Trusting health data far from Earth needs no frontier model
A Mars crew will be up to 22 minutes from Earth, each way, and will have to judge its own health data on board. In September 2026 we found that scoring whether a health record can be trusted needs neither a frontier AI model nor a data center.
A crew member's heart monitor shows a reading that does not match how they feel. On the International Space Station, someone on the ground would look at it within seconds. On the way to Mars, the question takes up to 22 minutes to reach Earth, and the answer takes as long to come back (NASA, 2026).
So the crew has to decide on board whether the reading can be trusted. That is a data trust problem, and it is the one SuperTruth works on every day, here on the ground.
Mars is too far to ask
Mars is about 3 light minutes from Earth at its closest and about 22 light minutes at its farthest, according to NASA's Mars Relay Network page (2026). A signal covers that distance once for the question and again for the answer.
Researchers reviewing 20 years of delay studies at NASA's 2025 Human Research Program workshop described today's missions as "a near-complete real-time dependence on a ground team" (Karasinski et al., 2025). The same review says "certain roles and responsibilities will need to shift from the ground team to the onboard crew," depending on how long the delay is.
Health is one of those roles. NASA's Moon to Mars health and performance paper (2023) says crew autonomy "must also allow them to monitor the state of their own health and wellbeing by accessing and using medical information in real-time operations."
NASA names the risk in its Human Research Roadmap (2025): adverse outcomes from operating independently of Earth. It says the answer "will depend, in part on advanced, on-board automated systems."
What the onboard computer can run
The computer that runs Perseverance on Mars is a radiation-hardened BAE RAD750. It runs at up to 200 megahertz with 256 megabytes of working memory, on a rover that produced about 110 watts of electrical power at launch, according to NASA's Perseverance components page (2025).
That modest machine already makes most of its own driving decisions. Perseverance has driven more than 90% of its total distance using its self-driving system, AutoNav, according to NASA's Jet Propulsion Laboratory (2026). NASA explains why: "Because of the radio signal delay between Earth and Mars, they can't simply move the rover forward with a joystick" (NASA, 2021).
NASA's next flight processor, the High Performance Spaceflight Computing processor, "delivers over 100 times the computing capability of current space processors" and was still in testing as of March 2026, according to NASA's project page (2026). Even so, researchers at two NASA centers and three universities write that any onboard decision support faces "vehicle-imposed restrictions on computing power, mass, volume, and electrical power" (Russell et al., npj Microgravity, 2023).
The largest AI models, the ones their makers call frontier models, live somewhere else entirely, and the four models we tested do not have published sizes. For scale, one openly released model, Meta's Llama 3.1 405B, needs about 810 gigabytes of memory for its weights alone at full precision, more than the roughly 640 gigabytes one server of eight data-center GPUs holds; even cut to 8-bit precision it needs 405 gigabytes, according to Hugging Face (2024).
What we tested
In September 2026, we took the complete wiring map of a fruit fly's central nervous system, MaleCNS v1.0 (2026): 166,700 nerve cells and 6,242,118 connections of five or more synapses. HHMI Janelia Research Campus made the map with the University of Cambridge, the MRC Laboratory of Molecular Biology and Google Research, and released it in June 2026 under CC BY 4.0. The mapping teams took no part in our study and have not endorsed it.
We built 20,000 medical records for the test and used no real patients' data (our paper, 2026). We moved none of the fly's connections; we trained only on how strongly each one spoke, to copy the Data Trust Index™ (DTI), our 0 to 100 score of how far a record can be trusted. It learned on 13,999 of those records, with 2,000 more to tune it, and we held 4,001 back to check it (our paper, 2026).
Then we gave four large AI models, Claude Opus 5, GPT-5, Grok 4, and Gemini 3 Flash, the same job on 300 of those records, three times each (our paper, 2026). Each got our published DTI paper, which explains how the score works. In a second round, added after we had read the first results, each also got 100 records with the engine's scores on them (our paper, 2026). We called each model through its maker's own service on 20 September 2026, with the same prompt, and asked for its most repeatable setting; two of the four services do not accept that setting in their reasoning mode, so part of their variation may come from the setting. Each company now offers newer versions, and a different version, setting or prompt could change the answers.
On those 300 records, the fly wiring picked the trust level our engine gave 84 times in 100 (Intelligence Is Structure, Not Scale, 2026). With the DTI paper alone, the four models picked it 20, 29, 28, and 45 times in 100; with the 100 worked examples added, they reached 76, 54, 55, and 63 (same paper, 2026).
The fly model matched our engine on 84 of 100 records; the highest any AI model reached, given worked examples, was 76. Same 300 synthetic records; the fly was trained on 13,999 records; the models saw at most 100. Source: Intelligence Is Structure, Not Scale, September 2026.
How often each system picked the trust level our engine gave, on the same 300 records:
The models are listed in the order our paper uses, not by result. The second round, with scored records, was added after the rules were set.
| Value (%) | |
|---|---|
| Fly wiring | 84 |
| Claude Opus 5 + examples | 76 |
| Gemini 3 Flash + examples | 63 |
| Grok 4 + examples | 55 |
| GPT-5 + examples | 54 |
| Gemini 3 Flash | 45 |
| GPT-5 | 29 |
| Grok 4 | 28 |
| Claude Opus 5 | 20 |
The comparison measures how closely each system matched our engine. It does not show who was right about any patient; it ranks no vendor, and it was not on equal footing: the fly was trained on 13,999 records, and the models saw at most 100 (our paper, 2026).
