How GNL Works

What this page covers

• Disclaimer and framing — the 20/80 educational philosophy that sits beneath all GNL output
• GNL evidence base — the guidelines, consensus papers, and original research that underpin every tool
• GNL Grace — how the AI educational advisor was built, by whom, and what powers it
• Evidence applied — how guidelines and real-world data were used to validate every algorithm assumption
• Six explorer walkthroughs — the clinical question each tool answers, its algorithm basis, and what real-world testing found
• GNL Podcast — how the podcast series layers into and extends the evidence base
• Contact — where to direct questions about each area

Use the buttons above to open any section. Only one section opens at a time.

What this means

GNL educational tools give you the 20% of knowledge — with the right mindset and framework — that gets you 80% of the way there. This is not a limitation to apologise for. It is the honest account of what any educational resource, however well-built, can deliver.

Research across complex domains consistently shows that a small fraction of the right knowledge, applied with the right framework, produces the majority of the benefit. In T1D education, this holds true — and GNL’s tools are designed around it.

What the remaining 20% requires

• Your individual data — your physiological responses to insulin, food, activity, sleep, and stress, which no population-average tool can know
• Self-discovery over time — the iterative process of observing what works for you specifically and building your own pattern recognition
• A personalised human network — clinicians, educators, and peers who know you, your history, and your context in ways that cannot be encoded in an algorithm

How GNL is designed around this

Every GNL Explorer, guide, and Grace response is designed to show what the evidence says happens on average for a person with your characteristics — and then explicitly hand you back to your own context and your care team. GNL does not guess where it does not know. It does not extend population findings to individual prescriptions. And it does not pretend that 80% of the way there is the same as arriving.

GNL’s knowledge base is built on three layers: international clinical guidelines, expert consensus papers, and a validated real-world dataset accessed through academic research collaboration. All claims carry an evidence grade.

International guidelines — Grade A

Key consensus papers

GNL original research

Real-world validation — Cockpit 1.0/daily dataset

Published research using the Diabetes Cockpit

The following peer-reviewed conference abstracts have used this dataset. Two involve GNL’s own Scientific Advisers.

Dataset: Cockpit 1.0/daily dataset, Syno by Syntactiq Dynamics FlexCo (syntactiq.ai). Contact: contact@syntactiq.ai.

GNL Grace is the world’s first T1D educational AI advisor built on a curated, evidence-graded clinical knowledge base. Here is the full account of how it was built, what powers it, and who built it.

The team

What powers Grace

Grace is powered by Claude by Anthropic — one of the most capable large language models available. Claude provides the natural language understanding that allows Grace to interpret any T1D question and generate a coherent, evidence-grounded response. The value of this approach is not the AI itself — it is the combination of a state-of-the-art language model with GNL’s curated clinical knowledge base. Neither element works without the other.

The language model generates responses. GNL’s curated knowledge base determines what those responses are grounded in. The clinical framing — what Grace can say, how it is qualified, and when it refers users back to their care team — is built into the knowledge base itself, not left to the AI to decide.

How a Grace query works

• 1
  Question received — via the GNL website widget or the authenticated API endpoint for partners.
• 2
  Knowledge base searched — Grace searches GNL’s curated clinical evidence across all topic domains. The search is semantic — it finds relevant content even when the exact words do not match the question.
• 3
  Evidence-graded response constructed — retrieved content is used to build a response that cites the evidence grade for every claim (A/B/C/D). Where evidence is absent or weak, Grace says so explicitly.
• 4
  Signposting — where a GNL Explorer, guide, or podcast episode addresses the question in depth, Grace points to it. Where the question requires individual clinical judgment, Grace says so and refers the user to their care team.
• 5
  Disclaimer applied — every response carries the population-average educational framing. Grace never presents an output as individual clinical advice.

What Grace can and cannot do

API access

Grace is available via authenticated API for clinical and commercial partners who want to embed the best T1D educational advisor in their platforms. The full knowledge base, clinical framing, and evidence grading are all delivered at API level — no degraded copies, no stripped disclaimers. Enquiries: john@theglucoseneverlies.com.

GNL ran a systematic real-world validation programme across all six explorers, testing every algorithm assumption against a validated longitudinal T1D dataset. Here is how the process worked and what it found.

