Methodology
What this site is
Sly by Design is a public registry of contract mechanics in South African financial products — the clauses that govern how rewards are earned, reduced, gated and lost. Every entry is built on verbatim quotes from the provider’s own published documents, each with a version date, a capture date, a page reference and a stable record ID. The registry describes what the documents say. It does not speculate about what anyone intended.
The first institution documented is Discovery, across the Vitality Health, Vitality Active, Vitality Money and Discovery Miles programmes.
How entries are produced
Document capture. Provider documents are downloaded from the institution’s own public channels — for Discovery, the published Vitality product-rules pages and Discovery Bank’s document server — and archived with the date of capture. The corpus for the first release comprises 273 PDF documents captured between 21 and 23 July 2026. The provider’s own website structure is mirrored exactly, including its naming inconsistencies, and filenames are preserved as served, because the name a document is published under is part of the evidence. Where the same document is published in more than one place, every copy is kept: parallel copies are how differences between product-combination tabs become visible.
A defined extraction schema. Clauses are extracted under a written schema (version 0.2) that defines eight neutral categories — consent, unilateral change, forfeiture, qualification, marketing supremacy, exit barriers, cross-product gating, and other — together with a set of clause types that must always be recorded, such as fees to participate, clawback of benefits, and acceptance-by-conduct terms. One record is one clause coded to one category; a clause that fits two categories is recorded twice. The schema forbids severity ratings, harm scores and opinions in the data.
AI-assisted extraction with machine validation. Clause identification was performed with the assistance of AI language models (Anthropic’s Claude), working from the extracted text of each document, every page of every document. This assistance is disclosed plainly because it is part of the method: the models proposed the clauses; the pipeline then machine-validated every proposed quote word-for-word against the source document’s text before it entered the dataset, computed its physical page reference, and verified duplicate documents byte-for-byte by checksum. Quotes that could not be verified against the source were excluded rather than published. The result is a dataset of 4,952 records across the 273 documents.
Human verification. The researcher reviews entries against the dataset and the source PDFs before anything is published. Every registry entry is written as a draft and only goes live after that review. Known limitations of the pipeline — including two footnote clauses that an over-aggressive filter initially dropped and that were later restored, and a small number of assumptions about capture dates — are recorded in the project’s extraction log rather than silently fixed.
Evidence standards
Every quoted clause on this site is verbatim from the source document. The only permitted change is the normalisation of font-encoding artefacts — for example, the stylised “Д glyph Discovery’s brand font uses for the letter D — and each such normalisation is logged against the document it affects. Nothing inside quotation marks is trimmed, paraphrased or cleaned up.
Every entry cites the source document by its printed title and archived filename, the version date printed in the document, the date the document was captured, the physical page on which each clause appears, and the dataset record IDs behind the entry.
Where a provider’s own documents disagree with each other — different values for the same limit, different versions of the same document under different names — the registry presents both quotes, both dated, and does not adjudicate which is correct. The disagreement is the finding.
What the registry is not
The registry contains no severity ratings, no scores and no rankings of institutions. It does not claim to know what any institution intended, and it does not assert that any documented mechanic is unlawful. It is not legal advice, and it is not financial advice; readers making decisions about a product should read the provider’s current documents and, where it matters, take professional advice. The registry’s claim is narrower and firmer: on the date of capture, the documents said what the entries quote them saying.
Corrections and right of reply
Anyone — including the institutions documented here — may request a correction by emailing corrections@slybydesign.com.
A correction request should identify the entry, the statement in question, and the document evidence it conflicts with. Every request is reviewed against the archived source documents and the extraction dataset. Where a provider has published a newer document, the review compares versions; a change in the current terms does not falsify what an archived version said, but it will be recorded as an update.
Where a correction is warranted, it is made in the entry itself, the entry’s status field is changed to corrected (or updated, where the change reflects a newer document version rather than an error), and the nature of the change is noted in the entry so that the correction is itself part of the public record. Corrections are not made silently.
Institutions documented in the registry may also submit a factual response to any entry concerning them. Responses that address the documented mechanics will be linked from or quoted in the relevant entry, verbatim and attributed, under the same evidence standards as everything else on this site.