The sustainability statement is laid out, the narrative reads beautifully, the tables are full, and the file is one click from going to the printer and the regulator. Before that click, one person runs a pass that asks a single, ruthless question of every line: can the file support this claim? Not "does it sound right." Can we, today, point to the evidence that backs it. That pass is the last gate, and it is the difference between a disclosure that survives assurance and one that becomes a restatement.

The Last Gate, and Why It Exists

Everything upstream of this moment can be done well and you still are not safe. You can have grounded your model, labelled your data, traced your factors, and drafted carefully, and a single unsupported claim can still be sitting in the final file because someone reworded a sentence, pulled in last year's number, or let an AI polish pass reintroduce a figure that had been cut. Disclosure runs on an unforgiving asymmetry: the report is easy to generate and hard to defend. The verification pass exists because the cost of one unsupported claim reaching publication, a restatement, a greenwashing headline, a regulatory file, is wildly out of proportion to the few hours it takes to check.

The governing principle is one sentence, and it is the whole lesson: no published claim the file cannot support. Not "no claim we think is true." Not "no claim the AI was confident about." A claim ships only if, standing in front of an assurer or a regulator with the evidence file open, you could put your finger on what backs it. If you cannot, the claim does not ship, regardless of how plausible it is or how good it makes the company look.

The verification pass does not ask whether a claim is true. It asks whether the file can prove it. An unprovable truth and a confident fabrication look identical on the page, and neither one ships.

Why AI Makes This Gate Non-Optional

This gate always mattered, but AI-drafted disclosure makes it non-negotiable. A human writer who is unsure of a number tends to hesitate, hedge, or leave a note. A model does not hesitate. It produces fluent, confident, finished-looking prose whether or not the claim underneath is supported, so the usual visual tells of an uncertain figure are gone. Worse, AI is often used in the final stretch, a polish pass, a reformat, a "tighten this section," and those passes silently reintroduce or alter content. A figure you deleted can reappear because the model considered it a natural fit. The last gate is where you catch what the convenience of AI quietly let back in.

There is a specific way teams lull themselves here, and it is worth naming so you can resist it. Because the upstream work was done carefully, the final file feels trustworthy, and a polished AI draft amplifies that feeling: it reads like an already-assured report. But polish is a property of the prose, not of the evidence behind it, and the two have no necessary connection. A sentence can be impeccably written and completely unsupported. So the confidence the file inspires is not a reason to relax the gate; it is a reason to run it, because the same fluency that makes the draft feel finished is what disguises the claim that cannot be backed. The verification pass exists precisely to refuse to be reassured by how the file looks and to insist instead on what the file can prove.

It also matters that this is genuinely the last gate, not the only one. Upstream you have labelled primary and secondary data, traced factors, grounded the model, and checked extractions. None of that is wasted; it makes the final pass shorter and cleaner. But every one of those controls operated on an earlier version of the file, and the file has been edited, merged, reformatted, and polished since. The job of the last gate is to verify the artefact that will actually be published, in the exact state it will be published, because that is the only version a regulator can hold you to. A control that ran on a draft three versions ago has not verified the bytes going to the printer.

The Four Things You Verify

A disclosure is not one kind of claim, so the pass is not one kind of check. Every published file decomposes into four classes, and each has its own evidence standard. Run them as four explicit sweeps, not one vague read.

Every Number to Its Source

Take every figure in the file, in the narrative and in the tables, and trace it to the source it came from: the inventory line, the source data, the dated record. The number in the disclosure must equal the number in the source, to the figure. This is where you catch the transposed digit, the stale prior-year value, the rounding that drifted, and the AI-inserted number that has no source at all. A number that cannot be traced is not "probably from somewhere." It is unsupported, and it comes out or gets corrected.

Every Factor to Its Provenance

Emission factors are a special class of number because they are the silent multiplier underneath your emissions totals, and a wrong factor corrupts a total without changing anything that looks wrong. Every factor used in the file must trace to a named, dated, authoritative database, the specific source, the specific version, the specific value. The failure mode here is the hallucinated factor: a plausible-looking emissions coefficient the model produced that exists in no database. You verify provenance, not plausibility, because a fabricated factor can look exactly as reasonable as a real one.

