BOM in. Matched xlsx out. Every dataset cell a link.

Input: An Excel file, a CSV, or a pasted list. Column names, part numbers, free-form specs — Cortex reads structure as it finds it.

Output: One .xlsx workbook with four sheets, delivered in the session window.

A BOM of any size — from a single material to thousands of rows — processed against Cortex's fourteen-database foundation in Cortex Cowork completes in minutes. The search runs on a unified index — not sequential queries across each source — so breadth of coverage does not multiply the time.

The manual equivalent: each row requires opening the database interface, entering search terms, comparing results, and recording the selection. At a conservative 3 databases per material and 3–5 minutes per lookup, any serious BOM runs into the days of cross-database lookups it would take by hand before a single comparison begins.

The speed difference is a function of parallel search workers running against a unified index. It is not a rounding claim.

Row 3, HiQLCD column: the dataset cell reads 钢坯, 转炉, 304不锈钢 and is underlined blue. Clicking it opens https://www.hiqlcd.com/dataset/HiQLCD/1.4.0/CUT_OFF/… in a browser — the exact record used. The adjacent GWP cell shows 1.95 kg CO₂e with a blue shading proportional to its rank in the column.

The Ecoinvent column for the same row shows market for steel, chromium steel 18/8, hot rolled — and the GWP cell reads Licensed, because Ecoinvent access requires a subscription. The ∆% column is empty; no comparison can be made without both values.

No literature substitutes. Ever. That Licensed cell is the correct behavior. Cortex does not substitute a literature estimate, a model-generated number, or a generic proxy to fill the gap. The dataset's name, region, unit, and link are shown. The GWP value waits for access.

The Methodology sheet closes the workbook: "GWP values are GWP100 (kg CO₂e) per IPCC AR6 characterisation factors." The parameters table records which databases were queried, which version of each, and which system model governed the run. The report is reproducible from those parameters.

When no single dataset matches a BOM line, Cortex may use a weighted average. Brass, for example: 65% copper cathode × 35% zinc ingot. The Notes column in the Results sheet shows the decomposition explicitly — each component's dataset name, its weight, its GWP value, and the arithmetic. Each component row still carries its own hyperlinked dataset cell. The proxy is traceable to its sources, not collapsed into a single opaque number.

Proxy rows carry ~ Proxy in the Status column. The Notes column names the bias direction: if the substitute is likely to understate embedded energy, that is written out. Proxy status is not a footnote condition — it sits in the row alongside matched rows.

When no proxy is defensible, the Status column reads ✗ Gap. The Gaps sheet collects those rows with the search terms tried and the reason for failure.

When a BOM batch returns coverage below 80% after three search attempts, Cowork pauses and presents the outstanding rows. The practitioner chooses:

• Continue with generic proxy values
• Mark remaining rows as not applicable
• Confirm row by row

That decision is recorded in the project's progress log with a timestamp. The export reflects what the practitioner chose. No silent default.

The same mechanism triggers for other conditions: a material name ambiguous enough to produce 3–10× GWP spread across candidates; a system model conflict between the active source and what the project requires; a row where the only available match is a proxy and the practitioner should confirm it before it enters the workbook.

Cortex explains its analysis before pausing — the reasoning appears in the session before the input fields. The practitioner sees the spread, the candidate options, and the implications. The decision form appears after that context, not instead of it.

Chinese factory BOMs use factory-floor names: "502 胶水," "IGBT 模块," "背钩." These are not LCA dataset names. From watch components to EV power modules to industrial adhesives — Cortex carries thousands of curated mappings from Chinese factory terminology to LCA dataset identifiers, built over years of practitioner decisions. "502 胶水" resolves to polyurethane adhesive (marked proxy, with the bias direction documented). "IGBT 模块" resolves to the insulated gate bipolar transistor dataset for electric vehicle applications.

Cortex searches both Chinese and English fields simultaneously — Chinese input terms are automatically matched against Chinese-language records while also reaching English-named databases like Ecoinvent and EF. The system recognizes tens of thousands of LCA-specific Chinese terms as distinct material identifiers, distinguishing between terms that look similar but refer to different materials.

HiQLCD covers every Chinese provincial grid. The difference between China's coal-heavy grid and the European mix is decisive for materials like primary aluminum: 14–20 kg CO₂e/kg in China versus 8–10 in the EU.

Getting the geography right is not optional for a product carbon footprint or an EPD targeting a Chinese supply chain. The database has the data; the match gets the geography.