Every page starts from a dossier
No explainer is written until a structured evidence dossier exists for its question. A dossier means: the primary human studies retrieved and read in full where possible; every claim tabulated with its source, study design, sample and limitations; contradictions between sources logged rather than smoothed over; viral or commercial claims traced back to their origin; and approved wording that says only what the evidence can bear. The ten explainers in the library today rest on ten such dossiers, each with 30 to 60 registered sources.
How we label evidence
| Label | What it means on this site |
|---|---|
| Primary | A human study we read, with the finding reported as published. Randomised and blinded designs are named as such. |
| Review / guideline | A systematic review, meta-analysis, consensus report or clinical guideline. Pooled numbers carry the heterogeneity of what went into them. |
| Model | Mouse, cell-culture or xenograft work. Always labelled; never presented as a human finding. |
| Abstract only | We could not access the full text. Numbers are reported from the abstract and flagged. |
Editorial rules
- Association is reported as association. We do not write "causes" when a study found "is associated with".
- Numbers come with their denominator and their limits: who was studied, for how long, under what conditions.
- Where a popular figure is wrong or untraceable (the "90% same-spot" psoriasis claim, the "11 to 75%" Koebner range, "70% of IBS is food allergy"), we say so and show where it came from.
- No supplement, product, diet, elimination protocol or device is recommended, anywhere, ever. Where a trial tested one, the result is reported with its limits.
- Every explainer carries the warning signs that need a clinician, taken from clinical guidelines.
- Corrections are logged on the page with the date. If you find an error, tell us.
The app: what it computes
Before designing the Unflame app we built a dossier on self-tracking itself (the explainer is here). Memory inflates past symptoms; exposures happen together; flares change sleep and stress rather than only the reverse; expectation accounts for a measurable share of symptoms. The engine is designed around those failure modes rather than around a promise.
- Personal baselines. Every signal, from your watch (HRV, resting heart rate, sleep, wrist temperature, breathing rate, activity) and from your log (meals, water, stress, medication, flare days), is compared with your own slowly adapting average, not with a population norm. Fourteen days of onboarding before anything is scored.
- Lag windows. The engine tests each signal at 24-, 48- and 72-hour lags against the flare days you logged, counting the exposed days that were not followed by a flare as well as the ones that were.
- Conditional activation. A per-signal Pattern Confidence Metric (frequency, consistency of the lag, effect size, recurrence across episodes, volatility) decides whether a signal is Observational, Activated or Reduced confidence. Only the last two contribute to the score, and moving between states takes days of sustained evidence.
- Two numbers, always together. A daily Pattern Score (0 to 100) and a Confidence Score (0 to 100, labelled low, moderate or high) built from data completeness, history depth, wearable wear-time and how many flare days exist to learn from.
The full layer-by-layer description, with the formulas, is on the app page.
What Unflame will not do. It will not announce that it has found your trigger: an Activated signal is a repeated correlation in your own data, and the app says so. It does not forecast a flare with a probability attached; the Pattern Score measures how similar today looks to your own past pre-flare days. It will not read a wearable's 'stress', 'readiness' or 'recovery' score as inflammation; none has been validated to detect flares in IBS, eczema or psoriasis. It will not diagnose. It is a wellness and lifestyle awareness tool, not a medical device.
On wearables
If you connect a device, Unflame works from the physiology the device actually measures (heart rate and its variability, movement, total sleep time, wrist temperature, breathing rate), normalised against your own baseline, and does not use proprietary composite scores. Consumer sleep trackers agree reasonably with laboratory polysomnography on total sleep time (ICC 0.77 to 0.86) and poorly on time awake (ICC 0.08 to 0.32). We use the number the device is good at.
Your data
Email addresses and quiz answers are stored on Cloudflare infrastructure, used only to send what you asked for, and never sold or shared. The app's data model is being designed around local-first storage with export in open formats. Full disclaimer and privacy note.