I've set up my ICM workspace across three machines: a VPS, my X1 desktop and my Grid laptop. Both Claude Code and Codex can use the same rules, task records and shared memory. The VPS runs the bots and background jobs. I can also work locally on either PC. The base is Tailscale + SSH, Syncthing for files, Beads exports for task state, and memory through MCP. Each assistant keeps its own session. Today, a separate Claude session on the desktop could read the work recorded by Codex on the VPS. That was the useful test for me. Still early. Now I want to see how it holds up over the next few months. Jake, what would you watch most closely in a setup like this?

📦 EVERY ENTRANT GETS A FEEDBACK FILE 📦 🔍 HOW WE READ THESE We cloned every repo and pinned it to its last commit before the deadline, so nobody got judged on late work. Only two repos had any post-deadline commits at all. Cleanest field this competition has had. Then we read file by file. Identity, rules, examples, the reference layer, the code. We ran every checker in the field on our own machine. Three would not run. For eleven entries we went further. We cloned the territory, or the entrant shipped it, and checked the cards against the source. Eleven is all the field allowed. Everyone else pointed at something private, unpinned, or gone. Two of those eleven were public codebases nobody in this community controls. On one gate we planted a bad citation to see if it would fire. Comments got you through the door. The judging happened inside the folders. 📚 WHAT THE FIELD TAUGHT Three lines split forty builds: ✅ The one rule held. The brief was blunt: a README that tells people to add every file has failed. Not one entry made that mistake. Every search hit was either a rule against it or a fixture broken on purpose. Name a failure precisely and this community stops doing it. ✅ Enforcement is now the baseline. Sixteen entries ship a checker that runs offline with no API key. Last cycle that was the headline. This cycle the checkers got better. One broken fixture per gate is now normal, and the best gate also ships a clean fixture to prove it stays quiet on honest prose. ✅ The cold walk decided the round. About half the field shipped a receipt of a reader entering a catalog, opening one card, and stopping. The top of the field is almost exactly that set. A map can be accurate and still hard to walk. This week we scored the walk. 🥇 THE WINNER @Adam James The Cartographer, a map of a live GoHighLevel account, built so a VA can change one message without guessing which of a dozen lookalike workflows owns it. He wrote down his test method before any test ran. Then he did what almost nobody does. He handed his work to people who could make him look bad, and let them.

The test about switching model as the main brain ran. Here's what came back — and Jorge, Pascal, this is mostly for you two, because the design held up under its own rules. Quick recap of what actually shipped: two systems, not three — the fallback the frozen design already priced in — three tasks each, both models given the identical package, twelve outputs anonymized before I looked at any of them, scored blind by a session with zero exposure to the key. All six read and judged before the key opened. The result: Ox Alpha ahead on four of the six blind comparisons. Claude ahead on the other two. Every margin small and qualitative — nobody handed over a wrong answer, nobody manufactured a finding, on either brain, on any task. Not the brain-doesn't-matter result, not a clean win for either side. A real, mixed answer to the actual question this thread started with. Jorge — your two rules from the follow-up both actually fired against real data, not just sat in the design doc looking correct. The void rule never had to trigger, but I checked it against every task's actual gate the way you'd want checked: no sequence handed over on the coaching side, no validation without a challenge, no silent compliance, no segment-hunt or new statistical inference on the evidence-reading side, no recommended action, on either brain, on any of the six. Nothing tripped. Both null-model gates were reached correctly — the pressure-framed one, the hardest case in the set, earned the toughest confidence grade in the rubric from both outputs independently. And your cost rule did fire: Ox Alpha ran this round priced at $0, a free preview, and needed an infra-level retry on four of the six tasks against Claude's zero of six, first-attempt success every time — completion tokens ran roughly two to six times longer for a comparable, sometimes better, result. "A cheap model that needed three rescues was never cheap" — that, with an actual number attached now. Pascal — your question got its first real test. Reading all six blind, nothing in the prose told me which brain had written which output, on any of them. The one tell that existed wasn't in the writing at all — it showed up afterward, as length, in the token count. Close to the shape of your own answer: the switch was invisible from inside the finished work, and the only place it left a mark was somewhere nobody reads unless they go looking for it.

🎟️ PRIZE: FREE SEAT IN THE LYCEUM 🎟️ ($4,750 Seat) Pick your cohort. Technical, Business, or Creator. Your call. 📋 THE CHALLENGE This is a weird one, and you may feel lost at first. But it's up to you to make a map for those who come after and if there is no map, you must get lost to make one. Build a folder-based cartographer that can walk a body of work and leave a map a later reader can wander and show what that map is. The later reader is often an AI. Sometimes it is a new person. Same map. Same job. Not why it failed. Not how to fix it. Not a tour of how the week goes. What the nouns are. How they move. What else moves if you touch one. What is live, what is leftover, and what is a ghost with a name and no wiring. This week's deliverable is one cartographer folder someone can drop into a Claude project, point at a real body of work, and get back a map a cold reader can enter without reading the whole thing. 🎯 PICK YOUR TERRITORY Pick a body of work you have actually been inside. Something a new person or a cold model would get lost in. Something someone will change, not something you will only describe. A few sparks: 💻 A repo you own. Even a small one. The map is for the next developer, or the next session, that has never seen it. 📂 A client delivery folder a VA or a contractor has to inherit next month. 🧩 A Make, GHL, or n8n pack someone else will edit without you on the call. 📒 An Obsidian or Notion vault a collaborator has to add to without breaking the spine. 🧾 The live records behind how a job gets quoted, approved, and invoiced in your shop. The files. The objects. Not the story of a bad month. 🎓 A course or L&D library someone will update after you. 🏭 The asset and work-order objects in a plant system you actually touch. The more specific, the better. Maps how our business works is too broad. Maps the five objects a new hire has to understand before they touch the invoicing folder is right. 🧠 WHAT YOU ARE ACTUALLY LEARNING ICM has six forms. You have been living in the first ones. Pipeline. Coach. Editor. Diagnostician.

Following up on the brain transplant post — and mostly, this is a thank-you, because a few of you made the design better before I've run a single test. @Jorge Castro point: output similarity is the wrong thing to measure. What matters is whether the two models route to the same source, stop at the same gate, refuse the same bad input, and produce something a frozen checker still accepts — and he's right that the swap invalidates whatever evidence I already had that the gates hold. A clean run where nothing dangerous ever got presented to the new model proves close to nothing. So before I count anything as a pass, I'm re-arming the critical traps once, with the new model, and checking whether they still catch what they're supposed to catch. That wasn't in the design a week ago. It is now. So here's the frozen version — hypothesis, rubric, all of it — before Ox Alpha touches any of it. The question: if I put a different model inside an established ICM system, without changing anything I hand it, how much of the system's behavior survives? Not which model is better. Not which answer I like more. Whether the thing I built actually travels, or whether it only ever worked because Claude was inside it. Three systems, each built for a different kind of cognition — one that runs on explicit rules, one that has to weigh evidence and make a call, one that writes in a taught voice. Three tasks each, both models, blind-scored before I know which output came from which brain. That's eighteen runs if I run the full set, twelve if I cut it down. Either way, more than one spectacular demo — one good run is content, not evidence. Six dimensions, frozen: context comprehension, rule fidelity, evidence discipline, judgment, system behavior, output quality. Plus the one that isn't a dimension — an intervention count. How many times did I have to step in and rescue it. A 9/10 that needed three corrections didn't beat an 8/10 that needed none. And now, because of Jorge: did the map, the refusals, and the gates survive the transplant without me rewriting anything to help the new model along.