An analyst needs one answer before the disclosure can move: what is the emissions figure for purchased electricity at the Spanish plant, and is the market-based method even allowed under the company's own policy? The answer lives in three different places. The factor is in the licensed database. The consumption is in the utility records. The method rule is in the GHG accounting policy. Ask an AI that can only see one of them and you will get a confident answer stitched from guesses about the other two. Ask one that can see all three at once, and you get an answer with its whole evidence base attached.

The Answer Lives in Three Places at Once

Reporting is not one library. It is a stack of them, and a single defensible answer almost always draws on more than one floor. Think of the reporting stack as three distinct evidence layers, each authoritative for a different kind of fact, and each useless for the others.

The factor layer is the emission-factor database: the licensed or government dataset that tells you how many kilograms of CO2e a unit of activity produces, with its release year and row. It is authoritative for coefficients and nothing else. It does not know how much electricity you bought or what your policy says.

The source-data layer is your own records: utility bills, fuel invoices, supplier statements, meter readings, the activity data that is the raw material of every calculation. It is authoritative for what actually happened in your operations and value chain. It does not contain factors, and it does not contain rules.

The policy and methodology layer is the basis-of-preparation, the GHG accounting policy, the boundary definitions, the framework requirements, the methodology decisions the company has committed to. It is authoritative for how you are allowed to calculate and disclose. It tells you whether a method is permitted, what the boundary is, and what must be reported. It contains no numbers of its own.

A real question, the electricity one, needs all three. You need the consumption from the source layer, the factor from the factor layer, and the method permission from the policy layer, and the answer is only assurable if it cites all three. This is the heart of the lesson: the unit of a defensible disclosure answer is not a number, it is a number plus the whole evidence base behind it, drawn from across the stack.

Most reporting questions are like this, even when they look simple. "What were our refrigerant emissions?" needs the leakage records from the source layer, the global warming potential from the factor layer, and the boundary rule from the policy layer that says which equipment is in scope. "What is our Scope 3 category 1 figure?" reaches into supplier data, spend or activity factors, and the method the policy commits the company to for that category. The pattern repeats because disclosure is, by design, the disciplined combination of what happened (source data), how it converts to emissions (factors), and how you are permitted to count and present it (policy). A tool that can see only one of those three can never give a whole answer, only a fragment dressed up as one.

Stitched Guesses Versus One Grounded Pass

There are two ways to assemble an answer that spans the stack, and they are not equally safe. The first is what most people do without realising it: they retrieve from one source and let the model fill in the rest from its training. The second is to retrieve from all the relevant sources in a single grounded pass and have the model answer only from what was retrieved. The difference between them is the difference between an answer that fails assurance and one that starts an audit trail.

The stitched answer and why it rots

Picture the stitched approach. The analyst gives the model the utility consumption, then asks for the emissions. The model has the consumption, so that part is grounded, but it does not have the factor database, so it supplies a factor from memory, a plausible-looking number that may be outdated, from the wrong country, or simply invented. It does not have the policy, so it assumes the market-based method is fine without checking. The result is an answer that looks complete and is two-thirds guessed. Worse, the guessed parts are invisible: the figure reads as confidently as the grounded part. Stitching real data to model guesses produces a number that no one can fully trace, and an untraceable number is a misstatement risk no matter how much of it happens to be right.

The single grounded pass

Now picture the grounded pass. The retrieval system searches the factor database, the source records, and the policy library together, pulls the relevant passages from each, and the model answers from those retrieved passages only, citing each. The factor comes with its database and release. The consumption comes with its bill reference. The method permission comes with the policy clause. One question, one retrieval pass, one answer, three citations spanning the stack. Nothing is filled from memory because the model is constrained to answer from what was retrieved, and where a source is missing, it says so instead of inventing. This is multi-source retrieval: pulling factors, source data, and policy in one grounded pass so the answer cites the whole evidence base at once. The retrieval-augmented pattern is the same one a careful analyst already follows by hand, go to each source, take the relevant passage, build the answer from those passages, and never from memory, only automated and applied across the whole stack in a single step.