Every Qualitative Claim to Its Evidence

The prose makes assertions: we have a policy, our board oversees this, we reduced that impact, we are committed to a goal. Each qualitative claim must link to the document or record that supports it, the policy itself, the charter, the minutes, the assessment, the adopted target. The claim must match what the evidence actually says, not what a well-meaning model assumes a responsible company would say. This sweep catches the invented target, the overstated commitment, the policy described as in force when it is still a draft, and the softened negative impact that reads more reassuringly than the underlying record warrants.

Every Estimate Labeled as an Estimate

Some of your numbers are not measured; they are estimated, modelled, or proxied, and that is legitimate. What is not legitimate is an estimate that looks identical to a measured figure. The fourth sweep confirms that every estimated number is labelled as an estimate, with its method and, where relevant, its uncertainty, so that a reader can tell measured from modelled. An estimate disclosed as an estimate is defensible. An estimate dressed up as measured activity data is a misstatement, and it is the easiest one to commit by accident, because the model that produced your table did not carry the label.

This sweep is easy to skip precisely because nothing about an unlabelled estimate looks wrong. A spend-based Scope 3 figure and a metered Scope 1 figure are both just numbers in a column, and a clean AI table presents them identically. The discipline is to refuse to let any number into the file without a determination of whether it is measured or estimated, treating the absence of a label not as "probably measured" but as "not yet verified." In a Scope 3 inventory, where so much of the footprint is unavoidably estimated, this is not a corner case; it is most of the file. The reader, and the assurer, needs to be able to weigh the inventory's reliability, and they cannot do that if the measured and the modelled wear the same uniform.

Verify From the Claim, Not From the Source

One method point ties the four sweeps together and decides whether they actually work. Verify outward from each claim, not inward from each source. The lazy version of verification reads the evidence file and confirms the disclosure is consistent with it, and that version misses every fabricated claim, because a fabrication is consistent with a source that simply never mentioned it. The reliable version takes each claim in the published file and demands the specific evidence that produced it. The invented target, the unsourced figure, the hallucinated factor: all three survive a consistency read and all three die when you point at them and ask, what produced this, and find nothing. Make every claim earn its place by naming its support; do not let the mere absence of contradiction pass for the presence of evidence.

The Checklist That Stands Between Draft and Disclosure

The pass becomes reliable when it becomes a checklist anyone on the team can run the same way, not a heroic act by the one careful person. A workable last-gate checklist reads roughly like this, and it is deliberately concrete.

• Every figure in the file traces to a source, and the disclosed value equals the source value to the figure.
• Every emission factor traces to a named, dated, versioned, authoritative database.
• Every qualitative claim links to the policy, charter, minutes, assessment, or adopted target that supports it, and matches what that document says.
• Every target or commitment traces to a formally adopted target document; no target exists that the company did not set.
• Every estimated number is labelled as an estimate with its method, and no estimate is presented as measured data.
• Every negative impact is described at the intensity of the underlying assessment, not softened.
• The narrative figures reconcile to the tagged or tabulated figures; the prose and the numbers tell the same story.
• Nothing reintroduced by a late AI polish or reformat pass has slipped back in unchecked.

A claim that fails any line is not a debate. It is fixed, traced, labelled, or removed before the file moves. The checklist's power is that it converts "be careful" into a finite, repeatable set of yes-or-no questions, and yes-or-no questions are auditable in a way that diligence is not. It also makes the pass transferable: a single careful person is a single point of failure, but a written checklist with defined evidence standards can be run identically by anyone on the team, and the log it produces shows exactly what was checked. That log is not busywork. It is the byproduct that doubles as assurance-readiness, because the same record that proves you ran the gate is the record an assurer uses to reconstruct your claims.

Worked Example: One Page, Through the Gate

A disclosure lead has an AI-assisted climate section ready to ship. It reads cleanly. Watch it go through the four sweeps.

The draft: "We reduced absolute Scope 1 and 2 emissions by 11% in 2025, using a grid emission factor of 0.233 kg CO2 per kWh. Scope 3 emissions were 412,000 tonnes CO2e. We are committed to net zero by 2045, overseen by our Sustainability Board. Our supplier engagement programme has substantially mitigated deforestation risk in our palm-oil supply chain."

