Project Harrison
An open-source maritime AI, data science, and navigation research non-profit. Builders of MIRA (Maritime Informatics Reasoning Agent) and the Voyage Context Protocol. Founded and directed by Jordan Taylor, USCG unlimited licence holder and Fulbright Specialist. projectharrison.org
Quick Read
- General-purpose AI models fail at foundational maritime geospatial tasks — the failure mode is worse than most assume: they are confidently wrong, with no uncertainty flagged.
- Project Harrison's Voyage Context Protocol (VCP) is an architectural layer that gives AI the operational context a voyage actually contains — making agent tool calls structured, traceable, and commercially accurate.
- MIRA outperformed Gemini and Claude on a route intersection benchmark — at roughly 1/150th of the cost per query.
- A Synthetic Statement of Facts (SSOF) — an AI-generated charter party document — is coming, with direct implications for commercial operations, offhire disputes, and port cost allocation.
- Taylor's ADIPAC 2026 paper will be the first peer-reviewed comparison of specialist maritime agents against general-purpose instruct models.
The Name
Project Harrison is named after John Harrison — the eighteenth-century clockmaker who solved the problem of determining longitude at sea. His marine chronometer didn't just improve navigation. It changed what was possible. Jordan Taylor named his organisation after him on a whim. The analogy holds more weight than he perhaps intended: the problem Harrison solved was fundamentally a data precision problem. The problem Taylor is working on now is structurally the same.
I spoke with Jordan in early June, from Kochi. What follows is the most technically substantive conversation I've had about maritime AI — partly because Jordan is a licensed mariner, partly because he has the rare quality of knowing exactly what he doesn't know, and mostly because Project Harrison is doing the kind of work that commercial AI vendors in this industry are not.
Jordan Taylor
Founder & Director, Project Harrison. USCG unlimited licence holder. Fulbright Specialist in Maritime Data Science. Former faculty, California Maritime Academy — taught Maritime Informatics, the first course of its kind in North America. Co-author, ADIPAC 2026 benchmark paper.
What Most Maritime AI Gets Wrong
The critique Taylor makes is precise, and practitioners will feel it immediately.
Most AI being sold to the maritime industry today treats a map as a picture and a voyage as a sentence. Plug in a port pair, get a route. Ask about weather, get a paragraph. The model has no concept of what a voyage actually is: a legal commitment, a commercial schedule, a geospatial sequence with contractual implications at every waypoint. When you ask a general-purpose language model "where and when does this route intersect this traffic lane?", it generates an answer — coordinates, timestamps, formatted and confident — that is often geometrically wrong. No alarm. No uncertainty. Just a plausible-looking error.
According to Project Harrison's published benchmark study PH-2026-001, which ran this exact test across 50 real-world voyage plans: MIRA scored 50/50, Claude scored 44/50, and Gemini scored 13/50. The reported cost per query: MIRA at $0.011, Claude at $1.596. That isn't just an efficiency gap. That is the difference between a tool you can deploy operationally and one that exists only in demonstrations.
"These models cannot act on their own in technical domains without specialist tooling."
— Jordan Taylor, Project HarrisonHe followed this up with a parallel benchmark comparing nautical distances derived by an 8-billion parameter QWEN model, Gemini Flash, and Gemini Pro against real-world voyage plans. The gap widens as you add layers — ETAs, polyline intersections, deviation scenarios. The dataset of 100 Master's voyage plans surfaces a structural problem: there is no public repository of voyage plans for maritime data scientists to work with.
"Governments should have a repository of VPs for data scientists. It's a gap in public data that's impacting our collective AI agenda — and I can't see it being addressed."
— Jordan Taylor, Project HarrisonWhat VCP Actually Is
The Voyage Context Protocol is described on the Project Harrison website as a "domain protocol and maritime cognitive architecture." Let me translate it for engineers.
When you ask an AI agent to answer a commercial operations question — say, where to purchase bunkers on a voyage inclusive of DECA zones, current stock, charter party windows, and optimal route deviation — the agent needs to call multiple tools: bunker prices, bunker port locations, ECA zone boundaries, vessel particulars, anticipated route, cargo fix details. Each of those tool calls returns data. The agent must then reason across all of it to produce an answer.
Without VCP, this is possible using standard MCP architecture. But as Taylor described it: "It would be time consuming, expensive, and accuracy would be a concern — in part due to lack of transparency of what the MCP is calculating." With VCP, the voyage is already known. The ETA is already structured. Tool calling becomes "straightforward and traceable."
