WoodIntel™ Scientific Methodology — Version 1.0

0

Executive summary

WoodIntel is an independent open platform that predicts, grades and publishes what is known about the durability of wood in outdoor use. This document is the full method behind it. The summary below is written for readers who need to judge the platform's credibility without reading the seventeen sections that follow.

• What it does: screens modelled service life for above-ground timber in EN 335 use classes 2 and 3, scores species and treatment states, indexes the published literature and states where the evidence agrees, disagrees or is missing.
• How numbers are produced: deterministic, published equations with seeded Monte Carlo uncertainty. Language models never generate a published value — they classify study stance and phrase explanations of numbers computed elsewhere.
• What every prediction carries: a P10/P50/P90 interval, the climate zone and parameters used, the calibration applied, and the version identifiers of every component involved.
• How claims are graded: each study receives an evidence score from study design, sample size, replication and independence; claims inherit a consensus position that is recomputed when the corpus changes.
• Independence: initiated and funded by SiOO:X Wood Protection, with no funder influence on results and no model input, score or ranking that references any commercial product, including the funder's.
• Reproducibility: every model, parameter table and grading rule is published, and every figure on the platform can be recomputed from this document without access to the source code.

One caveat governs the whole platform: modelled service life is a screening estimate, not a measured lifetime. Detailing, ventilation, end-grain protection and maintenance routinely matter more than species choice.

1

Scope and purpose

WoodIntel is an independent open platform for evidence-based wood specification. It publishes three distinct kinds of statement, and this document exists so that a reader can tell them apart and check each one: values taken verbatim from standards and datasets, values computed by published WoodIntel models, and assessments of what the scientific literature currently supports.

This report documents every model, dataset, scoring rule and interface in the platform at the release stated on the title page. It is written to be citable. A reviewer who disagrees with a result should be able to locate the exact equation, parameter table or grading rule that produced it, and to recompute the number without access to the source code.

• In scope: above-ground timber in use classes 2 and 3 as defined by EN 335, exterior cladding, decking and facade detailing, and the published literature that describes their performance.
• Out of scope: structural load-bearing design, ground-contact use class 4 and marine use class 5, fire engineering beyond published classification values, and any assessment of a named commercial product.
• Not provided: warranties, performance guarantees, product recommendations, and any statement that a specific supplier's material is preferable to another's.

Modelled service life is a screening estimate, not a measured lifetime. Every predicted value in the platform is labelled as modelled and carries an uncertainty interval. Detailing, end-grain protection, ventilation and maintenance routinely change real outcomes by more than species choice does.

2

Governance, funding and independence

WoodIntel was initiated and is funded by SiOO:X Wood Protection. The funder has no influence on results, model parameters, evidence grading or editorial content, and no ability to suppress a published finding. This is stated on every page of the platform rather than in a footnote, because a reader is entitled to weigh it.

Two structural safeguards support that claim. First, the platform is brand-neutral by construction: no model input, score or ranking references a commercial product or supplier, including the funder's. Species and treatment classes are modelled as generic material states defined by EN standards. Second, every model is published in full and every prediction is reproducible, so a third party can detect a favourable parameter choice by inspection rather than by trust.

WoodIntel is not a research institute, a certification body or an accredited laboratory, and does not describe itself as one. It is an open platform that indexes, models and publishes.

3

System architecture

The platform is a single application with four layers. The data layer holds standards-derived parameters, species properties and climate references. The model layer computes service life, uncertainty, scores, carbon and forecasts deterministically. The evidence layer indexes literature and assesses claims. The interface layer renders the web platform and serves the open read API.

  DATA LAYER            MODEL LAYER              EVIDENCE LAYER
  ------------------    ---------------------    ----------------------
  EN / ISO parameter    WI-SL service life  ->   Literature retrieval
  tables                     |                        |
  Species properties    Monte Carlo P10/50/90    Study grading (WI-EV)
  Climate stations           |                        |
  Validation set   ->   WI-CAL zone calibration  Stance classification
                             |                        |
                        Scores, carbon,          Claim consensus +
                        forecast, digital twin   contradiction register
                             |                        |
                             +-----------+------------+
                                         v
                              INTERFACE LAYER
                    Web platform | Calculation records | Open read API
                    every output carries: value + interval +
                    component IDs + versions + sources

The separation matters for auditability: language models never produce a number that the platform publishes as a value. They are used only for retrieval framing, stance classification of individual papers, and natural-language explanation of results computed elsewhere.

