Fox News and MSNBC broadcast the same facts and produce opposite realities. ChatGPT and CaribChat receive the same cancer screening questions and produce opposite answers. In one case the mechanism is attention: what you attend to determines what you believe. In the other the mechanism is governance: what you govern determines what is true. The transformer paper was called “Attention Is All You Need.” It was wrong. Commitment is all you need.
The Post-Truth Machine
The modern information crisis is not a crisis of scarcity. It is a crisis of governance. A cable news anchor can cite the same Bureau of Labor Statistics report and tell you the economy is thriving or collapsing depending on which network signs the paycheck. A search engine returns ten blue links for “is soursop good for cancer” and the patient in Trinidad has no way to know which one will kill her. A large language model generates fluent clinical guidance using a training corpus that does not contain the phone number of a single Caribbean hospital.
These systems are not malicious. They are ungoverned. They route information according to attention: whatever the algorithm, the advertiser, or the training data optimizes for. None of them optimize for governed truth, because governed truth requires something no model provides out of the box: domain knowledge committed to an evidence layer that any auditor can trace. That absence is how you arrive at a world where the same 28 cancer screening questions, posed to two AI systems on the same day, produce 100% verified facility citations from one and 46.4% from the other.
The Arc
Two papers define the arc that brought us here, and each answered a question about mechanism while ignoring a question about truth.
In 2012, Krizhevsky, Sutskever, and Hinton published the AlexNet paper and launched the deep learning revolution. AlexNet won ImageNet by a margin so large that the field abandoned decades of hand-engineered computer vision overnight. The question it answered was whether neural networks could see. The question it left open was whether you could trust what they see.
Five years later, Vaswani and colleagues published “Attention Is All You Need” and introduced the transformer. Where AlexNet proved deep learning worked, the transformer proved it could scale to language, to reasoning, to anything you could tokenize. GPT followed. ChatGPT followed. The title was a provocation: you do not need recurrence, you do not need convolution. Just attention.
Both papers were triumphs of mechanism. AlexNet was a mechanism for seeing. The transformer was a mechanism for attending. Neither was a mechanism for governing, and that gap produced the post-truth world. Attention is a routing function: it determines which tokens influence which other tokens without determining whether the tokens are true. A transformer trained on one cable network will attend to that network’s patterns with exquisite precision while a transformer trained on the rival network does the same. Neither will notice the contradiction, because attention cares about correlation, not truth. The machinery was built. The governance layer was not.
The Compiler Insight
Governance is compilation. That is the insight that started CANONIC: the same formal process that transforms source code into a trusted executable can transform domain knowledge into a trusted AI system. A compiler enforces contracts. It does not interpret intent or attend to what feels right. It checks what was declared against what was produced and fails if the two diverge.
Git is the compiler. That is the deeper claim, and the reason the title of this post is not a metaphor. A git commit is a compiled state of truth: a snapshot of every governed file, cryptographically hashed, signed, timestamped, and immutable. The commit history is the audit trail, the diff is the evidence of change, and the merge conflict is the structural detection of contradiction that no cable network and no language model provides. When two news anchors produce opposite framings of the same government report, there is no merge conflict because there is no shared repository. When two contributors to a governed evidence layer assert contradictory facts, the compiler catches it.
The architecture that follows is simple enough to describe in one paragraph. Every governed scope has three files: a contract declaring what the system knows, a coverage map declaring what it can and cannot do, and a learning ledger capturing what it discovers in production. Those three files compose into a trust chain where child scopes inherit constraints from parents, so a Caribbean cancer navigation platform inherits from oncology, which inherits from clinical medicine, which inherits from the root governance contract. The evidence chain traces every claim to a source, every source to a version, every version to a commit. The entire structure compiles into a graph of 916 governed nodes traversed by breadth-first search. The same BFS that compiles the developer documentation also compiles the clinical evidence that CaribChat serves to patients in Trinidad. The same contract language that declares what a permit app knows about California building codes declares what a cancer navigation platform knows about Caribbean screening infrastructure. The compiler does not care about the domain. It cares about the contract.
Domain knowledge trumps everything else when CANONIC is the trust layer. AlexNet gave machines the ability to see. The transformer gave machines the ability to attend. Git gives institutions the ability to commit. But a commit without domain knowledge is an empty snapshot, and domain knowledge without a commit is an oral tradition that degrades with every retelling. Ten years after AlphaGo proved intuition was computable, we still cannot prove most AI systems are trustworthy. The compiler is how you prove it. The commit is how you mean it.
The Hackathon
On April 1, 2026, Dr. Marisa Nimrod proposed a hackathon in the CaribChat launch group on WhatsApp. The University of the West Indies was interested. Guardian Group had a three-person team ready. A coding hackathon at a Caribbean university, organized by a physician who does not have a GitHub account.
This is not contradiction. This is the pattern. A lawyer won Anthropic’s hackathon by building a permit app in six days. A cardiologist in Brussels built clinical follow-up guidance. A road technician in Uganda built an infrastructure appraisal pipeline. Thirteen thousand applied, five hundred got in, and the winners were domain experts who had never shipped software. The lesson everyone drew was that domain expertise beats coding. The lesson nobody drew was that prompting is good for winning a hackathon but hardly how you formalize AI across an institution.
