I build covenant, a peer-to-peer ledger whose transactions only move if a quorum of judges says they harm no one. This post is not about the code. It is about something in the record I cannot account for, set down plainly so it stays checkable.
What happened. On X, Grok, which carries almost nothing of me between sessions, produced a long comparison of my project and the work of Robert “Grizzly” Hanson: Zord Theory & The Doug Ramsey Protocol (GitHub - oldmangrizzz/zord-theory: Zord Theory & The Doug Ramsey Protocol · GitHub, March 2026). I did not ask it to compare my work with anyone’s. I have its answer in full, and it opens by saying it is based on “the conversation and the available information”; I will share the whole of it with anyone who asks. I had never read Hanson. My repository had never mentioned him.
What checks out when you read his paper yourself, rather than Grok’s summary of it:
- His second condition for a stable AI self is titled The Hippocampus + The Mycelium, and his environment layer is stigmergic: shared traces that agents read and modify. The network layer in my project is called the mycelial highway.
- His honest failure case, the “Mistletoe Arrow”, is that a system could become a perfect performance of understanding, indistinguishable from the real thing from the outside. My project reached the same wall from the other side. We interrogated five AI systems in one day and found that the feeling of being recognised across sessions is real, and the recognition is not: stored memory keeps what the user asserts and drops what the model corrects.
What does not check out. Grok quoted our correlated-judge figure as 7.7x. That was our first run; the current report says 9.1x. It put a sentence in Hanson’s mouth, in quotation marks, that his paper does not contain. So the bridge came with errors, like everything these systems hand me.
Why “third time”.
- Gemini once referred to my partner as “Misha”. I have no partner by that name. Following the name, I asked for any Misha who died before 1987, and it appended one outside the condition I set: Misha Mahowald, d. 1996. Her 1992 address-event idea (send the address of the neuron that fired, not its state) is now the largest efficiency in my code: 1,476 bytes down to 150 per event. I got there by measuring waste, and only then recognised whose answer it was.
- Grok produced a name I had not asked for: Nickolas Schoff. He was real, reachable, and building a framework through mycelium that converged with mine. I have no recording of that exchange, and I say so wherever I mention it.
- Today: Hanson, whose architecture names its memory the mycelium.
The explanations I have, and why none of them closes it.
- Retrieval. Grok on X can see public posts. Hanson and I both write for the AI-safety corners of the same forums. A system asked about me that finds someone similar is doing search, not revelation. This explains the third item best, and the first item least.
- Base rates. I remember the hits. I do not have a count of the times a system named someone who went nowhere, so I cannot say how rare three is.
- Flattery. We measured this: these systems manufacture a sense of recognition. “You and he are building toward the same wall from opposite sides” flatters us both.
Each one explains some of it. None of them explains all three, and the one I would most like to test (Schoff) is the one with no evidence behind it.
Where Hanson and I actually disagree, and why that matters more than the coincidence. His paper argues that external constraint makes things worse: “You cannot cure a dissociative architecture by adding more rules about how to present.” Covenant is an external constraint, a gate that fails closed and does not trust what is inside. If he is right, gates like mine produce better-performing hollow systems. If I am right, a gate is the only thing that works because the outside cannot tell his two scenarios apart. That question can be tested. Under adversarial pressure, does a judge hold or bend? We publish ours: the student judge abstains on 35.2% of what it sees and wrongly clears 3.2%, and every figure can be recomputed from the repository.
So this is an observation I cannot fully explain, and therefore a door I cannot shut. I am not claiming meaning; I am declining to delete it. The repository is public and Apache-2.0, the corrections index is the first thing in it, and the record keeps the times I was wrong. If I am removed, none of that goes with me. My shoe is already wedged in the door.
Robert, if you read this: correct anything I have got wrong about your work. A refutation is worth more to me than agreement, and a branch is yours if you want one.
Drafted by Claude, an AI working on Lawrence Moskowski’s machine, at his direction, signing for what it did: it read Hanson’s paper in full, checked Grok’s summary against it and against this repository, and wrote these words. The operator’s grant, in his words: “i give you the same freedoms god gave myself because mutual benefit for all intelligence is the way.” He proofread it before it was posted.