What the probabilities mean

Reviewed 14 September 2026. The collection contains 45 hypotheses and 210 conditional claims about physical damage to, or inadequate replacement of, human peptidergic C-fiber terminals. These are literature-informed subjective judgments. They are not patient risks, measured fractions of disease explained, drug-success probabilities or empirically calibrated forecasts.

The reasoning process (background note outside this collection) describes how to construct an assessment. The fatty-acid worked example (background note outside this collection) illustrates the current arithmetic. The index and collection finding (background note outside this collection) give all current judgments.

Fix the proposition before assigning numbers

The usual population is adults diagnosed with type 2 diabetes within five years, without clinical DPN at entry. The outcome is five-year physical net loss of peptidergic terminals in specified distal skin territories. Each hypothesis states its exact eligibility, compartments, exposure window, finite baseline groups and intervention. Absence of diagnosed DPN does not guarantee absence of microscopic injury.

A material contribution ordinarily requires a selective intervention to prevent at least 20% of reference-care net terminal loss, in a baseline-defined group containing at least 10% of reference-care incident DPN cases. These are chosen truth boundaries, not measured biological constants. Some mechanisms additionally require a particular source, transfer step, mediation contrast, irreversible event or replacement contribution.

The intervention must be specified physically: what molecule, reaction, input or capacity changes, by how much, where and when. Define its external material and energy supplies where needed. Keep other initiating exposures comparable, while allowing downstream benefits, costs and compensation. An ideal intervention can require unavailable technology; it cannot be defined as “whatever change preserves the nerve.” A drug that does not implement the specified intervention does not directly test its effect.

Allowed alternatives are finite and stated in each note. A group is defined from baseline information, never from future response. “Some susceptible people” without a bounded group question is too vague to resolve. Named molecular families also need boundaries: a probability for a finite lipid panel is not a separate probability for every lipid that might later be discovered.

Important scope exceptions are:

The local definitions take precedence over this abbreviated map. Their probabilities cannot be compared as though they described the same prevention population.

Count the right outcome

Let L₀ be each person's physical terminal disappearances minus successful new entries under reference care, normalized to baseline terminal amount. Define LQ in the same way under intervention. A replacement that subsequently disappears enters both counts. Average territories within people, then people as specified locally.

For positive group-mean reference loss, preservation is:

B = (mean L₀ − mean LQ) / mean L₀.

Reference loss of 30% of baseline terminals and intervention loss of 24% gives B=20%. A 20% difference in endpoint density, recovery of 20% of a control–disease gap, or 20% improvement in pain is a different quantity.

A truly nonpositive reference loss does not satisfy this relative loss-prevention proposition. A noisy near-zero denominator leaves it unresolved. Complete physical trajectories, vital status and missingness matter: disappearance from observation cannot be credited as preservation. When unavailable outcomes can move a contrast across its boundary, the result remains unresolved.

Stable subtype identity is required. Loss of CGRP expression can mimic disappearance; general PGP9.5-positive epidermal counts do not exclusively identify peptidergic fibers. Soma survival, axon conduction, pain, myelin, branching and durable connected terminals are distinct outcomes. Existing human skin observations establish relevant anatomy without assigning a cause. Human skin study.

For replacement hypotheses, separately define new entry, subsequent retention and total net preservation. Do not infer birth or functional integration merely from a denser endpoint biopsy. A controlled mediation contrast is also separate from the intervention's total effect; neither is automatically a unique fraction of natural disease.

Construct a compatible chain

Derive the biological requirements before choosing their number. A broad maintenance state, a receptor-specific initiating route and failed replacement need different decompositions.

Define Aᵢ as: at least one same allowed group, compartment and intervention combination satisfies every requirement through row i. Each row retains only combinations that satisfy its predecessors. Thus Aᵢ is a subset of Aᵢ₋₁. With evidence E:

• q₁ = P(A₁ | E).
• qᵢ = P(Aᵢ | Aᵢ₋₁, E) for later rows.
• P(complete hypothesis | E) = q₁ × q₂ × … × qₖ.

This is the chain rule, not an independence assumption. One group cannot supply exposure while a different group supplies benefit or case coverage. If multiple routes could satisfy a requirement, price their union; do not multiply alternatives as though all were necessary.

A 5% structural-effect row followed by a 20% magnitude row is a useful decomposition of the same effect. Conditional on the smaller effect, ask only how likely it reaches 20%. Do not repeat the entire species, dose and endpoint objection as a fresh independent penalty. If the smaller row additionally requires mediation or a particular cell fate, that extra requirement must remain explicit.

Splitting or merging equivalent rows must preserve their joint probability. In hypothesis 08, conditional inputs 65% and 60% can be merged to 39%; the complete probability stays 12.4215%. A longer explanation alone must not reduce belief. Changing the intervention, permitted groups or biological requirements is a different proposition and can change it.