A hackathon is where domain experts discover they can build. What happens Monday morning, whether the thing they built can be trusted at institutional scale, is the question that separates a demo from a product.
Prompting vs. Governing
Dr. Nimrod uses ChatGPT. She prompts it with clinical questions, tests its answers against a decade of Caribbean oncology experience, and relays the results to her network. When her CAOH colleagues tested CaribChat on real patients in Trinidad and asked what made it different from ChatGPT, the answer was not the model. It was the governed evidence layer underneath it.
The numbers are stark. The same 28 screening queries, submitted to both systems: CaribChat cited a verified Caribbean facility in every response, ChatGPT in fewer than half. CaribChat provided actionable navigation (a specific facility name with address or telephone number) in 100% of responses; ChatGPT managed 35.7%. For “Where can I get screened for cancer in Toco?”, CaribChat returned the TT Cancer Society mobile mammography units, Sangre Grande Hospital for referral, and the Cancer Society telephone. ChatGPT returned “local health centers” and “the Ministry of Health” with no number, no schedule, and no Toco-specific navigation.
ChatGPT does not know the mobile mammography schedule for eastern Trinidad because no one committed it to a governed evidence layer. That is the gap. Prompts are recipes: useful to the person who wrote them, invisible to the institution that depends on them, gone the moment the model updates. Governance declares the invariant. The model can change, but the contract is fixed, reproducible, and transferable.
The Board Member
Dr. Marisa Nimrod, MD, MPH, joined Dr. Dexter Hadley, MD/PhD, as the first outside appointment to the Board of Directors of the CANONIC Foundation, to fully govern CaribChat and help grow the Foundation globally. Appointed March 31, 2026, as Director and Chief Governance Officer for Caribbean Operations, she brings the domain knowledge that no compiler can generate and no model can hallucinate.
She is CEO of the Caribbean Association of Oncology and Hematology (CAOH), the region’s convening authority for oncology standards, and former President of the Trinidad and Tobago Medical Association. She holds an MD from St. George’s University (2010) and an MPH in Global Health (2019), runs M.D. Medical Consultancy Ltd, providing executive medical assessments, health education, and occupational health services across the public and private sectors, and is organizing the CAOH 2026 Annual Scientific Conference at the Hilton Trinidad, July 17-19: “The Next Decade of Cancer Care: Collaboration for Regional Impact.”
She secured endorsement from the Organisation of Eastern Caribbean States for CaribChat deployment across fifteen sovereign nations and from CARICOM across fifteen full member states and five associate members. She confirmed as co-investigator on the CaribChat IRB protocol, filing for exempt determination in a jurisdiction that has no research ethics legislation. She co-invented three provisional patent applications covering community learning ledgers, cultural competency attestation, and federated health governance across zero-legislation jurisdictions. She brought NCCN resource stratification guidelines to the table through CAOH’s direct partnership and proposed federated registries across CARICOM, because Caribbean populations represent rare phenotypes that no single island can study alone.
She does not have a GitHub account. She has never made a commit. Her domain knowledge, committed to the CANONIC governance layer by others, is worth more than any commit history, because domain knowledge is what makes the trust layer true.
The Ledger
The CaribChat community learning ledger tells the story that the commit history cannot. Over 30 days across ten Caribbean jurisdictions, the ledger accumulated 136 governed sessions that matured from exploration to active care navigation: screening in weeks one and two, epidemiology and treatment in week three, survivorship and active treatment in weeks four and five.
A community member spontaneously contributed HIV testing sites, treatment centers, and emergency hotlines from the Trinidad Ministry of Health. A clinician corrected the system’s assumptions about oncology services at Mt. Hope Hospital. A woman wrote: “I have a lump in my breast. Did a biopsy at Mt Hope and it’s taking a long time.” Then: “I worried that they going to have to cut my breast off and it will look ugly.” By March 31: “I just got diagnosed with metastatic breast cancer.”
These are people navigating cancer in a system that no general-purpose AI was built to understand, finding answers on a governed surface that traces every claim to a source. The community is the AI in caribchat.ai: not artificial intelligence but accumulated community intelligence, the navigation knowledge of every patient and clinician who asks a question and improves the answer for the next person.
The Evidence
Dr. Nimrod is first author on a paper targeting JAMA Network Open: “Community Learning Ledgers for Cancer Navigation in Small Island Developing States.” Nimrod, Roach, Wray, Fraser, Calleja, Kalia, Bajnath, Hadley. The paper reports 136 sessions across ten jurisdictions, head-to-head comparison against ChatGPT (100% vs. 46.4% facility citation, P<.001), and a privacy architecture that collects no personally identifiable information by design: three fields per session (date, question text, random UUID), no linkage table, no IP addresses, no cookies. Re-identification is architecturally impossible, which means the system governs without data protection legislation in jurisdictions that have none.
