A Tool That Asks Before It Guesses

The first time a verifier asks you to defend an emission factor you didn’t choose, you find out what your tools are worth. A row in a 200-line BOM — 304 stainless steel, no region named — came back with a clean GWP100 value. One number, no hesitation. Then the verifier asked the only thing that mattered: where did this come from, and why this one and not the other three? The tool had no answer. It had an output.

That’s the gap. Between an answer and an answer you can defend — and it’s wider than most LCA software lets on. What closes it isn’t a faster model. It’s a tool that knows when to stop and ask.

The cost of a confident guess

A model that always answers is easy to build and dangerous to file. Take stainless steel. Ask fourteen LCA databases for the same material and the factors can differ by more than 2× — Ecoinvent at cut-off, an EF dataset, a CarbonMinds record, each modeling a different route, region, reference year. To hand you one number, a tool has to bury that spread. So it picks. It smooths. It moves on.

And you inherit the pick. When the auditor walks the chain back, the choice was made by software no one can cross-examine, on grounds it never wrote down. The number isn’t wrong, exactly. It’s undefended — and undefended is the one thing a CBAM declaration or an ISO 14067 calculation file can’t be.

This is the quiet way automation fails in LCA. The work that looks automatable — match the row, return the factor — is precisely where a silent wrong turn costs the most, because it resurfaces months later in front of someone paid to doubt it.

Where a snail would stop

Cortex slows down at the points where speed costs you the audit. It searches fourteen LCA databases — HiQLCD, Ecoinvent, EF, CarbonMinds, and others — in parallel, and returns top-k, not top-1. Every candidate arrives with its GWP100 value, region, system model, source record, and a DQI score across five dimensions — Temporal, Geographic, Technology, Completeness, Reliability — in the lineage of the ecoinvent Pedigree Matrix. The spread is on the table. So is the provenance.

Then, where automation would break an audit, Cortex stops and hands the decision back. When fewer than 80% of BOM rows match with confidence. When the closest dataset is a proxy, not an exact fit. When the same material comes back more than 2× apart across databases. When a dataset is licensed and you have no access — Cortex won’t quietly slot in a literature value to keep the run moving. When the system-model match is ambiguous — cut-off versus APOS versus consequential — and the wrong call shifts the recycled-content accounting enough to matter, roughly doubling the result in some boundaries.

This isn’t a fixed list, and Cortex doesn’t count them off. These are the rows a verifier will circle. You decide at each one, and the decision goes into the reasoning chain — what was asked, what you chose, why. The way a careful snail leaves a trail you can walk back.

An answer is a number. A defensible answer is a number with the decisions that made it still attached.

It doesn’t guess your system model

The market’s instinct is to ship the tool that decides everything, because deciding everything demos well. Cut-off or APOS? The eager tool picks for you, silently, and the demo glides. But the system model isn’t Cortex’s to choose for you — pick wrong and the recycled-content accounting can move the result by tens of percent, double it in some boundaries. So Cortex returns the system model with every dataset and won’t mix them behind your back. When the match is ambiguous, that’s a stop, not a default.

The same restraint holds at the edge of what Cortex is. It doesn’t run the calculation. It plugs into the engine you already use — openLCA, brightway, 积木LCA — matching background datasets, building the product system, driving the run. It doesn’t replace the engine; it drives it. For SimaPro and GaBi, the door is open and the invitation stands.

And it doesn’t file for you. Cortex produces a reasoning chain, aligned with ISO 14067, GHG Protocol, and CBAM reporting requirements. You file. Alignment is a property of the work; certification is the verifier’s act, and no tool hands you either at the click of a button.

What the trail is for

Automation in LCA always promised the machine would do the judgment too. That’s the promise to distrust. The judgment — which proxy is honest, which spread you can live with, which system model the study actually requires — is yours, and it’s exactly the part a verifier is paid to test.

So the design question was never how to remove you from the loop. It’s how to keep your judgment in it without making you re-derive every row by hand. Cortex runs the parallel search across fourteen databases, scores every candidate, surfaces the spread, then stops at the four or five rows per hundred where a human has to decide. The rest it carries through. What it can’t defend on its own, it returns to you — and writes down what you chose.

It does not replace experts. It scales their judgment.

Start with one row you have to defend. Bring the stainless steel. Ask Cortex what the spread looks like, watch where it stops — then file the number you can stand behind. cortex.hiq.earth/chat