# What these probabilities mean

For the conversational development, reusable steps and a worked example, see [Quantifying causal hypotheses: the reasoning exercise](../causal_hypothesis_quantification_process.md).

Literature and judgments reviewed 13 September 2026. These are hypotheses about structural damage to human peptidergic C-fiber nociceptors in diabetes. They are subjective scientific judgments, not patient risks, measured attributable fractions, treatment recommendations, or empirically calibrated forecasts.

## The proposition being priced

The default population is adults with type 2 diabetes before substantial distal neuropathy. The target is the population of peptidergic, unmyelinated sensory neurons supplying distal leg skin, including their cell bodies, axons and terminals. A mechanism can act through supporting cells or blood vessels, but its consequences must reach those neurons.

For comparability with the [preceding fatty-acid note](../fatty_acid_hypothesis_probabilities.md), a **material cause** means that selectively preventing the specified harmful route before injury would avert at least 20% of five-year net terminal-axon loss in an identifiable subgroup accounting for at least 10% of incident DPN cases in the stated population. This does not require the pathway to cause disease by itself. Necessary downstream execution mechanisms and upstream initiating exposures can both qualify.

The 20%, 10% and five-year choices are thought-experiment thresholds, not measurements. A YES probability of 20% is a probability *of that proposition*, not evidence that 20% of axons are lost through that route. A subgroup must be describable by an exposure or biological state independently of knowing who benefits; an unspecified collection of eventual responders is insufficient.

Files explicitly identify exceptions: type 1 diabetes, an acute treatment-associated syndrome, established injury/repair, or a diabetes-associated second insult. Those prices are not directly comparable with default-population prices. Their denominators and time windows are stated locally.

Peptidergic identity is not perfectly captured by one stain. Reduced CGRP or substance P expression can mimic disappearance of a fiber; PGP9.5-positive intraepidermal fiber counts do not identify peptidergic fibers specifically and can include terminals of other small sensory fibers. Human and mouse nociceptor classes are not interchangeable. Consequently, most cited structural studies support a broader small-fiber claim and leave the exact subtype uncertain. Human peptidergic-associated swellings have been observed, but do not identify their cause. [Cheng et al., human skin study](https://pmc.ncbi.nlm.nih.gov/articles/PMC3994562/); [human keratinocyte/C-fiber study](https://pmc.ncbi.nlm.nih.gov/articles/PMC8915676/).

Net terminal loss can arise from increased destruction, reduced replacement, or both. Repair hypotheses explicitly price the latter. Preserving a terminal does not necessarily restore its normal function. Pain relief, conduction velocity, stress-marker expression, nerve-fiber density and neuron survival are different outcomes.

## Derive each mechanism's logic before pricing it

There is no fixed number of facts. Start with the physical or biological requirements of the proposed cause: what has to reach what, what must change, why that change could defeat axonal survival or replacement, and what would distinguish cause from a response to injury. Then state only the extra propositions the particular hypothesis actually requires. Exposure chemistry, organelle transfer, enzyme-mediated destruction and impaired regeneration require different decompositions.

An initiating exposure may have multiple possible downstream routes. A signaling pathway may be protective below a threshold and harmful above it. A maintenance defect may matter only during an additional insult. These are branches or interactions, not automatically more events to multiply. Each file explains its causal structure before its probability table. Where several routes could satisfy a requirement, its price is for their union; no invented independence assumption turns them into separate factors.

Compatible timing and connected compartments are required throughout. Selective prevention is the causal meaning of the completed hypothesis. Neither is charged again as an extra generic “fact.” Holding other *initiating exposures* constant still allows consequences downstream of the blocked route to change. A repair hypothesis uses the loss-versus-replacement balance explicitly rather than pretending reduced outgrowth proves increased destruction.

For the particular requirements in a file, write A_i for the conjunction of requirements 1 through i. Each table gives q_1 = P(A_1), then q_i = P(A_i | A_(i-1)). The cumulative probability is q_1 × ... × q_i, using however many requirements that mechanism needs. **This multiplication is the probability chain rule; it does not assume the biological links are independent.** This is an ordering for assessing joint belief, not a claim that every biological mechanism is a linear pathway. Conditional prices already presume every predecessor is true in the same relevant human mechanism. Multiplying the cumulative column again would be incorrect.

The numbers are elicited judgments after reading the supporting evidence and its strongest relevant limitations. There is no fitted model that transforms a mouse rescue percentage into a human causal probability. The arithmetic makes judgments coherent and exposes where confidence is being lost; it does not validate the judgments.

**Detail must not create an automatic probability penalty.** Splitting one requirement into two equivalent subrequirements should preserve its joint price: a 60% claim might become 80% followed by 75% conditional on the first, still 60%. It should not become two fresh 60% discounts merely because the explanation became longer. Conversely, a broad claim with several alternative routes can be more probable than any specific route. The variable-length decompositions expose biological gaps; their lengths are not scores. The selected numbers remain contestable, and differences of five percentage points in an input are usually below what the evidence can resolve.

