The assurer put her finger on a single sentence in the sustainability statement. "The company strengthened its supplier code of conduct and achieved high compliance across its strategic suppliers." She did not argue with it. She asked one question: "Show me." Show me the strengthened code, dated. Show me how compliance was measured. Show me what high means and which suppliers count as strategic. The disclosure lead had the numbers in the report cold, every tonne of CO2e traceable to a factor and a source, but this qualitative sentence had nothing behind it but a confident verb. It took two days and three departments to assemble after the fact what should have travelled with the claim from the start. This lesson is about making a sustainability narrative as auditable as a number: every qualitative claim linked to its support, so that when the assurer says "show me," the answer is a click, not a fire drill.

The Asymmetry: Numbers Get Tested, Narrative Gets a Pass

Reporting teams have internalised that every figure must trace to evidence. A tonne of CO2e has an activity figure, an emission factor with provenance, and a source document behind it, and nobody publishes the number without that trail. But the same teams routinely publish qualitative claims, about policies, programmes, progress, governance, and impact, with no equivalent trail at all, as if words carry less weight than numbers. They do not. In an assured CSRD or ISSB statement a qualitative assertion is a disclosure exactly like a quantitative one, and it is testable exactly like one. The assurer who traces your Scope 3 factor will also test your claim that you "engaged stakeholders" or "remediated the incident," and the narrative that cannot answer "show me" fails in the same way the unsupported number fails.

The reason this asymmetry persists is that a number obviously demands a source and a sentence does not feel like it does. "We reduced emissions 4%" visibly needs the 4% backed up; "we are committed to responsible sourcing" feels like a statement of intent rather than a checkable claim. But in a regulated filing it is checkable: committed implies a policy or decision that exists somewhere, responsible sourcing implies a programme with content, and an assurer can ask for both. The discipline this lesson teaches is to treat every qualitative claim as if it were a number, because to the assurer it is.

The Claim-to-Evidence Map

The central artifact is the claim-to-evidence map: a simple, explicit linking of every material qualitative claim in the narrative to the specific piece of evidence that supports it. Not a vague pointer to a folder, but a precise link: this sentence is supported by that document, that board minute, that register entry, that measured result, at that location. The map is the narrative equivalent of the provenance trail you already keep for numbers, and it does the same job: it makes the claim reconstructable by someone who was not there when it was written.

A complete map entry has three parts. First, the claim, the exact assertion as it appears in the narrative, isolated from its surrounding prose so it can be tested on its own. Second, the support, the specific evidence that makes it true: a named, dated source with a location precise enough to find, a policy PDF and its approval date, a measured KPI and its calculation, a board resolution and its minute reference. Third, the link itself, the recorded connection between the two, so that the claim and its support are bound together in the file rather than living in separate worlds that someone has to reunite under deadline. A claim with all three is auditable. A claim missing the support, or missing the link to it, is an assertion floating free, and a floating assertion is what the assurer pulls.

What Counts as Support for a Qualitative Claim

Support for a qualitative claim is more varied than support for a number, which is part of why teams neglect it, so it helps to name the forms. A claim about a policy is supported by the policy document itself, dated and version-identified. A claim about a decision or commitment is supported by the governance record that made it: the board minute, the committee resolution, the approval. A claim about a programme or action is supported by evidence the action happened: the project record, the engagement log, the remediation file. A claim about performance or progress is supported by the measured result and its basis, which ties the qualitative statement back to a quantitative one. And a claim about an outcome or impact is supported by evidence of the effect, attributed accurately to the company's actual contribution. The form varies; the requirement does not. Every claim points to something specific that an assurer can examine.

A number without a source is a misstatement. A qualitative claim without a linked evidence item is the same misstatement wearing words instead of digits. Make the narrative as auditable as the inventory: every claim, its support, and the link between them, in the file before the assurer asks.

How an Assurer Tests a Narrative Assertion

To build the map well, picture exactly how the assurer works on a sentence, because the map is built to answer their procedure. Under a limited assurance engagement, the common level today, the assurer performs procedures sufficient to conclude that nothing has come to their attention suggesting the disclosures are materially misstated, and for a narrative assertion that means selecting claims and asking the company to substantiate them. They read the statement, pick the claims that are material or that catch their eye, and for each one they ask the "show me" question: what is the evidence, where is it, and does it actually say what the narrative says it says. Under reasonable assurance, the higher level the market is moving toward, they select more claims and probe the evidence harder, but the procedure is the same shape.

What the assurer is testing is the fit between claim and evidence, and there are three ways it can fail. The evidence can be absent: the claim has nothing behind it, the worst case. The evidence can be insufficient: something exists but does not actually establish the claim, as when a single pilot is offered to support a claim of broad action. Or the evidence can contradict the wording: the document exists but says something narrower or different from what the narrative asserts, as when a claim of "achieved compliance" rests on a document showing compliance was only assessed. The claim-to-evidence map defends against all three by forcing you, before filing, to find the specific evidence, confirm it is sufficient, and check that it actually says what the claim says. You run the assurer's test on yourself first.

