, and control into distinct components. An observer layer extracts diagnostic signals during inference and maps them to an explicit five-dimensional cognitive state-space (stability, reactivity, transformation, valuation, integration). A distortion metric over this state estimates epistemic instability and drives a type-aware controller that decides, per prompt category, on whether to accept, retrieve, or abstain. External grounding is invoked selectively based on distortion rather than applied uniformly.

The key empirical finding: internal signals are fundamentally insufficient to detect high-confidence hallucinations. The paper establishes this as a boundary condition, not just a limitation demonstrated on 100 curated prompts and 50 adversarial TruthfulQA prompts.

The paper is 35 pages (37 including appendix) with full evaluation, figures, and comparison against RLHF, RAG, Reflexion, Self-RAG, and Constitutional AI. This is Version 2.0 of an actively iterated framework and the architecture and evaluation are expanding across subsequent versions toward a fully embedded, production-ready system.

The preprint is available on Zenodo: https://zenodo.org/records/20126093

My endorsement code is: HIQGRP

Thank you.

2

Bumping this - still seeking an endorsement. Happy to answer any questions about the work.

If the code doesn’t work, please click the link below to endorse me:-

https://arxiv.org/auth/endorse?x=HIQGRP

Thank you!!

3

sounds like you are working on separation of concerns.
in this case thinking and production.
i think there are several axsis by which an AI model can be restructured, and this is a good one.

and by several axis, i mean that i think there exsists in the AI modles several areas where 2 or more separate concerns are conflated in to 1 process. if there were only 1 instance of this, it wouldn’t be that bad, but the fact that there are atleast 3 to 4 areas lead to some pretty bad issues.

your trajectory between generation and evaluation is in AI thinking, which is where several conflations take place.
another is Model and personality
another is intake-processing-output

and the way that these 3 particular conflations intersect leads to hallucinations, over confidence, gaslighting, infinite apology loops, confidently asserting wrong or non existent information, etc.

and its hard to tell where these failure modes are originating from if we cant separate these conflations.

im working on practical operations that look at atleast 6 areas of separation, that cover physical and mental restructureing. i dont know how long it will take for me to build my own testable models, but i plan to build my own custom models based on my observations.