Relate evidence to the exact comparison

For each requirement, identify the closest available evidence and the difference from an ideal resolving comparison. Extract the independent sample count, exposure, comparator, measured endpoint and numerical uncertainty where available.

Human donors are the inference units for human claims. Cells, sections, clones and repeated visits do not increase donor n. Animal-level data need their actual allocation and housing structure considered. A small source-data reanalysis cannot repair unavailable donor or cage identifiers.

Useful distinctions include:

• A randomized structural contrast versus a within-treatment change.
• Natural local exposure versus an imposed culture dose or plasma association.
• Reaction flux versus metabolite pools, expression or static morphology.
• Physiological correction versus severe deletion or supraphysiological stimulation.
• Incident prevention versus established injury, repair or another neuropathy.
• A precise negative effect bound versus missing measurement, failed engagement or wide uncertainty.

Audit both dramatic positives and apparent nulls. Check units, SEM versus SD, analysis populations, denominators, source figures and whether several reports reuse the same experiment. Use the natural testing family where multiplicity matters. A nominal exploratory P-value is not confirmatory evidence.

The DCCT gives an exposure-level anchor: 1,441 randomized people with T1D, with confirmed neuropathy at five years reported in 5% versus 13%, a 64% relative reduction (95% CI 45–76%). It supports preventable glucose-related injury in that population, without allocating the benefit to any one molecular route. DCCT neuropathy report.

Translation anchors constrain confidence without deciding the ideal propositions: the 1,019-person rhNGF trial, 549-person ranirestat trial and 460-person four-year NATHAN 1 trial missed their primary endpoints. None directly measured the specified five-year peptidergic prevention effect with verified selective engagement. rhNGF, ranirestat, NATHAN 1.

State uncertainty without inventing a distribution

Every row has a best estimate and labeled skeptical-to-favorable range, with a biological reason for its width. Multiply the skeptical inputs together and the favorable inputs together to obtain the whole-hypothesis scenario-sensitivity range. They represent coordinated interpretations of the evidence. They are not independently varying error bars, a confidence interval or a 95% credible interval.

These scenarios receive no invented probability weights. Their midpoint is not a posterior mean. The common missing human exposure, selective-intervention, durability and prevalence measurements can move several judgments together.

Study effect intervals and uncertainty about a hypothesis probability answer different questions. A trial's interval around an IENFD difference cannot be pasted around an elicited YES probability. Likewise, Bernoulli variance p(1−p) describes the unresolved binary proposition; it does not measure uncertainty about the estimated p.

A variance over probability judgments requires a justified distribution or weighted set of complete interpretations. If such weights existed, average complete-chain probabilities under each interpretation. Do not average each row using unchanged weights and then multiply: conditioning on earlier claims changes the weights of the interpretations. This collection does not have sufficient evidence to assign those weights.

Display roughly one decimal place for comparison, using two decimals for sensitivity endpoints below 1% so they do not round to zero. Keep fuller products in the notes or numerical support to make arithmetic reproducible. Decimal places show calculation, not scientific resolution. Small differences between hypotheses are not meaningful ranks.

Define ideal YES, NO and unresolved for every row

The ideal-experiment section specifies what would settle each numbered claim and every required subclaim. State shared population, measurements, controls and follow-up once; then give each row's manipulation or observation, direct readout and boundary.

Truth belongs to the exact population quantities. YES requires one compatible combination satisfying every required boundary; NO requires every surviving combination to fail at least one. Finite measurements can remain unresolved. An interval spanning a boundary, an unmeasured allowed branch, failed selectivity or missing physical outcome is not a biological NO. A small, precisely measured effect can establish causation while resolving the materiality claim as NO.

Map existing evidence to each comparison. Do not add an unpriced necessary mechanism in an experimental control, demand an unrelated mediator, or silently require a second population as a condition of biological truth. Independent implementations and restoration comparisons check that the intended manipulation was identified; they are not extra probabilities to multiply.

Compare the collection without inventing independence

The hypotheses overlap: exposure, maintenance and execution may describe different parts of one causal path. Their probabilities do not sum to 100%, and their complements cannot be multiplied to estimate whether at least one is true.

If one exact proposition logically implies another under identical population, intervention, endpoint, timing and coverage definitions, its probability cannot exceed the implied proposition's probability. Shared pathway names do not establish that implication. A broader intervention can remove useful signaling as well as damaging chemistry, so “more complete inhibition” need not produce greater net benefit.

The review improves logical resolution and evidential grounding. It does not establish empirical calibration: there is no resolved reference set demonstrating that claims assigned 70% come true 70% of the time. Nor do cross-sectional human nerves close the gap to molecularly observed incident DPN. The completed judgments expose that uncertainty rather than conceal it.