The Linking Discipline in Practice, Including With AI

The map is only as good as the habit that maintains it, and the habit has to fight a strong current, especially when AI drafts the narrative. An AI writes fluent qualitative claims at speed, and it writes them with no attached evidence, because generating a plausible sentence about supplier engagement does not require the engagement to have happened. So an AI-drafted narrative arrives as a stream of confident, unlinked assertions, which is exactly the failure mode the map exists to prevent, produced faster than ever. The linking discipline is what converts that stream back into auditable disclosure.

In practice the discipline has three moves. First, link as you draft, not after. The moment a claim enters the narrative, its support is identified and recorded, because the cost of finding evidence for a claim rises steeply with time and the two-day fire drill is what happens when linking is deferred to the end. Second, refuse the unlinkable claim. If a claim cannot be linked to specific evidence, it is not yet a publishable claim, and the response is to find the evidence, soften the claim to what the evidence supports, or remove it, never to publish it unlinked and hope. Third, make AI serve the map rather than undermine it. Instruct the model to produce, alongside each qualitative claim, the specific evidence it would need to be true, so that every drafted claim arrives with a slot for its support that a human then fills or fails, turning the AI from a generator of floating assertions into a generator of claims that announce what would substantiate them.

Why Linking as You Draft Beats the End-of-Cycle Scramble

The temptation is always to write the narrative first and assemble the evidence later, because drafting is fast and linking is slow, and the deadline rewards whatever is fast now. This is a trap, and naming why protects you from it. At drafting time, the writer knows exactly what they meant and what made the claim true, so the support is one step away. Weeks later, after the writer has moved on and the context has faded, reconstructing what a sentence was based on is archaeology, and sometimes the answer is that nothing solid was based on it at all, which is a discovery you want to make while you can still fix it, not when the assurer makes it for you. Linking as you draft also surfaces the unlinkable claim immediately, when it is cheap to soften or cut, rather than at the end when it is a published exposure. The discipline costs a little at every sentence and saves the catastrophe at the end.

A Worked Example: The Supplier-Compliance Claim, Linked and Unlinked

Return to the sentence the assurer challenged and watch the two versions diverge. The disclosure lead is finalising the governance narrative, and the AI has drafted: "The company strengthened its supplier code of conduct and achieved high compliance across its strategic suppliers."

Before, the unlinked claim. The sentence ships as written, with no map entry. It contains three distinct assertions, each unsupported: that the code was strengthened, that compliance was achieved, and that this holds across strategic suppliers. When the assurer says "show me," the file has nothing assembled. The team scrambles: someone has to find whether the code was actually revised and when, someone has to reconstruct how compliance was measured and what high meant, someone has to define which suppliers were strategic and confirm coverage. Two days and three departments later they assemble a partial answer, and it turns out compliance was assessed, not achieved, for 60% of strategic suppliers, not all. The narrative said more than the evidence supports, and the gap only surfaced under the assurer's question, which is the most expensive moment to find it.

After, the linked claim with a map entry. As the claim is drafted, the lead builds its map entry and the entry forces the truth out. Claim one, the code was strengthened: support is the revised supplier code of conduct, version 3, approved by the procurement committee on a specific date, minute reference recorded. That links cleanly. Claim two, compliance was achieved: the lead goes to find the support and discovers only a compliance assessment covering 60% of strategic suppliers, with results, not an achievement of high compliance across all. The map entry exposes the gap before filing. So the claim is rewritten to what the evidence supports: "The company approved version 3 of its supplier code of conduct in [month], and assessed compliance across suppliers representing 60% of strategic procurement spend, with the results disclosed below." Now claim two links to the assessment record, claim three links to the defined strategic-supplier list and its coverage figure. Every assertion in the sentence points to a specific, dated, locatable piece of evidence. When the assurer says "show me," the lead opens the map and answers in three clicks. Same AI draft, same underlying facts, and the difference is a narrative built to be tested instead of one that hoped not to be.

That is the linking discipline in one comparison. The map did not just prepare the answer to "show me." It changed the claim, by surfacing during drafting that the original wording overstated what the evidence could support, which is the deeper value: a claim built to link is a claim forced to be true.

The Three Fit Failures, and How to Catch Each in Your Own File

The assurer tests the fit between a claim and its evidence, and that fit fails in three distinct ways. Building a good map means running the same three tests on yourself first, so it pays to know each one by its signature and the question that exposes it.

Absent evidence is the simplest and the worst: the claim has nothing behind it at all. The signature is a confident sentence that, when you go looking, points to no document, no record, no measured result. The exposing question is the bluntest one, "where is it," and the only honest answers are to produce the evidence, soften the claim to something the file can support, or remove it. Absent evidence is what the flight-to-Singapore narrative had everywhere, and it is the failure the map most obviously prevents, because an empty support field is impossible to miss when the discipline requires the field to be filled.