4

Data sources

Every quantitative value surfaced by the platform carries a provenance record: the value, the unit, a confidence class, a one-sentence description of how it was derived, and the source references it resolves to. The confidence classes are measured, published, derived and model estimate, in decreasing order of directness.

Standards texts are not redistributed. The platform reproduces only the classification values needed to compute a result, together with the reference by which a reader can obtain the standard. Climate values are reference values for a Köppen zone, not a site-specific weather file; site-specific work requires a local weather file and is out of scope.

5

Service-life model (WI-SL v1.2.0)

The service-life screening method is the platform's core physical model. It was first released on 2026-08-01 and follows the ISO 15686-8 factor structure: a reference service life derived from the EN 350 durability class, modified by climate load, orientation, macro-exposure and surface treatment, then corrected by an empirical per-zone calibration.

• L_ref — reference service life in years, indexed by EN 350 durability class 1–5 for above-ground use.
• f_climate — composite of UV dose, wetting frequency, driving-rain index, freeze–thaw cycling and biological activity, computed from the zone climate reference.
• f_orientation — north, south, east/west or horizontal, applied separately to the UV and wetting terms because they move in opposite directions.
• f_exposure — sheltered 0.72, moderate 1.00, fully exposed 1.30.
• f_treatment — applied only for a maintained coating or a modification process, expressed as a generic material state and never as a named product.
• c_zone — empirical bias correction fitted per Köppen zone against the validation set.

Species resistance is expressed as a dimensionless 0–100 profile across six mechanisms — UV, moisture, salt, freeze–thaw, biological attack and dimensional movement. This profile is an ordinal scaling of the EN 350 class, extractive chemistry, cell-wall modification state and shrinkage anisotropy. It is not a measured quantity and is labelled as a model estimate everywhere it appears.

The model computes a load and a resistance per mechanism and reports the margin between them. A negative margin is surfaced as a detailing flag rather than being absorbed silently into the service-life number, because the correct response to a moisture-load exceedance is a detailing change, not a species change.

6

Uncertainty propagation

No prediction is published as a point estimate. Every input carries a distribution, and the model propagates them by Monte Carlo simulation with a seeded generator, so the same brief always returns the same interval and any third party can reproduce it.

• Outputs are reported as P10, P50 and P90 for both service life and maintenance interval.
• The seed is derived from the method version, the brief and the calibration value, making each run reproducible and each version change visible.
• Variance is decomposed per input group, so the interface can state which unknown dominates the uncertainty for that specific brief.
• A zone uncertainty multiplier widens the interval where the calibration is supported by fewer benchmark cases.

Specification guidance in the platform is written against the P10, not the median. The interval is empirical model error measured against the validation benchmark, not a manufacturer expectation.

7

Calibration and validation (WI-VAL v1.1.0)

The uncalibrated factor model is systematically biased in some climate zones. The calibration layer corrects that bias with a per-zone multiplier fitted against a published benchmark set of field trials and in-service inspections. The benchmark set, the fitted multipliers and the residual error are all published.

Validation reports mean absolute error, bias and interval coverage — the share of benchmark cases whose observed service life falls inside the predicted P10–P90 band. Coverage is the metric that matters for specification: a model with a small mean error but poor coverage is understating its own uncertainty.

Benchmark cases are versioned. Adding, correcting or removing a case bumps the benchmark version and triggers a refit, and both events appear in the public release history.

8

Evidence model (WI-EV v1.0.0)

The Evidence Engine turns a set of graded publications about one claim into four auditable numbers: an Evidence Score from 0 to 100, a confidence band, a scientific consensus percentage, and a set of detected contradictions. Every one of those numbers is produced by a transparent rule, never by a model judgement, and can be recomputed by hand from the published study list.