That is the post-truth exit ramp. Not better models or better prompts. Governed evidence: sourced, versioned, append-only, auditable. The same compiler that enforces CaribChat’s clinical contracts can enforce any domain’s contracts. The same architecture that governs cancer navigation in Trinidad satisfies GDPR data minimization and EU AI Act auditability requirements. The technology is domain-agnostic. The commitment to govern each domain’s truth is not.
The Close
CaribChat launches at CAOH 2026 in Port of Spain on July 17. The paper, the first to prove that governed AI outperforms general-purpose AI on health navigation in Small Island Developing States, using cryptographically ledgered community learning across ten Caribbean jurisdictions, goes to JAMA Network Open. The UWI hackathon will come, and domain experts at St. Augustine will build clinical tools in a weekend, discover they do not need developers, and wonder what comes next. What comes next is the question that every blog in this series has been building toward, from the compiler insight to the three files to the evidence chain to the graph to the lawyer who won: can you prove it is true?
Dr. Nimrod does not have a GitHub account. She prompts ChatGPT with clinical questions and waits for someone else to commit the results to the governed layer. ChatGPT is the middleman: a fluent, ungoverned intermediary between a domain expert and the truth she already knows. If she had a GitHub, the middleman disappears. Her domain knowledge goes directly into the evidence layer, cryptographically hashed, immutable, and traceable. No prompt required. No model required. No session that expires. Just a commit.
The UWI hackathon could teach a generation of Caribbean physicians, epidemiologists, and public health students to make that commit: not to prompt better, but to commit domain knowledge directly to a governed trust layer that outlasts any model, any session, and any middleman.
That is commitment. In a world drowning in attention, it is all you need.
Dr. Hadley asked Dr. Nimrod for her GitHub username three weeks ago. Maybe now she will finally get one and share it with him, so she can make that commitment herself at the speed of CANONIC itself.
Sources
| Claim | Source | Reference |
|---|---|---|
| Krizhevsky, Sutskever, Hinton, “ImageNet Classification with Deep Convolutional Neural Networks” | NeurIPS 2012 | papers.nips.cc |
| Vaswani et al., “Attention Is All You Need,” transformer architecture | NeurIPS 2017, Google Brain | arxiv.org/abs/1706.03762 |
| UWI hackathon proposed by Dr. Nimrod; Guardian Group and UWI interested | CaribChat.ai launch party WhatsApp group, 2026-04-01 22:41 TT | CAS: WhatsApp/120363426479033445@g.us |
| Dr. Nimrod’s CAOH colleagues asked what made CaribChat different from ChatGPT | Nimrod WhatsApp DM, 2026-03-27 16:48 TT | CAS: WhatsApp/45239075078224@lid |
| CaribChat vs ChatGPT: 100% vs 46.4% facility citation, 100% vs 35.7% actionable navigation (P<.001) | Nimrod, Hadley et al., JAMA Network Open (in preparation) | PAPERS/COMMUNITY-LEARNING/submission.md |
| 136 sessions, 10 jurisdictions, 30 days (March 2-31 2026) | CaribChat community learning ledger | TALKS/CARIBCHAT/LEARNING.md |
| OECS endorsement across 15 member states | Organisation of Eastern Caribbean States, March 2026 | oecs.org |
| CARICOM endorsement across 15 full + 5 associate member states | Caribbean Community, March 2026 | caricom.org |
| Dr. Nimrod appointed Director, CANONIC Foundation Board, 2026-03-31 | Board Resolution | CHARTER/BOARD/RESOLUTIONS/2026-03-31-NIMROD-APPOINTMENT.md |
| Dr. Nimrod: MD (SGU 2010), MPH (SGU 2019), CEO CAOH, former President TTMA | LinkedIn, CAOH institutional records | linkedin.com/in/dr-marisa-nimrod |
| CAOH 2026 Annual Scientific Conference, July 17-19, Hilton Trinidad | Caribbean Association of Oncology and Hematology | caohcaribbean.org |
| NCCN Resource Stratification Framework | National Comprehensive Cancer Network | nccn.org |
| PROV-007/008/009 provisional patents, Hadley and Nimrod co-inventors | CANONIC Foundation IP portfolio | PATENTS/PROV-007, PROV-008, PROV-009 |
| Mike Brown won Anthropic hackathon; 13,000 applied, 500 selected | “The Lawyer Who Won,” Hadley Lab, March 2026 | hadleylab.org |
| UWI AI Centre established | Trinidad Guardian | guardian.co.tt |
| Governance is compilation | “The Compiler Insight,” Hadley Lab, December 2025 | hadleylab.org |
| CANONIC graph: 916 nodes, BFS traversal | “MAGIC GALAXY,” Hadley Lab, February 2026 | hadleylab.org |
| AlphaGo proved intuition computable, not trustworthy | “AlphaGo at 10,” Hadley Lab, March 2026 | hadleylab.org |
| Prompts are recipes, governance is theory | “Stop Prompting, Start Governing,” Hadley Lab, March 2026 | hadleylab.org |
Figures
| Context | Type | Data |
|---|---|---|
| post | gauge | value: 255, max: 255, label: CONTRACT COVERAGE |
| *Commitment Is All You Need | GOVERNANCE + LAW | BLOGS* |