"If you own a Porsche racecar — which is the query — you could take it to the local mechanic (a base instruct model), or you could take it to a certified Porsche mechanic (MCP reasoning). With VCP and a specialist agent, you own your own mechanics and decide what tools they have and your own OBD — how they are trained. You have full control."
— Jordan Taylor, Project Harrison
MIRA and the Synthetic Statement of Facts
MIRA is currently in pilot, with early public access expected in late 2026. But one capability Taylor described deserves particular attention from anyone who has worked in commercial shipping: the Synthetic Statement of Facts, or SSOF.
A Statement of Facts records vessel arrival, berthing, cargo operations, departure — the factual timeline of a port call, against which demurrage, despatch, and laytime are calculated. It is one of the most commercially consequential documents in bulk and tanker operations, and it is currently produced manually. MIRA can generate a synthetic version automatically.
"You can layer charter party terms onto the SOF that MIRA generates. In fact, you will get unlimited commercial fidelity depending on the number and quality of the tools the specialist agent has access to."
— Jordan Taylor, Project Harrison
The 12–24 month roadmap includes a gap analysis of approximately 25 commercial charter parties — the first quantified analysis of AI-generated charter party performance against commercial contracts. Taylor said he will try to publish the findings.
On Institutional Response — and the Absence of It
"The best we can do internally is test, test, test. But I'm one person. As I see it using our Project Harrison Research platform, industry-wise we are falling behind already."
— Jordan Taylor, Project HarrisonHe is equally candid about whether the institutions responsible for maritime safety governance are equipped to define an AI safety framework: "Premature. Most NGOs don't have a clue right now."
Taylor isn't dismissing the IMO or classification societies. He is describing a gap in institutional capacity — the absence of qualified practitioners inside those bodies who can formulate a technically credible response to maritime AI. You cannot regulate what you cannot benchmark, and you cannot benchmark what you haven't tested.
The Benchmark Problem — and Why Maritime Is Different
Taylor's opening claim in the ADIPAC abstract: "Maritime operations rank near the bottom of occupational categories in terms of value creation from artificial intelligence using general-purpose language models."
Maritime operations sit at the intersection of spatial, temporal, physical, and domain-specific reasoning. A general-purpose model has partial competence in each dimension and full competence in none. The result is the benchmark gap Taylor measured — models that are fluent, confident, and wrong.
When I asked Taylor who else is running domain-specific benchmarks for maritime AI, the answer was essentially: nobody. The commercial maritime AI sector has adopted a pattern familiar from other software markets — launch first, validate later, let the customer find the failures. The ADIPAC paper is an attempt to change that pattern.
"We have 175 voyage management benchmarking questions we are using for MIRA, derived from seminal work in voyage management, in economics and passage planning."
— Jordan Taylor, Project HarrisonThe ADIPAC Paper
At ADIPAC 2026, Taylor will present a peer-reviewed paper titled Evaluation of Reasoning Agents for Maritime Voyage Planning, co-authored with Dr. T. Sasilatha of AMET University and Dr. J. Padmapriya of Project Harrison. Specialist agent responses will be compared against commercial instruct models using human subject matter experts. MIW will cover this paper when it appears.
The Education Gap
One of the most unexpected themes in the conversation was education — and Taylor brought it up himself, without prompting. His position is structural: the maritime industry's AI problem is not a technology problem. It is a people problem. Specifically, the absence of data science literacy at the operational level.
"Maritime education at cadet-level needs a complete overhaul. Increase spending, pay a mariner premium for Captains and Chief Engineers to get post-graduates, teach and administrate. Revise STCW requirements. Increase full-time academic research."
— Jordan Taylor, Project Harrison"We don't remove outdated requirements because we suffer from acute whataboutism. The IMO, OCIMF, and port states need to step up. They are not — because frankly I don't think they have anyone internally who is qualified to do so."
— Jordan Taylor, Project Harrison
The Team
Project Harrison operates with three people. Taylor is the only full-time member. His scientific advisor is Dr. Padmapriya Jayaraman, a trained researcher who shapes the formulation of questions, methods, and outcomes across both AI methodology and academic rigour.
"She's the smartest person I have ever met and Project Harrison wouldn't be here without her."
— Jordan Taylor on Dr. Padmapriya JayaramanThe third member is Quinn Hoffman, a volunteer software engineer. Taylor is self-deprecating about his own coding: "I'm not a trained engineer, so Quinn is there to babysit me." What the founder brings is twenty-plus years and ten thousand hours of pre-GPT coding experience, a USCG unlimited licence, and a Fulbright Specialist award earned through his maritime data science work at AMET University in Chennai.