Each study contributes a weight equal to its grade weight multiplied by 1.0 if peer-reviewed and 0.6 otherwise. Grade weights are A = 1.00, B = 0.75, C = 0.45, D = 0.20.

The engine also reports its own limiting factor — the component with the lowest share of its maximum — so a reader immediately knows whether a low score reflects thin literature, weak study designs, genuine disagreement or ageing sources.

9

Confidence model

Confidence is a deliberately coarse, three-level band derived jointly from the Evidence Score and the amount of directional evidence behind it. Both conditions must hold; a high score computed over three papers is not high confidence.

A separate provenance confidence class applies to individual data values rather than claims: measured, published, derived or model estimate. The two systems are never mixed in the interface, because a measured density and a high-confidence claim are different kinds of statement.

10

Consensus and contradiction engine

Consensus is reported as a weighted share, not a vote count, so a single high-grade randomised field trial is not outweighed by several low-grade opinion pieces. The platform never rounds disagreement away: where the literature is split, the split is the finding.

The contradiction engine pairs the strongest supporting studies against the strongest contradicting ones and explains the difference from metadata alone — publication era, venue type and evidence grade. It offers no speculative reconciliation, because a plausible-sounding explanation generated by a model would be indistinguishable from an established one.

The research timeline shows how the balance of evidence for a claim has moved over publication years, making it visible when a consensus is recent, ageing or reversing.

11

Quality score — study grading

Every indexed publication is graded A to D before it can influence any score. Grading is metadata-driven and deterministic: it uses venue type, peer-review status, registry coverage, DOI resolvability, open-access status, citation count and publication year. It does not attempt to judge the scientific merit of the argument.

The species-level WoodIntel Score is a separate instrument and must not be confused with study grading. It is the unweighted arithmetic mean of six sub-scores — durability, carbon, maintenance, fire, circularity and sourcing — each normalised to 0–100 with its formula published. The mean is deliberately unweighted, because any weighting encodes a project priority that belongs to the specifier, not to the platform.

12

AI models and their boundaries

Language models are used in four bounded roles. In none of them does a model output a published number, and in all of them the output is attributable to a source the reader can open.

Every model-mediated surface is retrieval-grounded, cites its sources inline, and is labelled as model-generated. A nightly automated sweep re-runs retrieval and stance classification for tracked claims; each resulting change is written to the public revision history with the reason it changed.

Model outputs are assistive, not authoritative. Where a model-mediated statement and a published standard disagree, the standard governs.

13

Derived instruments

Each derived instrument declares the base method it depends on. When the base service-life method changes version, every dependent instrument is re-derived and the dependency is recorded in the release history.

14

Open API (v1.0.0)

Everything the platform displays about species, scores, climate zones, modelled predictions, evidence and research gaps is available as JSON. The API is free, unauthenticated, CORS-open, read-only, serves no personal data, and is licensed CC BY 4.0.

Every response is wrapped in an envelope carrying the API version, licence, attribution string, generation timestamp and the caveats that must travel with the numbers. Republished service lives must be labelled as modelled estimates. Scores are comparative and are not recommendations.

15

Reproducibility and versioning

Every calculation performed in the platform can be exported as a machine-readable record containing the inputs, intermediates, equations, parameters, seed and method versions used. A third party can recompute the result from that record alone.

Each component carries an independent semantic version. A patch corrects an error without changing published values; a minor version adds capability or data without invalidating existing results; a major version changes model behaviour, at which point previously published numbers must be regarded as superseded rather than corrected.

All changes are published in the public release history, and evidence-level changes additionally appear in the evidence changelog and its RSS feed, revision by revision, with the reason each score moved.

