You can build a number that is physically sound, properly modelled, and honestly sourced, and still walk a client straight into the wrong decision with it.

That sentence is the reason this last gate exists. Part 1 asked whether the model makes physical sense. Part 2 asked whether the modelling holds at the speed AI now works. Part 3 asked whether the data underneath is sound. Three gates, all facing the same way; inward, at the work, checking it is built correctly.

This last gate is the only one that turns around and faces the other way, toward the person who paid for the thing and now has to act on it: can you actually make a decision with this? A model can pass all three earlier gates and still fail this one, and that failure is the most dangerous kind, because it looks exactly like success.

The first three gates are a climb. You build up the stack; physical, then modelling, then data, each layer resting on the one beneath. But the outcome gate is not the top of a ladder you climb and then step off. It is where you check the result against the world, and then go back down the stack to understand why the result is what it is, before you decide whether to release it. The series has spent three parts climbing. This part is the descent. That round trip is the thing that turns a correct number into a number you can stand behind.

This is the last piece in the series, and the gate that closes it.

There is a quiet assumption in how LCA results are received: that a confident-looking number is a trustworthy one. Clean formatting, decimal places, a fluent methodology section; these read as authority. AI sharpens the effect. A result produced fast and presented cleanly looks settled, and the polish of the output rises faster than the trustworthiness of the conclusion.

The outcome gate exists to puncture that. A number can be internally immaculate, through all three earlier gates, and still be the wrong number to act on, because nothing inside the model tells you whether it corresponds to reality, or whether the conclusion would survive a defensible change of approach, or whether it means the same thing as the number it is being compared against.

The gate does two things. First, it checks the number against the world. Second, it uses what that check reveals to send the practitioner back through the earlier gates, not to change the answer, but to understand it well enough to defend it.

This is the oldest check in engineering and the easiest to lose when a machine produces a confident figure quickly. Step back from the model and hold the result against reality.

Does it sit sensibly against comparable products? Against physical limits — no process emits less than its chemistry allows, no recycled content exceeds what mass balance permits? Against what the rest of the market reports? Against the per-unit sanity check of someone who knows the sector; does the figure per kilogram, per functional unit, feel right?

A hidden error does not announce itself inside the model. A units slip, a factor at the wrong order of magnitude, a migrated decimal; these reconcile perfectly internally and only show up when the result meets the world and someone says that cannot be right, nothing in this category is that low. That instinct is a check, and it is the most valuable thing the practitioner brings to this gate.

But here is the part that matters, and it is where most accounts of plausibility get it wrong. When the number comes out higher, or lower, than the world expects, the answer is not to move it until it matches. That is the amateur reflex, adjust until you sit near the benchmark and hope nobody asks why. The expert response is the opposite. You go back and find out why you are standing where you are standing.

This is the loop, and it is the heart of how the gate actually works in practice.

The plausibility check is not a pass or fail. It is a prompt to go back down the stack you climbed. The number is higher than the benchmark; why? Is it a data problem (back to Part 3), a modelling choice (Part 2), or something in how the system was scoped and the goal defined in the first place (Part 1)? You descend through your own work with the outcome in hand, re-interrogating each gate now that you can see where the result landed.

One of two things comes back. Either you find an error — and you have just caught the kind of error the first three gates structurally cannot, because they each assume the layer beneath them was right. Or you find that your number is legitimately different from the world because your scoping was right and theirs was different — you included a life-cycle stage the market omits, you drew the boundary more completely, you allocated on a basis you can defend.

That second outcome is the powerful one, and it inverts the usual instinct. Being different from the world is not a failure to be corrected. It is a position to be understood and defended. A number that differs from the benchmark, where the practitioner can explain exactly why and trace it to a deliberate scoping decision, is more trustworthy than a number that matches the benchmark for reasons nobody ever examined. Conformity is not confidence. The practitioner who knows precisely why their number sits above the market, because they did the work the market cut corners on, has the stronger position, not the weaker one.

The decisions that put you there are the scoping decisions, and they were largely set back in Part 1: the goal, the functional unit, the allocation method, the cut-off rules, what co-products were included, how substitution and crediting were handled. These are the choices that move results, and they are the ones the plausibility check sends you back to examine. The series ends where it began, at scope, because the fitness of the outcome was determined, more than anywhere downstream, by how the thing was scoped at the start.

Three things get run at this stage and get muddled, so it is worth separating them cleanly, because they answer different questions.

Uncertainty is the band around the number that comes from the quality of the data feeding it. It belongs to Part 3 — it is a property of the inputs, and a Monte Carlo analysis propagates it through a fixed model. It tells you how wide the number is. It tells you nothing about whether the choices behind the model were right.

Sensitivity flexes a parameter inside the fixed model — vary a node up or down and see how much the result moves. It tells you what the result is most driven by. The model structure stays put.

Scenario changes the structure itself — a different boundary, a different allocation basis, a different crediting approach. It shows you whether a different but defensible model would give a different answer.

The outcome gate uses sensitivity and scenario. But, and this is the point most easily lost, running a scenario does not mean adopting it. The assessment is still the assessment. The practitioner makes one call, on one defensible basis, and signs their name to it. Scenario analysis tells you whether your conclusion hangs on that call; it does not turn the deliverable into a menu of every possible answer handed to the client. That would not be transparency. It would be abdication. The job of the gate is to know where the conclusion is load-bearing, defend the call you made, and disclose the dependency, not to refuse to make a call at all.