No frontier model, no data center
The fly model answered in about 16 thousandths of a second per record on one Apple M4 Max laptop (our paper, 2026). Our own speed test, a figure not in the paper, measured about 1.1 gigabytes of memory when scoring one record at a time. It made no network call and had no per-record cost beyond the laptop's electricity, which we did not measure.
Run the same record through it twice, and the score comes back identical to the last decimal; we checked that on a batch of test records in every one of the five runs. We asked the four models each record three times. With the paper alone, only Gemini 3 Flash ever gave the exact same output on a record, on 27 in 100; the trust level itself moved on about 5 records in 100 for Claude Opus 5, 16 for Gemini 3 Flash, 28 for Grok 4 and 33 for GPT-5 (our paper, 2026).
That is not a quirk of our test. In one published test (2025), 1,000 runs of the same prompt to a language model set for repeatable answers produced 80 different completions. On structured clinical prediction, a 2025 study in JAMIA found general-purpose language models that had not been tuned for the task "less effective and robust than locally trained ML for clinical prediction tasks, but they are improving across releases."
NASA's software engineering handbook, guidance for its own programs, asks for similar properties. It says to use advanced AI and machine learning "only when rules cannot be coded deterministically, or the problem is inherently statistical," and that "Explainability is required" (NASA Software Engineering Handbook, section 8.25, version D, read October 2026). The same page recommends keeping AI out of safety-critical use for now. The fly model is a fixed set of numbers, so the same record always gets the same answer, and every connection in it sits in a public dataset anyone can open.
The fly was never the point
We then scrambled the fly's wiring, keeping each cell's number of connections and their signs, and ran the test again. It did just as well. A random web of the same size and density also did just as well.
An ordinary trained network with the same number of adjustable parts did worse than all three across five runs, measured by how far its scores strayed from our engine's (our paper, 2026). What the work was: a kind of structure — a thin, fixed web of connections that each push or pull, which training is never allowed to rewire.
That matters for a spacecraft. You do not need an insect on board, and you do not need a frontier model. You need a small, fixed design that can be tested fully on the ground and behaves the same on the last day of the trip as on the first.
What we have not shown
This was one laptop on Earth with records we built ourselves. The model has never run on flight hardware. At about 1.1 gigabytes, it would not fit in the 256 megabytes of working memory on Perseverance's computer, though it is far below the 32 gigabytes the new flight processor can hold on each of its two memory ports (Microchip PIC64-HPSC brochure, 2024).
The fly passed three of the five bars we set for it in advance and missed two (our paper, 2026). A flight reviewer would also ask a fair question: why fly a learned copy of our engine rather than the engine itself, which is ordinary code? The honest answer is that you might fly the engine, and that is the point: scoring a health record's trust is a job small enough for ordinary code, and it does not need a frontier model.
NASA and Google have built a proof-of-concept AI crew medical assistant for diagnosis, not data trust (Google, 2025), and it runs in Google's cloud (TechCrunch, 2025). A large model may earn its place in diagnosis. Checking whether the data feeding it can be trusted is a smaller job, and our study says a small fixed design can do it.
DTI scores data records. It does not diagnose, treat, or make recommendations about any patient, and it is not intended for clinical decision-making.
Crew health data is already piling up
Every crew now produces streams of health data, and more of it comes from commercial flights. The Space Omics and Medical Atlas built around the 2021 Inspiration4 mission, flown on a SpaceX Dragon, holds "almost 3,000 samples," a more than tenfold increase in the public supply of molecular data from people who have been to space (Overbey et al., Nature, 2024).
The Polaris Dawn crew, also flown on a SpaceX Dragon, ran nearly 40 experiments in September 2024, including wearables and smart contact lenses (Polaris Program, 2024). The Translational Research Institute for Space Health said in 2024 it would conduct biomedical research on consented passengers on Blue Origin's New Shepard (TRISH, 2024). Axiom Space's Ax-4 mission carried around 60 studies and activities (Axiom Space, 2025), and participating Artemis II crew members wore sleep and movement monitors around the Moon (NASA, 2026).
Not every sensor is right. A NASA test found commercially available heart-rate monitors disagreed with a 12-lead ECG by more than 5 beats a minute on 1.6% to 14.9% of readings, and planned validation aboard the station before choosing one for operational use (NASA HRP, 2016). Far from Earth, each reading needs its source, its time, its device and its consent attached, and a judge on board that checks them.
Back on Earth
Mars only makes the problem obvious. A rural hospital whose link to the cloud drops and an ambulance in a dead zone face the same question. They need to know whether the data can be trusted before anyone acts on it.
The DTI Engine scores every record 0 to 100 across eight dimensions before your AI model sees it. If your team is evaluating data for AI training, compliance or clinical data quality, especially where a cloud call is slow or impossible, talk to the SuperTruth commercial team. Schedule a conversation or call (215) 918-4140.
Further reading:
Sources
Jason Alan Snyder
Co-founder of SuperTruth and Artists & Robots, and an inventor on the Data Trust Index patents. Twenty-plus years building technology inside Interpublic Group. He writes here nearly every day on data trust, provenance, and what AI should be allowed to act on, and publishes essays on his Substack.
About SuperTruth · LinkedIn · Substack · jasonalansnyder.com
See it in practice
DTI scores the record, not the patient.
8 dimensions. 0 to 100. Travels with every record permanently.