Methodology
How the figures are made, and what they are not
Written for a buyer, a reviewer or a verifier to read before they rely on anything this site produces. Every number on this page is read from the code that uses it.
Method version
tg-ea731277.cs-e8aabfeb
The version is a fingerprint of the reference data itself — every factor, every process band and every benchmark distribution — so it changes whenever any of them changes, including an edit nobody remembered to announce. A run identifier ends in the same fingerprint, so a result can be matched to the version it used.
Decision support, not certification. Every figure here is a screening estimate built from published data, labelled with how directly it was read from its source. Nothing here states that a product, a building or a business complies with a law, a standard or a customer's requirement. Where a decision turns on a rule — a tax threshold, a legal limit — the rule's value is entered by the user from the primary source, and the decision is made by a named person against evidence.
TrueGauge: a product footprint
- 01
Read the bill of materials
A spreadsheet or a written description is read into rows by deterministic parsing. A line that names a material without a mass becomes a question. It is never given an assumed mass, because there is exactly one way to be missing a quantity and it is to be asked for it.
- 02
Match each line to a published factor
Matching is deterministic and refuses rather than guesses: a description that could be two materials, or two routes of one material, becomes a question with the candidates and the reason for each. How often it is right, and how often it declines, is published separately.
- 03
Price each line from the best-fitting source
8 factor libraries are on the public prototype. It prices on the UK set, and a material that set does not carry is priced from the library whose row best fits where the product is made, labelled on the line, with its share of the total stated. The signed-in engine searches every library and gives each line exactly one factor — the one whose geography and age best fit the client — and says which library answered. Both carbon totals add each line's own published GWP100 figure, and the engine records each source's IPCC vintage. Every other impact category is reported only within its own method and never mixed, because a total built from two methods is not a quantity either method defines.
- 04
Add the conversion, scrap and freight
101 manufacturing processes are modelled as screening bands of energy and scrap, with heat duties computed from first principles where the physics allows. Energy is attributed along the route, including the material lost to scrap downstream. A metered figure supplied by the client outranks the band — and is first checked against the physical floor for the process, so an impossible figure is flagged before it becomes the most trusted number in the model.
- 05
Carry how each figure was obtained
Every line carries a tier: traced (read directly from a source row), derived (calculated from source figures) or modelled (a screening estimate). The tiers set the dispersion used for the interval — a geometric standard deviation of 1.05, 1.15 and 1.35 respectively, roughly ±5%, ±15% and ±35% at one standard deviation.
- 06
Propagate the uncertainty
The 90% interval is propagated across every input, weighted by its share of the total, rather than set by the worst line. Each input's share of the interval is reported, which is also how the next measurement worth making is ranked. In the signed-in engine each factor's range comes from its data-quality pedigree, with geography and age rescored against the client, and the carbon range is widened by how far independent publishers disagree about the same material where more than one publishes it.
- 07
Refuse to print a total that does not describe the product
When less than half of the product's stated mass could be priced, the figure is shown as a partial figure and not as a footprint. A total built from a minority of a product is arithmetically correct and describes something else.
- 08
Recommend with the working shown
A recommendation's carbon figure is computed from the same factors. Its commercial effect is stated as a direction unless the client has supplied prices and volumes, and each one carries the steps, the trial measurements and the point at which the trial should stop.
CarbonSpec: a building's materials
- 01
Estimate at concept stage as a range
At concept stage a class has been chosen and a product has not, so each schedule line is a range read from real declaration distributions — 1,778 construction EPDs across 9 categories, at module A1-A3. The spread inside a class is the decision still open, not noise to be averaged away.
- 02
Report coverage beside the figure
A benchmark is declared per kilogram and a schedule is in square metres. A line that cannot be converted without an assumed density is excluded and counted, not filled in, and the share of the schedule priced is part of the headline.