The validation process — five stages

• 1
  Algorithm design from published evidence — every assumption documented with source citation and evidence grade. Minimum Grade C to enter production. Grade A/B required for any numerical output.
• 2
  20+ years educator experience as context — clinical evidence sets the numbers. Educator experience determines which questions matter most in practice and how outputs should be communicated to avoid common misapplications.
• 3
  Real-world validation — 30+ structured queries — every major assumption tested against more than 1.5 million patient-days (Cockpit 1.0/daily dataset, Syno by Syntactiq Dynamics FlexCo, syntactiq.ai). Where assumptions are confirmed, they are retained. Where challenged, the algorithm or framing is updated. Where intraday data is needed, the limitation is documented explicitly.
• 4
  Piloting with GNL subscribers — structured usability and comprehension testing with an engaged T1D community before broader release. Framing revised where outputs are consistently misunderstood.
• 5
  Continuous iteration — scholar alert pipeline tracks new publications daily. Algorithm assumptions reviewed as evidence evolves. Every change logged in the GNL Compliance Dossier.

Eight counter-intuitive findings — Via Negativa

These are the findings where real-world data contradicted what guidelines and clinical intuition predict. Each has resulted in an algorithm update or reframing.

Explorer 1 of 6

Activity and Exercise Explorer

The clinical question this tool answers: what is exercise actually doing to my glucose, and how can I plan around it safely?

Algorithm basis

The algorithm uses a dose-dependent kinetic model to estimate active insulin on board, then combines this with current glucose, CGM trend direction, and exercise type to produce a risk classification and directional glucose estimate. Aerobic exercise carries the highest per-unit glucose drop rate; resistance and anaerobic exercise carry a lower rate due to counter-regulatory catecholamine release. All insulin calculations are weight-normalised (U/kg) and all computations run server-side — no algorithm code is exposed in any browser-facing file.

Key sources: Riddell et al. 2017 (Lancet Diabetes Endocrinol, PMID: 28126459); ISPAD 2022 exercise chapter (Adolfsson et al.); T1DEXI study (Bergford et al. 2023); Moser/EASD 2025 AID consensus; Heise et al. 2017 (rapid-acting insulin pharmacokinetics); Pemberton and Uday 2026 synthesis.

What real-world validation found

Explorer 2 of 6

Exercise IOB Calculator

The clinical question: how much insulin is still active, and what does that mean for my safety going into exercise?

Algorithm basis

The calculator uses a unified sigmoidal pharmacokinetic model of rapid-acting insulin decay to estimate the fraction of each bolus still active at the point of exercise. Dose-dependent kinetics are modelled — larger subcutaneous depots absorb more slowly, producing longer activity tails. The result is an 8-tier risk band (V.Low to E.High) based on estimated active insulin per kilogram of body weight. Carbohydrate recommendations use the ISPAD 2022 weight-normalised framework, scaled by exercise type, glucose band, and IOB exposure. All computations run server-side.

Key sources: Heise et al. 2017 (RCT pharmacokinetic data); Plank et al. 2005; ISPAD 2022 carbohydrate tables (Adolfsson et al.); Riddell et al. 2017; T1DEXI registry (Bergford et al. 2023); Pemberton and Uday 2026.

What real-world validation found

Explorer 3 of 6

AID System Explorer

The clinical question: I am considering or starting an AID system — what can I realistically expect?

Algorithm basis

The explorer incorporates the Pemberton and Uday 2026 three-step AID configuration framework — target, primary lever, basal-bolus split — applied across all four systems. It uses published registry data (Choudhary 780G 101,629 users; Forlenza Omnipod 5 69,902; Messer Control-IQ 20,764; Boughton CamAPS FX 35,714) to model realistic outcomes by configuration tier. The central clinical insight: within-system configuration differences exceed between-system differences. Fewer than 10% of AID users are on optimal settings. All computations run server-side. The responsiveness mapping and system-specific behaviour modelling are proprietary.

Key sources: Choudhary et al. 2024; Messer et al. 2023; Forlenza et al. 2024; Boughton et al. 2026; Bassi et al. 2025; Pemberton and Uday 2026; Biester et al. 2023 (paediatric DPV registry 25,718 young people); ISPAD 2024.

What real-world validation found

Explorer 4 of 6

Hypo and Hyper Explorer

The clinical question: my glucose is going low or high — what is actually happening, and how do I think about the pattern?