An answer stitched from one real source and two model guesses is not faster. It is a misstatement wearing the costume of a complete answer. The grounded pass is the only one whose every part you can hand to the assurer.

Why One Pass Beats Three Separate Trips

You might object that you could just retrieve from each source separately and combine the results yourself. You can, and sometimes you must, but the single grounded pass has advantages that matter for assurance, not just convenience.

First, consistency of the question. When one pass retrieves across all three layers against the same question, the factor, the data, and the policy are all answering the same thing: electricity, Spain, this period, market-based. Three separate trips invite drift, where you pull a location-based factor on one trip and ask about market-based method on another, and the mismatch hides in the seams. One pass keeps the whole answer aligned to one question.

Second, completeness is visible. A single pass that retrieves across the stack can report what it could not find. If the policy library has no clause on market-based method for Spain, a grounded pass returns "factor and consumption found, policy clause not found" rather than silently assuming the method is allowed. The gap becomes a flagged, fixable hole instead of an invisible assumption. That visibility is exactly what an assurer rewards and what a stitched answer destroys.

Third, the citation set is the evidence base. The output of one grounded pass is not just an answer, it is an answer plus the exact set of sources it rests on, ready to drop into the working papers. You are not reconstructing later which factor release and which bill and which policy version supported the number. The pass assembled that evidence set at the moment of answering, across the whole stack, which is the same efficiency-meets-assurance win that runs through all of this work.

Grounding Is Still the Non-Negotiable Rule

Multi-source retrieval only works if the model is genuinely constrained to the retrieved evidence. The whole technique collapses if the model is allowed to "help" by supplementing the retrieved passages with its own knowledge. A factor pulled from the database and then "corrected" by the model toward a number it remembers is no longer the database factor: it is a hallucination with a real citation stapled to it, which is more dangerous than an obvious guess because it looks sourced. The instruction must be strict: answer only from the retrieved passages, cite each, and where the evidence base does not contain the answer, say so and stop. A grounded pass that quietly ungrounds itself is the worst of both worlds.

Building the Stack Into a Searchable Knowledge Base

For a single pass to reach across the stack, the three layers have to be made searchable together, and how you assemble them is itself an assurance decision. The factor layer should be the licensed or government dataset, loaded at a known release, so that "the 2025 release, Spain row" means exactly that and cannot drift. The source-data layer should be your actual records, organised so a retrieved figure carries its bill or meter reference back to the original document. The policy layer should be the current, version-controlled basis-of-preparation and accounting policy, not last year's, because a method permitted under the old policy may not be permitted under the new one. The quality of the answer is capped by the quality of what you made searchable: a grounded pass over a stale factor release or an outdated policy will cite the wrong thing with total confidence. Curating the knowledge base is not IT housekeeping, it is the foundation the whole technique stands on, and a wrong version in the base is a wrong citation in every answer that touches it.

This is also where vendor tools enter, and where the obligation does not. Reporting platforms and carbon-accounting tools increasingly offer retrieval across exactly these layers, which is genuinely useful as an orientation to the category. But a platform retrieving across your stack does not transfer the accountability for the answer to the platform. If the factor release it indexed is out of date, or the policy it points to is superseded, the misstatement is still yours. The platform is a faster way to search the evidence base; it is not a reason to trust the answer without confirming the citations, and no tool's name makes a number assurable.

A Worked Example: The Spanish Electricity Figure

Watch the electricity question resolve through a single grounded pass across the stack. The analyst needs the Scope 2 figure for purchased electricity at the Spanish plant, and needs to know the method is permitted.

The question, posed to a multi-source grounded system:

For the Spanish plant, this reporting period: what is the Scope 2 emissions figure for purchased electricity, which method applies under our GHG accounting policy, and what factor and consumption support it? Answer only from the retrieved factor database, source records, and policy documents. Cite each source with its reference. If any of the three is missing, say so and do not fill the gap.

