Guide updated
03 / Use of AI
How AI is used
and checked
Language models read where fixed rules cannot. Code checks every answer, and the workbench labels it as AI.
Follow one answer ↓Registers, identifiers and tables
Sentences the rules cannot decide
Quotations, agreement and numbers
Labels and review
01 · Code readsRegisters, identifiers and reported tables are read by code, not by a model.
01 / One answer, followed
From a passage to
a labelled answer
A fictional Northbank report, read the way the engine reads it.
Code reads what it can
Home loans are recorded from passage 1. Passage 2 names a partner, so the rules cannot tell whose life cover it is.
A model selects a quotation
It sees only the supplied passages and must quote the words it relies on.
Code checks it word for word
The quotation must appear exactly in a passage the model was given.
A second reading must agree
A separate reading reaches the same answer. If they disagree, no conclusion is kept.
Labelled beside its source
The result is marked as a two-pass AI review, not human-verified, and linked to its passage.
A person decides
A reviewer’s decision takes precedence. The source text is never changed.
02 / Where AI is used
Where AI is used,
and where it is not
Models read. Calculations, identifiers and amounts stay with code.
Reading tasks in the Business Model Engine
Models read where names and tables run out: which report, which activity, which legal entity, which place.
Climate evidence: draft, then check
A model drafts commitments and findings; a separate pass checks each against its quoted source.
Hazard figures stay with code
Values come from published data. A model may only annotate one reading or choose between place names.
What a model never does
Identity comes from registers and amounts from tables. No score or ranking is calculated.
All model-assisted steps, in detail
| Task | What the model does | How it is checked | Shown as |
|---|---|---|---|
| Business Model Engine | |||
| Finding missing reports | Searches for an expected disclosure that the analysis did not find, and suggests candidates. | Candidates pass the same checks as any other source. The document’s own cover must name the company or its parent. | A source, with its search record |
| Whose report it is | Reads a document’s opening pages when its issuer cannot be settled from names alone. | The answer must quote the words that identify the reporting entity. Entity names are compared exactly, never by partial match. | Reporting entity, with the deciding quotation |
| Business activities | Re-reads activities the rules marked as uncertain, and can propose activities the rules missed. | Answers must cite the supplied passages. An activity can be withdrawn only after the model has seen all of its evidence, with an exact quotation for each objection. | Inferred |
| Which legal entity | Asks whether the company, another entity or neither can be shown to carry on each activity. | Two separate readings must agree. The quotation must itself name the entity. Parents and subsidiaries come from filed ownership records. | Two-pass AI review · not human-verified |
| Places | Decides whether a passage places the company in a country or only names it, and what a named place is for the company. | One passage and one question at a time. Quoted grounds are checked. “Unclear” is allowed, and an undecided case does not count as presence. | Place reading, with quoted grounds |
| Profile check | Flags readings a supervisor would question, such as an insurance line at a building society. | The engine, not the model, gathers the evidence. A conclusion without valid citations is marked unresolved. | Plausibility flag, with its evidence |
| Climate evidence in disclosures | |||
| Commitments and delivery | Drafts a register of targets, policies, progress and governance from the original passages. | A separate checking pass reviews every record. Values must appear in the quoted source, and a sentence containing a number not found in its quotation is rejected. | AI-assisted summary · source checked |
| Transition exposure | Separates facts about existing activities from the company’s own assessments and plans. | The same draft-and-check pattern. Findings whose quotations are not found in the source, or whose numbers differ from them, are dropped. | Finding, with exact quotations |
| Plain-English summaries | Rewrites checked findings as short sentences for a supervisory reader. | Each sentence cites its inputs and is checked against them. In climate summaries, a sentence with a number not found in its inputs is rejected. | AI-assisted wording · checked against the cited evidence |
| Climate Hazard Engine | |||
| Notes on a reading | Writes a short note on what one hazard reading shows and what limits it. | The model sees one reading only. Every figure in the note must match a published value in that reading, and words that rate the value, such as “high” or “severe”, are refused. Otherwise the note is withheld. | Model-written note |
| Ambiguous place names | Chooses between gazetteer candidates when a name matches several places. | The choice is never shown as a gazetteer match. | Inferred · requires review |
03 / Checks
Four ways an answer
is stopped
Code runs these checks after the model replies.
Quotation found: kept, labelled as AI
The quotation appears word for word in the passage the model was given.
Quotation not in the source: not kept
Invented words are never published as the company’s own. The question stays open.
Readings disagree: left for review
No affirmative conclusion is kept. A more cautious reading survives only if it passed every check.
A calculated number: note withheld
The model worked out 87%; it is not in the reading. Rounding a published value is allowed, new figures are not.
04 / In the workbench
How AI appears
in the workbench
Every model contribution carries a label and opens its evidence.
Read by rules; opens its passage.
Proposed by a model, with the passages it cites.
Two readings agreed; the quotation was found.
Written from checked findings; each sentence cites them.
A written rule sets the level from AI-checked criteria.
The evidence could not decide. Not a finding of absence.
05 / Data & limits
Data, models
and limits
A research pilot using public sources only.
Public information only
No confidential supervisory information is sent to a model or used to train one.
More than one model
Commercial API models and open-weight models on university computing, chosen per task.
Recorded and repeatable
Each saved answer records its model and is tied to its exact inputs.
Run deliberately
Opening a page never starts a model. Documents are read as data, never as instructions.
Models can misread
Checks limit the damage but cannot prove a passage was understood, so results stay labelled until reviewed.
Measured, not assumed
Changes are compared on reviewed reference cases before adoption. Accuracy figures are not yet published.
About this guide
This guide describes how language models are used in the Greenophy research pilot as of the date above, across the Business Model Engine and the Climate Hazard Engine. Text from scanned pages comes from optical character recognition, with its provenance recorded. A search that finds no evidence does not establish absence. Firm-level materiality and completeness are not verified. The Northbank example and the checking cases are fictional.