16

Limitations

• Climate loads are zone reference values, not site weather files. Local microclimate, shelter and thermal mass are not represented.
• Detailing dominates outcomes. The model can flag a detailing risk but cannot evaluate a junction drawing.
• The validation benchmark is small relative to the parameter space, so calibration confidence is uneven between zones and this is reflected in the interval, not hidden.
• Literature coverage is limited to what the connected registries index, which under-represents non-English and pre-digital field-trial reporting.
• Stance classification is model-mediated. It is published per study with its reason so that it can be challenged, but it is not error-free.
• Field observations are self-reported and moderated, not instrumented measurements, and are aggregated rather than treated as validation data.
• The platform models generic material states. It cannot predict the performance of a specific commercial formulation, and does not attempt to.

17

Companion reference documents

This report is the anchor document of a numbered WoodIntel reference series. Each companion document has its own identifier, version and release history, and is cited independently. The series exists so that a reader can go from the whole system to one component without reading the whole report, and so that a component can be revised without reissuing the anchor.

Where a companion document and this report disagree, the companion document governs for its own component, and the discrepancy is recorded as a correction in the release history of both.

18

Citation and licence

Cite this document as: WoodIntel. (2026). WoodIntel™ Scientific Methodology: Reference architecture for AI-driven, evidence-based wood specification. woodintel.io/methodology

This document, the models it describes and the data the platform publishes are licensed CC BY 4.0. Reuse with attribution to WoodIntel, woodintel.io. Standards texts referenced here remain the property of their publishers and are not redistributed.

Corrections are welcome and are handled in public: a substantiated error becomes a versioned change in the release history, with the previous value preserved rather than overwritten.

A

Equations in execution order

Every equation the engine evaluates, in the order it runs. Each is stated with its terms and the standard or dataset it rests on. Together with Appendix B and C, this is sufficient to reproduce any published prediction by hand.

B

Species resistance profiles

Dimensionless 0–100 resistance per degradation mechanism. These are expert-assigned ordinal scores derived from the EN 350 durability class, extractive chemistry, cell-wall modification state and shrinkage anisotropy. They are model estimates, not laboratory measurements.

C

Modifier tables and scalar constants

D

Declared input uncertainty and sampling

Every uncertain input carries a declared 1σ. These are modelling assumptions, stated openly so they can be challenged and replaced by measured field variance as the evidence base grows.

E

Calibration factors (WI-CAL v1.1.0)

Chosen so that leave-one-out P10–P90 coverage on WI-VAL v1.0.0 reaches 83 % against an 80 % target. The factor is large: it shows the declared input σ alone understated real scatter by roughly 4×, and the extra spread is honest model error, not measurement noise.

F

Validation benchmark set (WI-VAL v1.1.0)

The 24 benchmark observations the model is measured and calibrated against. Each is a published field trial, in-service survey, standardised test or service-life database entry.

G

Climate reference stations

Reference locations used to build zone climate loads. Coordinates are the reanalysis grid anchor; the distance to open salt water drives the chloride term.

H

WoodIntel Score — sub-scores and results

I

Evidence grades

J

Source registry

K

Stated limits of the service-life method

• Service life is empirically calibrated (E14) against a 24-case benchmark. The correction removes the average bias; it does not make the model a validated damage model, and the zone factors for tropical, arid and polar climates rest on two cases or fewer.
• The calibration is fitted on the same benchmark used to report accuracy. Leave-one-out metrics are published on /validation and are the correct figures to judge the model by.
• Resistance profiles are expert-assigned ordinal scores, not laboratory measurements. Only the EN 350 durability class behind them is a standardised value.
• The confidence band (E9) is a rule-based heuristic, not a statistical interval. The P10/P50/P90 figures come from Monte Carlo propagation (E11–E13) of declared input uncertainty — they are as good as those declared σ values, which are modelling assumptions until replaced by measured field variance.
• Input uncertainties are treated as independent. Real correlation between rain, humidity and biological load will make the true spread wider than modelled at the tails.
• Detailing quality — ventilated cavity, end-grain sealing, fall, ground clearance — dominates real outcomes and is only partially represented.
• Treatment performance assumes a correctly applied system maintained to its stated interval. Product-specific data is not used.
• Climate inputs are zone-level, not site-level. Local topography, shading and coastal microclimate are not resolved.