What the grounded pass returns:

1. Method (policy layer). The GHG accounting policy, section on Scope 2, permits the market-based method where contractual instruments exist and requires a location-based figure to be reported in parallel. Citation: GHG Accounting Policy v4, Scope 2 section. So both figures are required, and the answer must produce both.
2. Consumption (source-data layer). 6.2 GWh of purchased electricity at the Spanish plant for the period. Citation: utility invoices for the site, reconciled to meter readings, listed by invoice reference. This is the activity data, traced.
3. Market-based factor (factor layer). The contractual instrument for two of the three supply contracts gives a supplier-specific factor; the residual supply uses the residual-mix factor for Spain. Citations: the supplier disclosures for the instruments, and the residual-mix dataset, 2025 release, Spain row.
4. Location-based factor (factor layer). The Spain grid-average factor, named dataset, 2025 release, Spain row, applied to the full 6.2 GWh for the parallel location-based figure the policy requires.
5. The flagged gap. The retrieval could not find a policy clause confirming the vintage rule for the contractual instruments. Rather than assume the instruments are valid for the period, the pass flags it: "instrument vintage rule not located in policy, confirm before finalising." That flag is a gift, not a failure.

Compare that to the stitched alternative. A single-source approach would have taken the consumption, applied a remembered Spain factor of uncertain vintage and basis, ignored the market-versus-location requirement entirely, and produced one confident number with no policy check and no parallel figure. It would have looked done and been wrong in at least two ways the analyst could not see. The grounded pass produced a longer, messier, and entirely assurable answer: every figure cited, the method confirmed against policy, the parallel figure the policy demands surfaced, and the one genuine gap flagged for a human rather than papered over.

Reading a Multi-Source Answer

When a multi-source answer lands, your job is to check that each citation actually comes from the layer it claims and supports the part of the answer it is attached to. Does the factor citation point to a real database release and the right row? Does the consumption citation point to an actual bill? Does the policy citation quote a clause that genuinely permits the method, for this case, this period? And critically, are there parts of the answer with no citation, because an uncited sentence in a grounded answer is a smuggled guess. The discipline is the same as everywhere in this work: the citation is a lead to confirm, not a fact to trust, and a multi-source answer just gives you three leads to confirm instead of one.

There is a specific cross-layer check worth naming, because it is the one a stitched answer can never fail visibly and a grounded answer makes easy. The three layers must agree with each other for the answer to hold. The factor must match the method the policy permits: a location-based factor is the wrong evidence for a market-based claim, and the answer is wrong even if both the factor and the method are individually real. The consumption period must match the reporting period the question asked. The factor's geography and vintage must match the source data's location and period. These are not checks within one layer, they are checks across layers, and they are only possible because all three citations are present in one answer. A multi-source pass does not just give you the evidence; it lays the three pieces side by side so the seams between them become inspectable. That inspectability is the deepest reason one grounded pass beats stitched-together guesses: the guesses hide the seams, the grounded pass exposes them.

Key Takeaways

• A defensible disclosure answer usually lives in three layers at once: the factor database, your own source records, and the policy and methodology library. Each is authoritative for one kind of fact and useless for the others.
• The unit of an assurable answer is not a number, it is a number plus the whole evidence base behind it, with a citation from each layer the answer touches.
• Stitching one real source to model guesses for the others produces an answer that looks complete and is partly invented, with the guessed parts invisible and untraceable.
• Multi-source retrieval pulls factors, source data, and policy in one grounded pass, so the model answers only from retrieved evidence and cites each layer at once.
• One pass beats separate trips because it keeps the whole answer aligned to one question, makes missing evidence visible as a flagged gap, and assembles the citation set as the evidence base at the moment of answering.
• Grounding is non-negotiable: a model allowed to supplement retrieved passages with its own knowledge produces a hallucination with a real citation stapled to it, which is more dangerous than an obvious guess.
• In the worked example, the grounded pass surfaced the required parallel location-based figure, confirmed the method against policy, and flagged a genuine gap, where a stitched answer would have produced one confident, wrong number.
• Reading a multi-source answer means confirming each citation comes from the layer it claims and supports its part of the answer, and treating any uncited sentence as a smuggled guess.