Sweep one, numbers. The 11% reduction: the inventory shows 9%. Corrected. The 412,000 tonnes Scope 3: traces to the inventory, but it is a spend-based estimate, flagging it for sweep four. Noted.

Sweep two, factors. The 0.233 grid factor: which database, which year? The analyst checks and finds the model used a plausible but unsourced number; the correct published factor for the relevant grid and year is 0.241. Corrected and provenanced.

Sweep three, qualitative claims. The 2045 net-zero commitment: no adopted target document exists. Removed, replaced with a disclosed statement that no target is yet set. The "Sustainability Board": no such body exists. Corrected to the actual oversight body. The deforestation claim: the assessment says the risk is identified and partially mitigated, not "substantially mitigated." Rewritten to match the assessment, un-softened.

Sweep four, estimates. The 412,000 Scope 3 figure, flagged earlier, is now labelled as a spend-based estimate with its method, so it no longer masquerades as measured data. Labelled.

The shipped version: 9% reduction (inventory), 0.241 grid factor (named database and year), Scope 3 of 412,000 tonnes labelled as a spend-based estimate, no net-zero target with the absence disclosed, the correct oversight body named, and a deforestation risk described honestly as partially mitigated. Five sentences in, the draft had at least six problems, every one of which would have been an assurance finding and two of which were greenwashing exposures. After the gate, every claim on the page can be defended with the file open. The page is less impressive and entirely shippable, which is the trade the whole discipline is built around.

Who Runs It, and When

The last gate works only if it is the last thing, run after all editing, including any AI polish, is finished, on the exact bytes that will be published. Run it earlier and a later change escapes it. It is run by someone empowered to hold the file, who treats their sign-off as a statement that every claim is supportable, because that signature is what an assurer relies on and what a regulator will trace back. The pass does not make the discloser less accountable; it is how the discloser earns the right to be accountable, by knowing the file can carry every word.

The "who" matters as much as the "when." The runner needs the authority to stop publication, because a gate whose failures can be overruled by the deadline is not a gate at all. If the only response available when a claim fails is to note it and ship anyway, the pass becomes theatre. So the role carries a real power: a failed claim halts the file until it is fixed, traced, labelled, or removed. That power is also what makes the sign-off meaningful. When the signer says the file is ready, they are not expressing optimism; they are stating that they ran a finite set of yes-or-no checks and every one came back supportable, and they are putting their name to that as the person an assurer and a regulator will trace the claim back to.

What the Gate Does and Does Not Guarantee

Be precise about the gate's promise so you do not over-trust it either. The last gate guarantees that every published claim traces to evidence in the file and is correctly labelled. It does not guarantee that the underlying source data is itself free of error, which is why upstream data-quality controls still matter; a perfectly traced number is only as good as the source it traces to. What the gate buys you is a different and specific thing: it ensures the disclosure never says more than the file can support. That is the line between an honest disclosure that might still contain an upstream measurement error, which is a normal and explainable event, and a disclosure that asserts things the company cannot back at all, which is the misstatement that becomes a restatement and a headline. The gate cannot make your data perfect. It can make your disclosure honest about what your data shows, and that is exactly the property assurance is built to confirm.

Key Takeaways

• The last gate enforces one principle: no published claim the file cannot support, where support means you could point to the evidence with an assurer and a regulator watching.
• The pass asks whether the file can prove a claim, not whether the claim is true; an unprovable truth and a confident fabrication look identical on the page, and neither ships.
• AI makes this gate non-optional, because models produce confident, finished-looking prose whether or not the claim is supported, and late polish passes silently reintroduce or alter content.
• Verify four classes separately: every number to its source, every factor to its provenance, every qualitative claim to its evidence, and every estimate labelled as an estimate.
• A wrong emission factor corrupts a total without looking wrong, so factors are verified for provenance to a named, dated, versioned database, not for plausibility.
• An estimate dressed up as measured data is a misstatement and the easiest to commit by accident, because the AI table did not carry the label.
• Convert the pass into a concrete, yes-or-no checklist anyone can run identically; a claim that fails any line is fixed, traced, labelled, or removed before the file moves.
• Run it last, on the exact bytes to be published, by someone whose sign-off means every claim is supportable, because that signature is what assurance and regulation trace back to.