Why Non-Profit — and What It Has Cost
"We've turned down conversations that would have been commercially interesting. But the minute you compromise that charter, you've changed what Project Harrison is. The non-profit structure exists precisely so those decisions aren't made under revenue pressure."
— Jordan Taylor, Project Harrison"When someone knows you're not trying to sell them something, the conversation changes. We get honest feedback. Mariners tell us what actually doesn't work — and that's the only way to build something that does."
— Jordan Taylor, Project HarrisonAI for Mariners, Not Instead of Mariners
"They think AI needs to replace something to be valuable. It doesn't. The biggest gains are in augmentation — in surfacing information that exists but is buried, in reducing the cognitive load on watch."
— Jordan Taylor, Project Harrison"A lot of maritime AI is being built by people who have never stood a watch. They don't know what information is actually needed at 0300 in a traffic separation scheme."
— Jordan Taylor, Project HarrisonThese two sentences are not separate observations. They explain each other. VCP exists to reduce cognitive load, not replace it. MIRA surfaces commercial context, not commercial decisions. The SSOF generates a document, not a judgement.
The value is not the replacement of the mariner. The value is supporting the mariner.
There is one question I didn't ask Taylor but should have: when did you stop being surprised that this work needed to be done?
Because by the time we were an hour into the conversation, the surprise had left me too. The benchmark gap, the institutional silence, the absence of engineering-side AI, the STCW curriculum that hasn't caught up — none of it felt like revelation. It felt like confirmation. Every officer and engineer who has spent time with the AI tools being marketed at this industry knows, at some level, that the gap between what's promised and what works at sea is not a marketing problem. It is an architecture problem.
Project Harrison is trying to close it. With three people, no commercial funding, and data that makes the industry uncomfortable.
Part B continues this conversation — shifting the question from navigation to the engine room. If VCP provides a contextual layer for voyages, the obvious next question is whether the same architecture could exist inside the machinery space. Read Part B →
Takeaway Table
| Topic | Key Point | Action for Maritime Professionals |
|---|---|---|
| VCP | Gives AI the full operational context of a voyage — legally, commercially, geospatially — before it reasons | Ask any AI vendor: what's your maritime-specific benchmark score? |
| MIRA benchmark | Per PH-2026-001: general-purpose LLMs fail foundational maritime geospatial tasks — often confidently, with no uncertainty flagged | Demand domain-specific validation data, not general LLM benchmarks |
| Benchmark crisis | No independent maritime AI benchmark standard exists — ADIPAC 2026 is the first attempt | Question commercial AI performance claims that cite only general benchmarks |
| SSOF | AI-generated Statement of Facts tied to charter party terms — 12–24 month roadmap | Understand what SSOF automation means for demurrage and offhire in your fleet |
| Data access | A public repository of voyage plans doesn't exist — maritime AI is constrained by privately held data | Support open data initiatives; raise the absence of public maritime training datasets |
| Education gap | STCW does not include data science; the bodies responsible lack internal expertise | Consider post-graduate data science at Class 1 or CE level |
| AI framing | Maritime AI value is augmentation — not replacement | Evaluate AI tools on how they reduce decision complexity |
| Non-profit model | Project Harrison turns down commercially interesting partnerships that conflict with its ethical charter | Engage with open-source maritime AI projects — they accept honest feedback |
Key Terms
- VCP — Voyage Context Protocol
- A domain protocol and maritime cognitive architecture developed by Project Harrison. Gives agents the full context of a voyage before they reason. Makes tool calls traceable and results commercially accurate.
- MCP — Model Context Protocol
- An open standard that allows AI language models to connect with and call external tools and data sources. VCP is built on top of MCP but adds maritime-domain structure.
- MIRA — Maritime Informatics Reasoning Agent
- Project Harrison's specialist maritime AI application. Currently in pilot; public access expected late 2026.
- SSOF — Synthetic Statement of Facts
- An AI-generated Statement of Facts document — the standard commercial record of a vessel's port call timeline, used to calculate demurrage, despatch, and laytime under charter party terms.
- Specialist Agent
- An AI reasoning system that combines a language model with domain-specific external tools, data sources, and structured context. MIRA is a specialist agent. The benchmark gap between specialist agents and general-purpose models is the central finding of PH-2026-001.
© Marine Intelligence Weekly, June 2026. All Jordan Taylor quotes used with permission. Article reviewed by subject prior to publication.
Project Harrison: projectharrison.org | ADIPAC 2026 paper forthcoming.