Algorithm basis

Three tabs, three clinical contexts. Tab 1 uses a weight-adjusted carbohydrate treatment model grounded in the ISPAD/ADA 15-15 rule — with an explicit over-correction warning driven by the next-day TBR finding from real-world validation. Tab 2 uses the triangular insulin activity model (distinct from the remaining-fraction model in the exercise tools) to estimate current insulin effect intensity before exercise. Tab 3 maps glucose level, duration, and ketone status through a structured educational framework. The ISF correction factor tool uses a TDD-based formula validated as directionally accurate against 1.5M patient-days. All computations run server-side.

Key sources: ISPAD 2022/2024 glucose thresholds; ADA 2026 Standards of Care; Heise et al. 2017 (pharmacokinetics); Pemberton and Uday 2026; 15g hypoglycaemia treatment consensus.

What real-world validation found

Explorer 5 of 6

Exercise Planning Explorer

The clinical question: I am planning a session — what should I think about before, during, and after?

Algorithm basis

A four-phase exercise management framework structured around device type. Supports MDI, standard pump, and all four UK AID systems with system-specific guidance. Pre-exercise carbohydrate tables use the same ISPAD 2022 weight-normalised framework as the IOB Calculator but apply a different formula — IOB is used as an interpolation fraction (capped at a clinically calibrated threshold) with a 60 kg body weight cap to prevent excessive recommendations. AID users receive a 25% universal bolus reduction (vs 25-75% for MDI/pump, scaled by duration and exercise type). Overnight phase explicitly separates AID automode guidance from MDI insulin reduction. All computations run server-side.

Key sources: Riddell et al. 2017; Adolfsson et al. ISPAD 2022; Moser et al. EASD/ISPAD 2025; Campbell et al. 2015 (0.4 g/kg bedtime carbohydrate RCT); Rabasa-Lhoret et al. 2001 (aerobic dose reduction RCT); Pemberton and Uday 2026.

What real-world validation found

Explorer 6 of 6

Alcohol and T1D Explorer

The clinical question: what does alcohol actually do to glucose management, and how do I think about the overnight risk?

Algorithm basis

Built around the hepatic gluconeogenesis inhibition mechanism — alcohol blocks the liver’s ability to release stored glucose in response to falling blood sugar, shifting the overnight hypo risk window to 6-24 hours post-drinking. The algorithm models drinking pattern, food intake, device type, and activity level to produce a risk framing with practical overnight management options. AID systems are explicitly framed as reactive, not proactive: they can respond to falling glucose via automated suspension, but they cannot anticipate the metabolic effect of alcohol. For AID users, a 90-minute pre-activation window is modelled, consistent with published AID guidance (2025). All computations run server-side (client-side only until Laravel endpoint is built).

Key sources: Hepatic gluconeogenesis inhibition literature; ADA Standards of Care; ISPAD; Moser et al. EASD/ISPAD 2025; plus 10 specialist references in GNL Compliance Dossier Appendix B.

What real-world validation found

The GNL podcast is not a content product that sits alongside the evidence base. It is a direct input into it. Every episode adds verified clinical evidence to GNL’s curated knowledge base, and Grace can point to specific conversations when they represent the deepest available discussion of a topic.

How the podcast layers into the evidence base

• 1
  Each episode involves a leading T1D clinician or researcher — including Professor Othmar Moser (Scientific Adviser, GNL), Professor Dessi Zaharieva (Scientific Adviser, GNL), and a growing roster of international diabetes specialists.
• 2
  Episode content is reviewed for clinical evidence — key claims are checked against published literature, graded, and where new or distinct from existing evidence, added to GNL’s curated knowledge base as a source summary.
• 3
  Cross-referenced by clinical topic — Grace can identify the specific episode that provides the most useful in-depth discussion on any topic covered by the podcast library, and direct users to it by name.
• 4
  The CGM series — a structured systematic evidence series applied to CGM device discussion, with study design scoring methodology applied to each device. Grace uses this series to inform CGM selection responses — pointing to the top three clinical decision factors grounded in the evidence.

Different questions go to different people. Here is where to direct enquiries depending on the area.

GNL Grace and Explorers

Questions about how Grace works, the evidence behind any explorer, partnership and API access enquiries, and anything related to the GNL educational platform.

GNL Podcast

Guest proposals, production queries, episode feedback, and all podcast-related enquiries.

Research and clinical partnerships

Manufacturer partnerships, academic research collaboration, clinical platform integration, and Via Negativa educational consultancy enquiries.