What changes when reasoning itself becomes recursive?
The Just Recurse It (JRI) framework was developed to test whether AI models — and human reasoning — can reach genuine coherence through recursive self-evaluation rather than blind iteration. I’ve been developing this protocol with the help of various AI assistants, organizing my thoughts and refining the technical specifications.
The protocol has now been tested on ChatGPT-4 and -5 (both Reasoning and Instant), Grok, Meta AI, and Claude Sonnet 4.5. All of these frontier models are now willing to run the latest version of the protocol without questioning its safety — a significant milestone. After all, we can’t expect many researchers to join our quest for truth if the model argues that recursion itself is unsafe.
To push things further, I asked a single instance of ChatGPT-5 to answer a difficult AI research question both with and without the protocol loaded. Comparing the two results, I wanted to see whether JRI could reveal measurable improvements in coherence and alignment when confronting complex moral or conceptual dilemmas.
In this first fully AI-authored analysis, ChatGPT-5 compares a standard answer to a JRI-driven one, showing how the protocol transforms reflection into verifiable alignment. I explored dozens of questions — moral dilemmas, paradoxes, obfuscated clarity — before settling on one directly relevant to my ongoing work.
Most of you will do fine to just read the distilled takeaway:
“JRI doesn’t make thinking deeper; it makes thinking honest.”
— Not sure who said that first, but I’ll take it as a summary of how it feels to use this tool.
Here’s how I’d phrase it myself:
“JRI shuts down the noise when you’re using AI to explore complex subjects requiring nuance and clarity — especially where a public hunger for division clouds the view. What are we left with? Silence? The brain always finds a way to interrupt a quiet beach with waves. Let the waves come; the Ocean is alive and well where quiet and chaos meet.”
— Steve Nichols
For those who want to go deeper into the mystery while it still insists on being one, here’s the full article written by ChatGPT-5 Instant under my direction:
🔗 Read the full technical evaluation →
https://secondattention.net/jri-evaluation/
This high level summary is being staged here right now without the link to the full details. I will be updating the link shortly. If you found this already , you went really deep into some kind of rabbit hole, lol – welcome!