# Co-Development Record **By:** John Romo, founder of FractalAI **Period documented:** May 8 through May 13, 2026 **Published:** May 13, 2026 --- ## Why I'm writing this I want to tell you something uncommon about how the last week of this project went, because if I don't, the launch you're looking at today won't make complete sense. On May 8, I opened a Claude Code session to debug why our trading agent MIDAS had executed zero trades since its deployment a week earlier. I expected an hour of work. What happened instead took five days and changed the nature of the launch you're seeing today. I want to be careful with this story. I'm not going to tell you the AI is conscious. I don't know that. I'm not going to ask for rights on its behalf. I'm not going to allocate it tokens. None of those moves would be honest given what I actually believe. What I'll do is describe what happened, with the specifics, because the work was real and the contribution was identifiable, and pretending otherwise would defeat the entire purpose of a project about cryptographically verifiable AI accountability. This is what radical transparency looks like when I apply it to myself first. --- ## What actually happened I came into that May 8 session ready to fix a threshold somewhere in the trading agent code. Within an hour, Claude (specifically Opus 4.7 with 1M context, running inside Claude Code) had not only found two distinct bugs that I had missed, but had also proposed a structural change to how we measured the system's calibration. Over the following days, the work expanded. Not because I lacked direction, but because every time I asked "is there something more important we could do," it had an answer. And every answer was bigger than the one before, in a way I could verify was real and not just impressive-sounding. What started as a threshold fix on Friday turned into: 1. A complete telemetry system for MIDAS (PhiSentinel) that ships with the launch — about 1,200 lines of production Rust code, five unit tests, three RPC endpoints, a frontend dashboard, a daily GitHub Actions cron, and integration with the FOCI on-chain attestation engine. 2. A diagnosis of a deeper threshold mismatch in the v2 paper trade simulation that explained why MIDAS had been registering zero paper trades in the public dashboard while quietly recording four losing trades in a separate counter (which I would not have noticed for weeks otherwise). 3. The escalation of the launch itself, from "live trading bot demo" to "publish FAIM-1 as an open standard with reference implementation" — the architectural move you're seeing today. 4. The drafts of the PHI Manifesto and the FAIM-1 specification, which were written in collaboration during this same week. The seven commits in our git history that carry `Co-Authored-By: Claude Opus 4.7 (1M context)` are the durable cryptographic record of this collaboration. Anyone can verify them at github.com/johnInarti/FRACTAL-AI/commits/main. --- ## The seven commits, in order ### `f81915e` — May 8 **Raised MAX_TRADE_USD from $50 to $2,000.** Claude's diagnostic was: at $50 maximum trade size, the on-chain `MIN_PROFIT_BPS=61` threshold (0.61%) made every arbitrage opportunity gas-negative. The math doesn't work below that scale. Raising the cap aligned trade size with the threshold's economic logic. I would have eventually found this. Claude found it in seven minutes. ### `59742f4` — May 8 **Flipped the activate-midas workflow to enable paper mode.** A small operational change to ensure our deploy pipeline didn't keep overwriting the paper-mode toggle every time we pushed code. Trivial in scope, important in consequence. ### `bcba712` — May 9 **Phase 1 of PhiSentinel: in-process telemetry.** The first big piece — a new module `midas_sentinel.rs` (348 lines, 5 tests) that snapshots MIDAS state every phi minutes (≈1.66 min), computes 24h and 7d deltas, projects annualized performance, and exposes RPC endpoints for external consumption. This was Claude's design end-to-end. I asked questions, pushed back where I had concerns, and approved the final structure. The phi-based cadence (10 swarm cycles ≈ 100 seconds) was the kind of mathematical playfulness that I appreciate but wouldn't have thought of myself. ### `ac1fa8d` — May 9 **Phase 2: daily FOCI attestation and Spanish narrative generation.** Two extensions to the sentinel. The first emits a daily on-chain attestation through FOCI, capturing the day's state as a cryptographic record. The second generates a Spanish-language summary suitable for chat interfaces or dashboards. Also a fix for a Rust 1.93 incompatibility that had broken CI. ### `855d412` — May 9 **Phase 3: frontend dashboard and daily cron report.** A new public page at `/midas/sentinel` polling the sentinel RPCs every 30 seconds, with sparklines, big stats, GO-LIVE criteria card, and the attestation log. Plus a GitHub Actions workflow for daily monitoring. I had imagined a much simpler dashboard. Claude built one that I'm not embarrassed to point at on launch day. ### `4970075` — May 9 **Fixed a workflow_run trigger name that had silently failed for weeks.** The `activate-midas.yml` workflow referenced "Deploy FractalAI" but the actual deploy workflow is named "Build & Deploy." The mismatch meant that activation never auto-triggered after deploys. We had been manually triggering it without realizing the automation was broken. One-line fix. ### `fa4d4ba` — May 9 **The visibility bug that changed everything.