Numbers come from tools, not from the model.
GWP values, DQI scores, and dataset descriptions come from curated databases. The LLM handles question parsing, candidate ranking, and explanation — not emission factor generation.
Methodology
How Cortex produces an answer you can defend.
For reviewers, auditors, and procurement teams who need to understand what Cortex does before they trust a number it produced.
Section
§ I
Principles
Domain knowledge informs commentary, not data.
When Cortex says "1040 is typically for automotive leaf springs", that is commentary. When Cortex says the factor is 2.31 kgCO₂e/kg, that comes from a cited, clickable dataset row.
Restricted rows stay restricted.
For licensed datasets the user hasn't accessed, Cortex returns a truthful "⛔ no access" marker — not a substituted literature value passed off as data.
A chain of reasoning is part of the deliverable.
Every Excel export ships with match type, proxy notes, and source URLs. A downstream reviewer can reproduce or challenge any cell without re-running the system.
Section
§ II
Data sources
Cortex searches curated LCI and LCA databases, each indexed with its own unit, region, and system-model vocabulary. Nothing is merged blindly.
01 source
HiQLCD
China-focused primary data, 800k+ unit processes
02 source
Ecoinvent
Global LCI, cutoff + APOS system models
03 source
EF (Environmental Footprint)
EU-authorized datasets for PEF and CBAM scope
04 source
CarbonMinds
Chemicals and polymers, process-specific
05 source
Industry secondaries
Supplier-issued EPDs and verified PCFs, on request
Section
§ III
Data Quality Indicator (DQI)
Cortex uses DQI as its single confidence vocabulary because it's what reviewers already speak. A low score on any axis appears in the export — not hidden inside a probability number.
§ 1 axis
Temporal
Is the dataset's reference year within the target window?
§ 2 axis
Geographic
Does the region match the BOM row, or is it a reasonable proxy?
§ 3 axis
Technology
Does the production route match — BF-BOF vs. EAF, virgin vs. recycled?
§ 4 axis
Completeness
Are the upstream cradle-to-gate steps modeled, or scoped out?
§ 5 axis
Reliability
Source type: measured, calculated, expert estimate, or literature.
Section
§ IV
Human-in-the-loop gates
Three conditions are hard-wired to pause the agent and return control to the user. These are the moments where silent automation would break an audit.
Trigger 01
Coverage below 80%
Cortex pauses and lists unmatched rows with candidate datasets. The user picks: expand databases, accept a proxy, or defer.
Trigger 02
Proxy substitution required
Cortex names the proxy explicitly — which dataset, what differs (region / process / DQI), what the delta implies. The user confirms or replaces.
Trigger 03
Cross-database disagreement > 2×
When the same material returns factors that differ by more than 2× across databases, Cortex surfaces the spread rather than averaging silently.
Section
§ V
Standards alignment
Cortex is designed to produce outputs compatible with the standards below. Formal certification against a specific standard is scope-dependent and available on request.
ISO 14067 Carbon footprint of products — boundary, allocation, reporting std.
ISO 14044 LCA requirements and guidelines — the common substrate for tooling and auditors std.
GHG Protocol Product Life Cycle Accounting and Reporting Standard std.
PEF / EF 3.1 EU Product Environmental Footprint method and EF datasets std.
CBAM Emission factor selection for embedded-emissions reporting std.
Last reviewed: 2026-04-25. Updates tracked in the changelog.