## How calibration was attempted, and its limit

No labeled set of resolved DPN mechanism forecasts exists here. Therefore no Brier score, reliability curve or demonstrated calibration is claimed. “More calibrated” in this collection means more explicit resolution criteria, stronger attention to reference examples and negative evidence, consistent conditional arithmetic, and wider uncertainty when the human bridge is missing. It does **not** mean that 70% forecasts have been shown to resolve YES 70% of the time.

The following anchors constrain the judgments:

| Evidence anchor | What it should do to a judgment | What it cannot establish |
|---|---|---|
| Human randomized prevention | Strongly support the causal importance of the altered exposure | Identify one molecular mediator or one fiber subtype |
| Human pathogenic genotype producing sensory axon loss | Strongly support human biological capability of that defect | Establish the same defect at ordinary diabetic exposure or its prevalence in DPN |
| Selective manipulation in diabetic animals with axon counts and unchanged glycemia | Support a structural causal route in that model | Supply a human effect size or eliminate species, sex, dose and disease-stage differences |
| Human diabetic nerve chemistry or pathology | Support anatomical/exposure plausibility | Establish timing, direction or cell autonomy in cross-sectional tissue |
| Prospective blood association | Support temporal plausibility if baseline neuropathy was actually excluded | Identify nerve exposure, mediator, or independence from obesity, kidney function and treatment |
| Culture toxicity, omic enrichment or analgesic rescue alone | Support possibility and locate a candidate link | Establish material structural injury in humans |

For the strongest exposure anchor, the DCCT randomized 1,441 people with type 1 diabetes. Confirmed neuropathy at five years was 5% under intensive treatment versus 13% under conventional treatment, a reported 64% relative reduction (95% CI 45–76%). This strongly supports a preventable glucose-related component in that population; it does not divide the benefit among polyol, carbonyl, vascular or mitochondrial routes. [DCCT neuropathy report](https://pubmed.ncbi.nlm.nih.gov/7887548/).

Negative translational anchors matter too. The 1,019-person rhNGF trial did not meet its primary efficacy endpoint; the 549-person ranirestat trial did not significantly improve its primary sensory-conduction endpoint; and the 460-person, four-year NATHAN 1 antioxidant trial missed its primary composite endpoint (P = .105), with some secondary benefits. These are reasons to distrust a near-certain human structural claim inferred from attractive animal rescue. They are not proof that NGF, polyol flux or oxidative damage never matters: drug exposure, selectivity, disease stage and outcome sensitivity limit what a negative trial rules out. [rhNGF](https://pubmed.ncbi.nlm.nih.gov/11056593/); [ranirestat](https://diabetesjournals.org/care/article/32/7/1256/27025/Ranirestat-for-the-Management-of-Diabetic); [NATHAN 1](https://pmc.ncbi.nlm.nih.gov/articles/PMC3161301/).

There is deliberately no numerical “animal-study success rate” prior: the examples above are a selected reference set, not an unbiased denominator of all mechanisms tested. Claiming a 10% or 30% empirical translational rate from them would add false precision.

## Quantifying uncertainty without inventing a confidence interval

Each link has three elicited conditional prices: skeptical, base and optimistic. Each complete scenario is multiplied separately. The skeptical scenario puts more weight on confounding, exposure mismatch, compensation and model-specific effects; the optimistic scenario gives the favorable human extrapolation more weight. Files explain which concrete uncertainties drive those changes.

The resulting span is a **model-sensitivity range**, not an 80% or 95% credible interval, not a sampling confidence interval, and not a claim to cover every defensible view. No probability weights have been assigned to the three scenarios, so the central price is the base-scenario product, not their average. Coordinated changes preserve the fact that species transfer or subtype identification can affect several links together. Sampling each link independently from an invented distribution would obscure that shared uncertainty.

Link prices use coarse increments. Products are displayed to one decimal place only so that the arithmetic can be checked. A product of 23.7% is a judgment of roughly one chance in four; the decimal is not scientific precision. Small differences between files do not constitute a ranking. Overlapping scenarios are a direct reason to resist such a ranking.

Every file also identifies the most consequential uncertain link, explains why the base judgment is neither near zero nor near certainty, and gives numerical updates under specified hypothetical public-evidence outcomes. Those updates are sensitivity exercises, not claims that the evidence already exists.

Removing one uncertain link entirely has a transparent maximum effect: replacing its q with 100% changes the complete price from P to P/q. This illustrates why perfect evidence for one enzyme can leave a whole human chain uncertain. An actual new result can update several dependent links and should be reconsidered jointly.

## Dependencies and scope discipline

These hypotheses overlap. Lipid exposure may trigger ER stress, calcium disturbance, oxidative damage, PARP activity and SARM1 in one neuron. Schwann-cell stress can cause both impaired support and toxic lipid release. The probabilities must not be summed, and the chance that “at least one is true” cannot be obtained by multiplying their complements. More broadly phrased pathways also have more ways to be true than narrow molecular branches.

For human tissue, donors are the inference unit. A hundred fibers from five people remain n = 5 for between-person inference. When only an abstract was accessible, the file does not invent animal numbers, human exposure measurements, blinding, or target engagement. A missing detail is a limitation of the assessment, not proof that the study omitted it.

The intended common market concerns incident disease or early loss. Most human nerve studies concern established DPN. Linking the two remains an inference. A selective intervention improving established disease would support causality there but would not by itself settle prevention before onset.