- 03
Show and check the arithmetic
Every figure has a ledger entry that recomputes it from the steps shown and compares the result with what was reported. A disagreement is presented as a defect in this product. Each run has an identifier whose second half is the data fingerprint, so a figure that no longer reproduces announces itself.
- 04
Gate a comparison before making it
Two declarations are compared only after checking the declared unit, the product category rules, the electricity mix and the service life. Where the difference changes the answer, the direction is named if it is knowable and refused if it is not.
- 05
Give three numbers, not one
Each product carries its declared value, the range its evidence has not ruled out, and a conservative figure for procurement, so a manufacturer who measured is not ranked below an industry average that merely sits in the middle.
- 06
Score more than carbon
Shortlists weigh carbon beside health, fire and durability. Where a product's own rating is not recorded, the class-typical value is shown and labelled as such, so a missing rating is not mistaken for a poor one.
Where a language model is used
No language model is involved anywhere on this site. Reading a bill of materials, matching a line to a factor, every calculation and every recommendation is deterministic code: the same input gives the same output, every time, and the matching accuracy is published with that determinism tested rather than promised. A test fails the build if a model client is imported into any calculation module.
The signed-in TrueGauge application uses one for three bounded tasks:
- Proposing bill-of-materials rows from a written description. It is instructed to extract only materials the text implies and to return an empty quantity where none is stated, never a guess. Every row is shown back for the client to confirm before it enters an assessment.
- Choosing between factors the deterministic matcher has already shortlisted. Its answer is constrained to the shortlist, so it cannot name a factor that does not exist. Below a fixed confidence floor it abstains, and the line goes to a person.
- Writing an interpretation of a finished result, on request. It reads the calculated result; it does not change any figure in it, and nothing it writes is used as an input.
In neither place does a model supply an emission factor, a quantity the client did not state, or any figure in a total.
Not suitable for
An Environmental Product Declaration
An EPD is written to product category rules, registered with a programme operator and verified by a third party. A result here is none of those, however it is formatted.
A verified product carbon footprint
Verification is an independent reviewer's opinion on data this product has not had reviewed. The PACT draft it produces leaves the conformance field empty for that reason.
A comparison published to support a claim that one product is better than another
A comparative assertion disclosed to the public needs an ISO-conformant study with a critical review by a panel. A screening comparison can tell you where to look, not what to announce.
A compliance certificate or verdict
Nothing here returns "compliant". It reports what is recorded, what is missing and what a change would do, and leaves the judgement to the person accountable for it.
Tax or legal advice
Tax thresholds and rates are entered by the user from the Act, not held here, because they are amended and a stale rate stated confidently is worse than none.
Embedded-emissions reporting under a carbon border mechanism
Those regimes ask for installation-level data from the producer. A library average for the same material is a different quantity, and submitting it as the producer's figure would misstate what it is.
A marketing claim as it stands
A figure published without its scope, boundary, date and basis beside it says more than the figure does. Any public use needs those qualifications carried with it.
Known limitations
- Most figures are global warming potential only. Other indicators appear where the source publishes them and are never estimated from carbon.
- TrueGauge is cradle-to-gate. The use phase is outside the boundary; end of life (by route) and delivery to the customer (per 100 km, by mode) are reported beside the total rather than inside it.
- A line priced by what it cost rather than what it weighs uses US supply-chain factors per 2022 dollar: a proxy outside the US, restated in 2022 prices only where a producer price index for the supplying sector is held, and nominal otherwise.
- CarbonSpec's benchmark distributions are German declarations. They say whether a product is unusual for its class; they do not describe the UK market average.
- Process energy and scrap are screening bands at the modelled tier until the client supplies a metered figure.
- Prices are labelled bands until the client supplies their own. A saving stated in pounds is only as good as the price behind it.
Beyond these, 66 capabilities that competitors offer or buyers have asked for are not built. Each is listed with the reason and the condition that would change it on the capability register. Mistakes found after publication are recorded on the errata page, and how much the published factor libraries disagree with each other is measured in the state of the data.