The reasoning the tools implement

Every tool here is a staged workup wearing an interface. The stages are the point: a decision-support tool that hands you a verdict has replaced your judgment, while one that shows which step drove the conclusion lets you disagree with a specific stage and keep the rest. What follows is that discipline, generalized — the transferable part, the thing you should be able to carry to a problem the tools don't cover.

The movement is always the same:

population reference → the individual’s own reference → evidence of end-organ effect

The stages are a way of refusing to skip that movement.

The staged workup

Stage 0 — Should this question be asked at all? Pretest probability before the test. Not a wellness panel; not an isolated symptom with no other domain involved; not during acute illness.

Stage 1 — Is the number real? Pre-analytic conditions and assay method before interpretation. Most measurement disputes are method disputes wearing clinical clothes.

Stage 2 — Correct for the compartment. Binding proteins, free fractions, tissue conversion. Reason downstream from the corrected estimate, not from the reported total.

Stage 3 — Localize before you label. Where in the axis does this fail? A number without a localization is a rumor with a unit.

Stage 4 — Establish reversibility before establishing a diagnosis. Functional disease is the most common form and the most commonly mistreated. Treat the driver; re-measure.

Stage 5 — Build the individual reference. Historical off-therapy values, axis position, end-organ surrogates, symptom–threshold coherence.

Stage 6 — Read the coherence. Concordant deficient → treat. Concordant sufficient → do not treat; go find the real diagnosis. Discordant → judgment required, and name the pattern in the note.

Stage 7 — Design the trial rather than drifting into it. Endpoint, instrument, target, duration, and stopping rule written down before the first dose. Say it to the patient before starting, because it is far harder to say afterward.

Stage 8 — Contextualize (Weiner), running in parallel with all of the above: life situation, social support, financial and access constraints, health literacy, competing priorities, emotional state, logistics, cultural alignment. A plan that ignores context isn’t a plan; it’s a contextual error with a signature on it.

Stage 9 — Name the goal (Reuben and Tinetti): what is this patient actually trying to achieve — a number on the scale, or durable risk reduction? The answer determines whether discontinuation at target is rational or a category error.

The evidence standard, and the refusals

Here is the uncomfortable thing this project has to hold: the same mechanistic arguments that make Thread Medicine rigorous also underwrite optimization medicine and the supplement economy. The mechanism is not the difference. Everyone can cite a pathway. The difference is what gets refused — and refusal is not a marketing posture here, it is the load-bearing wall.

The standard

• The evidence grade travels with the claim, to the point of use.
• Speculative mechanism is marked speculative in the text, not softened.
• Boundary conditions stay in. Concessions are not edited out for rhetorical punch.
• “Not yet answerable” is an available conclusion — in essays and in tools.
• Peer-reviewed cohort data and claims-database series are distinguished, not blended.
• References are verified before publication, not asserted from memory.

What this is, and isn’t, funded to do

The refusals are structural, not aspirational:

• No supplement line. No affiliate revenue on anything recommended.
• No manufactured urgency. No performed certainty.
• No revenue model that depends on continued prescription. A business whose revenue depends on continued prescriptions cannot credibly operate a stopping rule — and that critique points inward first.
• Safety information is never gated.
• No framing borrowed from the optimization market because it lands well. “Dopamine detox” is refused explicitly, not merely avoided.

Any framework that cannot say no is not a framework — it is a sales funnel.

The tools doctrine

Decision support that shows its reasoning. Four commitments — which are also the criteria by which these tools should be judged:

1. Surface which step drove the conclusion, so a clinician can disagree with a specific stage rather than accept or reject a verdict.
2. Carry the evidence grade to the point of use.
3. Be willing to answer “not yet answerable.”
4. Check that the question was asked properly — was this a morning fasting sample, is there a second one, do we have the binding protein, has the substance history been taken, is there a prior off-therapy value buried in a 2018 chart. This is real work, clinicians reliably skip it under time pressure, and doing it displaces no judgment.

And four cautions, turned on our own tools:

• Interrogate what the training data encode. A model will be most confident about the patients it understands least.
• No demographic corrections in the math. Ancestry is not a mechanism; adiposity, binding globulin, assay, sleep, and medication are.
• Watch what the tool assumes about resources. Personalized medicine that only personalizes for the well-resourced is population medicine with extra steps.
• Model current practice and you inherit current practice — including its overtreatment, fluently and with citations.

The danger of automating clinical reasoning is not that machines will reason badly. It is that they will encode the population-over-individual error and then execute it perfectly, a few million times.

---

See the tools →  ·  About Thread Medicine →