The standards already require a sensitivity check; ISO 14044 mandates it. But in practice it is often run thin; one parameter, varied lightly, reported to satisfy the requirement and never interrogated. A sensitivity analysis that exists to tick a box is not the check. The check is the one that actually tells you whether the conclusion the client is about to act on would survive the choices it rests on.

This round trip, build the model up, check it against the world, descend through the gates to understand and defend your position, was punishing by hand. Going back to re-examine the scope after the model was built meant redoing weeks of work, so most of the time nobody did. The first defensible number got shipped, and the deeper question of why this number and not another went unasked because there was no budget to ask it.

That is what modern compute and AI change at this gate. Re-running the model under a different boundary, pulling comparative benchmarks from a portfolio of prior work, tracking far more of the available data than a manual process could ever hold, interrogating why the number sits where it does; these become fast enough that the round trip is standard practice rather than a luxury. The descent becomes affordable.

The judgement does not move to the machine. The practitioner still decides where the boundary belongs, still defends the allocation, still makes the call. What changes is that the loop good practice always wanted, and rarely had time for, can finally be run every time, on more data, against more comparison points, in the time the work actually has.

After the climb, the plausibility check, the descent, and the rebuild of confidence gate by gate, the practitioner makes one final judgement before release. Not is the number precise. Not is it conformant. The question is: is this decision-grade, can someone act on it, for the decision it was built to serve, and can I stand behind it.

That judgement is informed by all four gates at once. A brilliant model on poor data is not decision-grade. A perfectly-sourced model answering the wrong scoped question is not decision-grade. The grade is the synthesis; the practitioner's release call, made in light of how the work stands at every gate, expressed as a single verdict on whether the number is fit to leave the building.

It is worth being clear about what this is and is not. It is not verification. Verification confirms a result is conformant; that the choices were permitted under the standard and executed correctly. That is a check on the floor. Decision-grade is a judgement about the quality above the floor: of the permitted choices available, were these the right ones for this decision, and does the conclusion hold well enough to act on. Two assessments can both pass verification and differ entirely in whether they are fit for the decision in front of them. Verification cannot separate those two. The practitioner's decision-grade judgement is what does.

Here is the uncomfortable part, true even of good work.

If a proper EPD or standard-compliant LCA is used, none of these choices are hidden. EN 15804, the relevant PCR, the standard itself all require the goal, the boundary, the allocation, the cut-off rules to be declared. The information is there. The discipline does not have a disclosure problem.

It has a communication problem. The choices, and the reasoning for why the number sits where it does, are declared in a methodology section, an annex, a referenced PCR. Technically present. Practically invisible to the person holding the number. And then the number travels, onto a tender or a comparison or a label, while the reason stays behind in the document.

This actively punishes the rigorous practitioner. A number that is higher than a competitor's because the work was done more completely reads, stripped of its reason, as a worse product. The practitioner who included the stage the market omits is penalised for it the moment the number leaves the page without its scoping attached. Declared-in-an-annex is not the same as communicated, and the gap does its most damage to exactly the people doing the most honest work.

So the open question, and it is genuinely open, is whether the practitioner's release judgement can be made into something that travels. Not a methodology annex nobody opens, but a structured, declarable signal of how decision-grade a number is, and for what scope, that moves with the number to wherever it ends up. The four gates each produce a judgement of quality. The question is whether they can be made to compose into one, a decision-grade that carries, that a non-expert can read, and that can be re-applied to every future scenario and restatement the work produces. We do not think the field has answered this yet. We think it has to.

The Digital Product Passport is where the number travels furthest from its context. A passport carbon figure is scanned by a buyer, a regulator, a competitor; none of whom will open a methodology annex, most of whom could not interpret a PCR reference if they did. It is the purest form of the communication problem: a number, properly declared somewhere, presented for comparison entirely stripped of the scoping that makes comparison valid.

If the DPP is to do what it is for, let products be compared on environmental performance, then the scoping that determines whether a comparison is even valid cannot stay in the annex. Goal, functional unit, boundary, allocation, cut-off, co-product treatment, substitution: the decision-critical choices this gate identifies have to travel with the number, headlined, not buried. A carbon figure on a passport without its scoping is the confident, precise, incomparable number this entire gate exists to catch, now printed on millions of products and offered to the market as if it meant something on its own.

That is the case for a decision-grade signal that travels. Not as commentary on the regulation, but as the thing that would make the regulation's data trustworthy: a number that carries its own answer to the question why is it what it is.

That closes the series.

Physical System Integrity — does the model obey physics, scoped to the right goal. Modelling Integrity — are the factors and the modelling sound, at the speed AI now works. Data Integrity — is the data honest and documented. Outcome Integrity — is the result real, understood, defensible, and fit for the decision it was built for.

But the four were never a straight line. You climb them to build the work, and then, at the outcome gate, you turn around and descend through them to decide whether to release it. The number meets the world; the world sends you back to your own scope; you return to the surface knowing why your number is what it is, and whether you can stand behind it. That descent is the practitioner's real work, and it is the thing AI finally makes affordable to do properly, every time.

Trust in AI-enabled LCA was never going to come from trusting the AI. It comes from a production system where the work arrives as a checkable artefact, the gates are real, the practitioner climbs them to build and descends them to release, and the number that leaves the building is one someone can act on; carrying, with it, the reason it is what it is.

The AI made the production fast. The gates make it trustworthy. The round trip is what makes it decision-grade.

If you want to talk through what decision-grade LCA looks like in practice, work that survives the descent and carries its scoping to wherever the number ends up, that is what we do at LCA Factory.

— Roger