Method.
Same metric. Same year.
Different rulebook.
Companies.
Compare.
Where Kestrel reads the words, Strix reads the numbers.
Strix is a structured ESG-metric engine. It extracts the actual figures companies disclose — with full provenance — normalizes them, records the reporting basis behind each one, and tracks how they change over time.
Pull every disclosed ESG metric from factbooks, data tables and reports — each tied to its source document, sheet, cell and page.
Convert every figure to a canonical value and unit, so numbers compare cleanly across companies, years and formats.
Capture the reporting basis behind each number: consolidation boundary, GWP set, assurance, managed-JV and contractor inclusion, reporting standard.
Follow targets and commitments year to year — and flag the ones that quietly disappear from later reports.
Surface restatements and disclosure events where a previously reported figure silently changes.
Each observation is pinned to the exact place it came from and converted to a comparable unit — so a figure can always be audited back to source.
Two companies can report “emissions” on entirely different bases. Strix records the boundary, standard and assurance so figures are compared like for like.
Because Strix holds every figure across years, it catches what a single read-through never could — figures silently restated, and targets that quietly vanish from later reports.
Every figure shows which version of our extractor produced it.
We improve how figures are read over time. When we do, the figures that change are marked with the new version, so you can always see how a number was produced. Whatever the version, a figure is published only if it can be traced to the exact words or cell in the company’s own document; anything we cannot confirm is held back until a person checks it.
Our extraction model reads each company’s published data books and reports and records every ESG figure it finds, together with where it came from.
- Each figure is stored with the document, the sheet and cell or the page it was read from, and the words around it.
- Units are converted so figures can be compared, while the number as the company printed it is kept.
- Figures read from PDF reports are checked against the page they cite.
We re-checked figures against the companies’ own documents and fixed cases where a correct number had been given the wrong label.
- Percentages written as fractions (0.29 instead of 29%) are now read correctly.
- Units taken from the wrong place in a table, and site-level figures filed as company totals, were corrected.
- A correction is made only when the source document itself shows the right reading. The number the company reported is never altered.
- Every change is logged and can be traced or reversed.
Some figures look wrong because the company published them that way, not because we misread them. We now check which it is.
- Each unusual figure is compared with the cell or page it cites, to tell a company’s own error from a reading error, which we fix.
- We look through the company’s later reports for a correction or restatement and record what we find.
- No reported number is changed by this check.
Tighter rules on how a figure is identified, so the same figure is never shown twice or under the wrong year.
- When a company reports one measure in two units, both are shown instead of only one.
- The year of each figure is taken from the source document itself, never guessed.
- A figure that appeared twice under two labels is shown once.
- If a newer report’s figure cannot yet be confirmed, the older report’s figure is shown only when the two agree exactly.
A figure is now checked against the row label, table title and section heading printed around it in the company’s own workbook.
- A part of a table that had been filed as its total (for example contractors only, freshwater only, one division) is now named as that part.
- Tables printed side by side are read as separate tables, each with its own title and year heading.
- Units are read from the source, never assumed; figures whose unit the source does not state are held until it is confirmed.
- Restatements the company marks itself (italics, shading, a “restated” heading) are recorded as restatements.
Every large change between one year and the next was traced back to the source, to tell a real change from a reading error.
- Totals read from an intensity table, or from one column of a table’s total, are now read from the right place.
- Percentages stored as fractions, kilolitres and water reported by weight are converted correctly.
- Figures read from a chart’s labels or from an unrelated table are held until a person checks them.
A random sample of published figures was checked one by one against the source. Every kind of error it found was then searched for across all companies and corrected.
- A part shown as a whole is now named as the part: contractor-only or employee-only safety figures, a single disease, one energy source, one water source, one region or one programme.
- Waste sent for disposal is no longer shown as waste generated, and the right row is used where a different measure had been read.
- Tables that combine Scope 1 and Scope 2 emissions are filed as combined emissions, and figures stated on an equity-share basis are labelled as such, with the operational-control figure read from its own column.