Insufficient evidence is subtler and traps careful teams: something exists, so the support field is not empty, but what exists does not actually establish the claim. A single pilot is offered for a claim of broad action; one workshop is offered for a claim of extensive engagement; a plan is offered for a claim of completion. The signature is a support that is real but smaller, narrower, or earlier than the claim it is asked to carry. The exposing question is "does this actually establish that," and the fix is to either find sufficient evidence or shrink the claim to what the existing evidence genuinely proves. Insufficient evidence is dangerous precisely because the presence of some support creates a false sense of safety.

Contradicting evidence is the most insidious: a document exists and is even on point, but it says something narrower or different from what the claim asserts. The classic case is a claim of "achieved compliance" resting on a document showing compliance was only assessed, or a claim of "completed remediation" resting on a remediation plan. The signature is a support that looks right at a glance but, read carefully, does not match the claim's wording. The exposing question is "does it say what the claim says," and the fix is to align the claim's wording to what the evidence actually establishes. This failure is insidious because the team feels covered, a relevant document exists, yet the claim is a misstatement, and only a careful reading of the evidence against the exact wording catches it.

What Makes a Good Link, and Why Precision Is the Whole Point

A map is only as useful as its links are precise, and imprecise links quietly defeat the entire exercise. A link that points to a folder, a system, or "the supplier files generally" is not a link; it is a deferral of the search the map was supposed to eliminate. When the assurer says "show me," a folder-level pointer means someone still has to hunt, and hunting under the assurer's gaze is exactly the fire drill the map exists to prevent. A good link is precise enough that a stranger, following it, lands on the exact item that supports the claim without further searching: not the policy library but version 3 of the supplier code, approved on a named date, at a findable location; not the inbox but the specific supplier response; not the inventory but the exact line that produces the figure.

Precision in the link does a second job beyond speed: it disciplines the claim. To write a precise link you have to identify the exact evidence, and identifying the exact evidence forces you to confront whether it actually exists and actually supports the claim, which is where insufficient and contradicting evidence are caught. A vague link lets a weak claim hide, because "see the supplier files" never has to confront whether the supplier files really establish what the sentence says. A precise link cannot hide anything, because pinning the claim to a specific item is the act that tests it. This is why the linking discipline is not bureaucratic overhead; the precision that makes the link useful to the assurer is the same precision that makes the claim honest in the first place.

Links That Survive Time and Reorganisation

A link that works today and breaks next month is a latent fire drill, so a good map link is also durable. Two things threaten durability. The first is reorganisation: files move, systems migrate, folders are restructured, and a link tied to a transient location dies. The defence is to anchor links to stable references, a document identifier and version rather than a path that will change, so the link survives the next migration. The second is time: the evidence valid at drafting may be superseded before the statement is filed or queried, so a link must capture the dated version it relied on and be re-confirmed at the filing cutoff. A narrative is queried not only at filing but potentially years later in a restatement or a regulator's reopening, and a link that pointed to evidence as it stood, dated and stably referenced, is what lets the claim be substantiated long after the drafter has gone, which is the same reconstructability the inventory's provenance trail provides for numbers.

Key Takeaways

• Reporting teams trace every number to evidence but routinely publish qualitative claims with no equivalent trail; in an assured CSRD or ISSB statement a qualitative assertion is a disclosure and is testable exactly like a figure.
• The claim-to-evidence map links every material qualitative claim to the specific evidence that supports it, making the narrative reconstructable by someone who was not there when it was written, just as the provenance trail does for numbers.
• A complete map entry has three parts: the exact claim isolated from its prose, the specific dated locatable support, and the recorded link binding them, so the claim and its support do not live in separate worlds.
• Support takes varied forms: a policy document for a policy claim, a governance record for a commitment, a project record for an action, a measured result for a progress claim, and attributed evidence for an impact claim, but every claim must point to something specific.
• An assurer tests a narrative assertion by selecting claims and asking show me; the fit can fail three ways: evidence absent, evidence insufficient, or evidence that contradicts the wording, such as achieved compliance resting on a document showing compliance was only assessed.
• The linking discipline has three moves: link as you draft not after, refuse the unlinkable claim by sourcing softening or removing it, and make AI produce the evidence each claim would need rather than a stream of floating assertions.
• Linking as you draft beats the end-of-cycle scramble because at drafting time the support is one step away, while weeks later it is archaeology, and it surfaces the unlinkable claim when it is cheap to fix rather than as a published exposure.
• In the worked example, the map entry revealed that achieved high compliance across all strategic suppliers was really assessed compliance across 60%, forcing the claim to be rewritten to what the evidence supports: a claim built to link is a claim forced to be true.