§ LENS · NET-ZERO · IN BUILD
What a company promised, what it disclosed, and the distance between them.
The third lens. It reads net-zero and interim targets, transition plans, emissions inventories, capital allocation and the governance that sits over them, out of the documents a company publishes, and holds each claim against the passage it came from.
In build
This lens is not available yet and carries no date, because the work that decides the date is methodology work and it has not finished. What follows is where it actually stands, measured. The numbers below are generated from the test corpus and the last evaluation run rather than written by hand.
What it reads
66 fields across 20 indicators: the net-zero claim itself and its scope coverage, near-term targets with their base year, scopes and denominators, scope 1, 2 and 3 inventories with the basis each is measured on, transition-plan content, capital allocation, and the assurance a company obtained over any of it.
Every field is answered from a quoted passage or it is not answered. A figure with no span behind it does not enter the record, and a document that does not state something is recorded as not stating it, which is a different fact from a document nobody read properly.
Where it stands, measured
The test corpus is 33 companies and 2,364 labelled passages, against a frozen registry (9cc9c6dea08ad492). Every change to the lens is scored against it.
THE LABELS ARE MACHINE-READ
All 33 test companies carry the status machine-read, unreviewed, and no passage in the corpus has been reviewed by a person. So every figure below is agreement with a machine reading rather than accuracy in the sense the word usually carries. Expert review is a later stage of this work and the corpus will say so on its face when it has happened.
On the last evaluation run (2026-09-08-b, reader MiniMax-M3):
| agreement, where the passage was retrieved | 53.0% |
| agreement end to end, retrieval included | 32.2% |
| fields the reader declined to answer | 11.0% |
| answers with no supporting passage | 13.9% |
Those figures do not clear the bar this lens set for itself before it measured anything. The release gate has 3 reader-side conditions and 3 of them fail. That is why the lens is in build and not available, and it is the whole reason the bar was registered in advance.
What it publishes
A level for each of the six outcomes, on a 3-point ladder: 1 minimally disclosed (0% and above), 2 partially disclosed (50% and above), 3 substantially disclosed (80% and above). The percentage is how much of what the methodology asks for a company actually published, so a level means the same thing this year as last and does not move when a peer moves.
The boundaries are derived from the distribution across all 33 test companies rather than chosen, and the methodology records the measurement behind them. They are calibrated-provisional and expected to move when expert review lands.
There is no single company score. Six levels are the finding, and one number averaged over them would be a different claim that nobody has made. An outcome nothing could be scored for carries no level at all, which is a statement about our reach and is shown as one.
What it will not do
Estimate an emissions figure a company has not published. Where a disclosure is missing the lens records the absence and says which document it looked in, because a modelled number sitting beside a disclosed one, in the same column, is how a reader stops being able to tell them apart.
It also does not convert a range into a point. A company that says it will cut emissions by 15 to 25 per cent has said something different from one that says 20, and the record keeps the range.
What is known to be wrong with it
8 of the 33 test companies are affected by a defect in our own document extraction: pages that a company did publish were stored blank, so text that exists is missing from the corpus. One report lost its entire independent assurance section that way. Those companies are flagged in the corpus, and an absence recorded against them may be ours rather than theirs.
It is listed here because a reader who finds it themselves is entitled to ask what else was not mentioned.