About these numbers

What we extract

For every UK company we surface, the latest annual accounts (AA) filing is processed through both pipelines:

• iXBRL parser — free, deterministic, runs on the tagged XML version of the filing. Reads the structured concepts the filer chose to tag.
• AI PDF extraction (Gemini 2.5 Flash) — paid, probabilistic, runs on the human-readable PDF version of the same filing. Reads every line item and disclosure the document carries, including auditor / going-concern / subsidiary metadata that iXBRL doesn't tag.

Both views are cached. The displayed numbers come from iXBRL where available; the AI-extracted PDF view fills in qualitative metadata and serves as a cross-check. For consolidated group filers, the PDF view is preferred because iXBRL on these filings often reports only the parent-company-only standalone numbers.

How reliable is the AI extraction?

We continuously run a "truth-check" that compares the two extractors filing-by-filing on the latest matched fiscal year. On a randomized n=200 cross-hub sample drawn 2026-05-09:

• Where both extractors return a comparable value (~60% of schema fields), the two views agree on the exact integer 80.2% of the time.
• A further 5.0% of disagreements are parent-vs-consolidated scope differences on group filers — both views are right at their own level. The PDF view gives the consolidated group number, which is usually the operationally correct one.
• 1.8% of disagreements reflect a UK iXBRL convention split on cash-flow outflows like dividends_paid: filers tag it as positive (magnitude) or negative (cashflow direction) about 50/50. The number's meaning is invariant under sign.
• The remaining ~10% are taxonomy edge cases (intangibles split, fixed assets gross vs net) and a low-single-digit residual of genuine extraction error.

Across all field-comparisons, the AI PDF view independently provides ~25× the unique coverage of iXBRL alone (38.5% PDF-only vs 1.5% iXBRL-only). Most companies' financial records on StatDesk would be substantially less complete without it.

Why "Group / consolidated" appears on some pages

When a company is the holding company of a group, its annual accounts typically report two views:

• Parent-only standalone — the holding company on its own, often with little operating activity (because the business is run through subsidiaries). Can show small or negative net assets.
• Consolidated group — the holding company plus all subsidiaries, which is the operationally relevant view of the business.

iXBRL filings for these companies frequently tag only the parent-only numbers. The AI-extracted PDF view captures the consolidated group view, which is what we display when the badge appears in the page header. The official Companies House filing will still show both — what we're surfacing is the financially meaningful one for risk / valuation / headcount purposes.

What we do NOT extract

• Multi-year history — the latest filing typically carries the current year + 1 prior. We don't currently walk older filings to backfill 5+ years of trend data per company.
• Full footnote text — we capture short verbatim excerpts of going-concern, related-party, accounting policy changes, and other high-signal notes. The full footnote corpus is not in the structured record.
• Unaudited filings beyond the rich tier — micro-entity and abridged filings deliberately omit P&L disclosure; we cover them with metadata only.

If a number looks wrong

Click through to the Companies House profile from the page header to see the original filing. If you find a meaningful discrepancy between StatDesk and the original, it's worth knowing about — the truth-check tooling continues to validate the pipeline against every randomized sample we draw, and confirmed corrections feed back into the extraction prompt or parser logic.