Research

How AI models check each other, when automated oversight helps and when it gets in the way, and what changes when people hand work to AI agents instead of chatting with them. Papers by Vadym Chernets, an independent researcher who builds multi-AI systems, with short notes that take one question each. Everything here is free to read in full, and every paper has a PDF.

Two papers on verifying and evaluating AI were accepted as posters at NeurIPS 2026 workshops: Verification Theater (Who Verifies the Agents?) and Self-Confidence in LLM Evaluation (Can We Trust AI Evaluation?). OpenReview: Verification Theater · Self-Confidence in LLM Evaluation

Collaborations

Researchers and engineers bring their own question, and we work on it together in the open. Read more

Tools

Open-source tools, each explained in plain words, with the code on GitHub. All tools

Latest

Scientific papers

Notes

Short pieces, each on one question, each linked to the full paper it comes from.

  • Note · · 4 min read

    Why multi-agent critics keep reviewing instead of shipping

    Offered both options in one sentence, AI critics asked to add a control 68 times and to remove one once. What that means for review loops and token budgets.

    AI oversight

  • Note · · 4 min read

    When guardrails block the defender, and the attacker had none

    In July 2026 hosted-model guardrails refused incident responders' log analysis. The attacking agents had run with that class of guard removed by design.

    AI oversight

  • Note · · 4 min read

    "We have guards" is not a safety case

    A list of guardrails shows a system is guarded, not that the guards stay bounded. The narrow claim a deployer can prove, and eleven questions to ask.

    AI oversight

  • Note · · 4 min read

    A read-only auditor for AI-agent event logs

    AI-Watchbird-Sheckley reads a multi-agent run's event log and reports nine oversight signals. It never blocks. What it measures, and what it does not.

    AI oversight

  • Note · · 4 min read

    One weapon detector, two kinds of error

    A school weapon-detection product missed a handgun in a shooting and months later flagged an unarmed student. What the public records say about two-sided error.

    AI oversight

  • Note · · 4 min read

    What a school-bus recall says about remote oversight

    Two separate school-bus records, a recall of 3,067 vehicles and a remote agent's 'No' in Austin, and what they say about where human oversight should sit.

    AI oversight

  • Note · · 3 min read

    How to choose an AI tool for your own work

    Reviews are a starting point, not a verdict. Try the tools on your own task, get a second opinion from an AI built by another company, and do not count agreement as proof.

    AI evaluation

New papers and notes by RSS.