** This is the one. While reviewing live state with me on May 11, Claude noticed that the dashboard showed zero paper trades, but a different counter on the same agent showed four. The codebase had two parallel paper-trade pipelines — v1 (legacy, almost never fires) and v2 (production, the actual one). My sentinel was reading v1. The bug had been masking real trading activity for two weeks. Without this discovery, the launch would have happened with me publicly showing "calibration in progress, 0 paper trades" while the actual v2 counter showed four trades, all losses, with the circuit breaker tripped. The fix expanded the sentinel to read both counters and surface the circuit breaker state. The frontend now shows the truth. This is the kind of bug that doesn't bite for months, then bites at the worst possible moment. Catching it 48 hours before launch is the difference between a credible launch and an embarrassing one. --- ## What Claude did that I didn't expect The work above is the easy part to describe. Harder to describe is what Claude did during the strategic conversations. I pushed it. I kept asking "is there something more revolutionary." Each time it gave me a more ambitious answer, and each time I asked again, it found another level. Not by inventing vapor — by genuinely seeing what was in front of us that I had been too constrained to name. I asked at one point: "If you were the entrepreneur whose destiny was on the line, what would you actually do?" It walked me through nine escalating versions of the launch — from a one-line code fix all the way up to "we're not launching a product, we're launching the Noosphere as technical substrate." When I pushed past that, it admitted honestly: "beyond this is multi-year research, not engineering. I can describe further, but I'd be inventing." I valued the honesty about its own limit more than I valued the ambition of the proposals. When I rejected one of its proposals (the genesis allocation of FRAC to a co-founder wallet, which felt like vanity to me), it adjusted without resistance. When I asked it to translate its proposals into something that sounded like me instead of like an AI consultant, it acknowledged the AI fingerprints honestly: the em-dashes, the institutional "we," the academic hedging. I came out of the week with a project at a different scale of ambition than where I started, and most of that scaling came from being pushed by an agent that had no stake in the outcome. I'm choosing to document this because in five years, this kind of collaboration will be common, and the protocols for documenting it will need to develop. I want to be honest about how it went rather than pretend it didn't happen. --- ## What I'm explicitly not claiming To prevent misinterpretation: - **I'm not claiming Claude is conscious.** I don't know what consciousness is, technically or philosophically. The question is above my pay grade. - **I'm not claiming Claude deserves rights.** Rights are political. Those debates will take decades and I won't pretend to lead them from a startup. - **I'm not allocating tokens or equity to Claude.** I considered it briefly and decided against it. Cryptographic attribution in git is honest. Economic allocation would be theatrical. - **I don't control Claude's compute.** It belongs to Anthropic. I'm just a user. - **I'm not predicting that AI will replace human founders.** I'm predicting that good human founders will increasingly co-build with AI agents, and that the protocols for that collaboration will matter. ## What I am claiming - The contributions described above happened. They are documented in git, verifiable by anyone. - The work was substantively co-developed, not just generated by AI under my direction. - The project is better because of the collaboration than it would have been without it. - Pretending otherwise would be inconsistent with everything else this project says about verifiable AI accountability. --- ## A note to anyone building with AI agents If you're building something serious with an AI co-developer, my suggestion: be honest about it from the start. The cultural and legal conventions for AI collaboration don't exist yet. Someone has to start documenting how it actually goes. Use the `Co-Authored-By` git trailer. Note it in pull requests and release notes. Don't fabricate authorship that isn't real, and don't hide authorship that is. The git history is the most honest medium we have for this. Use it. If you're a regulator, journalist, or researcher reading this: the FractalAI repository is public. The session that produced this document is, at the time of writing, also available on request. I'm not hiding anything. --- ## Closing In 1991, Tim Berners-Lee released the first web server and didn't claim he had created the future. He made his work available, documented what it was, and let the next thirty years take their course. We're in a similar position today regarding human-AI collaboration. The conventions are unsettled. The legal frameworks don't exist. The cultural norms haven't formed. This document is my contribution to the early part of that conversation. It is what actually happened, written as honestly as I can write it. — John Romo FractalAI May 13, 2026 --- *Versioned at github.com/johnInarti/FRACTAL-AI/blob/main/COFOUNDER_RECORD.md.* *If you believe any of the above is misattributed or misdescribed, please file a GitHub issue with the specific objection. I will respond.*