- Each figure takes the year its own column states; injury and illness rates show the base the company prints (per 200,000 or per million hours, per 1,000 workers).
- Intensity figures are shown with the number and unit the company prints.
- Every new extraction is now checked for these patterns before it can be published.
A small number of figures that the extractor could not read were entered by hand from the source document, with the same source details.
How Kestrel reads the words — AI-extracted evidence, rules-based judgement.
Kestrel is the platform’s qualitative engine. It reads the narrative side of mining reporting — the sustainability, social-performance and governance prose — and turns thousands of pages of it into structured, comparable data: every passage scored against four explicit Creating Shared Value criteria, separating genuine, evidenced community and environmental benefit from PR. It is deliberately conservative — designed to err toward false negatives, never false positives — so only passages where real shared value is simultaneously and explicitly evidenced qualify as positive.
Read the PDF and segment it into discrete passages of 100–220 words — one passage per request, so each judgement is independent.
A language model extracts evidence from the passage — the four CSV definitional elements, the linkage claimed, and the supporting text. It does not assign the label.
Deterministic, human-authored rules turn that evidence into a label. Positive only when all four elements are simultaneously and explicitly present; any ambiguity resolves downward.
Passage labels roll up into company and sector profiles: share of true shared value, initiative types, CSV pillars and evidence strength.
A two-layer instrument. The model extracts evidence; human-authored rules — versioned and applied deterministically — turn that evidence into a label. No model is ever asked to decide what qualifies as shared value.
Five structured judgements per passage — here is how all 105,227 fall.
Kestrel’s headline call on every passage — Positive: genuine, evidenced shared value; Ambiguous: shared value claimed but not clearly evidenced; Negative: no measurable shared value in the passage.
How far an initiative actually goes — Transactional (one-off giving or compliance), Transitional (improving existing operations), Transformational (reshaping the wider system). “No initiative” means the passage describes none.
Which lever of shared value the activity pulls (the three ways business and society gain together) — Products & markets (serving unmet needs), Value-chain productivity (greener, fairer operations), Local cluster development (strengthening the surrounding economy).
How clearly the passage ties the activity to a real outcome — Explicit (stated and evidenced), Implied (suggested but not shown), None (no link made).
How hard the supporting evidence is, graded Tier 1–4 — Tier 4 is independently verified data, Tier 1 a bare assertion. “None” means there is nothing to check.
We ran every passage 5 times. The labels hold up.
Inter-run agreement was measured with Fleiss’ κ across 5 independent classification runs per passage. Overall agreement is substantial; on the headline question — whether shared value is present at all — it is almost perfect.
Initiative type is the least stable dimension: the boundary between Transactional and Transitional is the main source of disagreement, and those definitions are the top candidates for refinement.
of passages received the identical label across all 5 runs (5/5).
received the same label in at least a majority of runs — a clear consensus for nearly every passage.
Kestrel agrees with expert coders 85% of the time.
On a held-out sample, V2.0 matched expert human labels on the headline is this shared value? question 85% of the time. Exact three-category agreement is lower at 38% — and that gap is itself informative.
Humans are becoming the limiting factor. Raters routinely infer beyond what a report actually discloses and lean generous — awarding shared value where the evidence on the page doesn’t support it.
V3.0 targets >90% — with a refined rubric around Transactional vs. Transitional and a larger, multi-coder adjudicated sample.
A Conservative, Rule-Governed Classifier for Creating Shared Value in Corporate ESG Reporting
How Kestrel reads the narrative: the CSV classification schema, the conservative default, the text-only evidence rule, the two-layer decision architecture (the model extracts evidence; rules decide), and the validation approach — every design decision stated and defended.
The Evidence Layer Behind Kestrel — Structured, Audit-Traceable ESG Extraction
How Strix reads the numbers: the four-layer data model, provenance to the source cell, additive unit normalisation, basis-aware comparability, and how restatements and quietly dropped targets are surfaced.