A claimant-intelligence platform for a trust that already knows what it doesn't know.

Why they go missing

Mine employment records, TEBA records, ID data and medical files sit in different systems with inconsistent name spellings, mine numbers and dates. A person who worked under three different mine numbers across two provinces doesn't look like one claimant. They look like three partial ones.

Why it's a data problem, not just an outreach problem

Outreach without matching wastes field time on people already paid, already deceased with no traced dependant, or duplicated across source systems. The Trust needs to know where to look before it decides how to reach.

Reading it honestly: this isn't one straight pipeline

Registrations (414,885) run far ahead of Contact Made (53,945) and Appointments (191,592) exceed Lodgements. These are 8 process counters with different denominators, not one claimant flowing cleanly through 8 gates. The MCP-vs-BME question I flagged before now has a real answer: the Trust's own medical eligibility breakdown shows 119,555 MCP decisions splitting into 31,375 medically eligible, 78,144 medically ineligible, and 9,401 deferred. MCP counts decisions, including historical-TB cases that skip a fresh BME, which is exactly why it doesn't have to track BME volume 1:1.

A found data-quality note

The homepage banner still reads "over R2.5 billion paid," while the live counter beside it reads R2.73 billion. A stale hardcoded banner next to a real-time figure. Small, but exactly the kind of inconsistency a monthly ICT/data QA pass (Section 07) is built to catch before a trustee or journalist does.

How the labour system worked

Gold was discovered on the Witwatersrand in 1886. To staff it, the mines built two recruiting networks: the Witwatersrand Native Labour Association (WNLA, "Wenela," est. 1900) drawing from Mozambique and further afield, and the Native Recruiting Corporation (NRC) drawing from Basutoland (Lesotho), Bechuanaland (Botswana), Swaziland (Eswatini) and South Africa itself. The two merged in 1977 into TEBA. The Employment Bureau of Africa, the same archive this proposal's matching pipeline depends on.

Workers signed annual contracts and lived in single-sex hostel compounds. Over 97% of a roughly 500,000-strong workforce, cut off from family for the length of each contract, then rotating home before the next one. By the 1970s, about 70% of gold-mine labour was foreign-born. That oscillation between mine hostel and rural homestead is precisely why claimants and their descendants are hardest to find in exactly the regions that supplied the labour. The regional rankings in Section 06 aren't arbitrary. They're the recruitment map, inverted.

Why the disease burden is what it is

Autopsy studies of South African gold miners found silicosis prevalence rising from roughly 3% (Black miners) and 18% (white miners) in 1975 to 32% and 22% respectively by 2007. And in miners with around 40 years of dust exposure, silicosis was found in up to 52% at autopsy.

Tuberculosis followed the same exposure. South African gold-mine TB incidence has run at roughly 3,000–7,000 cases per 100,000 workers a year. Against a South African national average of 981 and a global average of 128 per 100,000, and more than ten times the World Health Organization's emergency threshold of 250 per 100,000. Rates peaked above 4,000 per 100,000 in 1999, compounded by HIV co-infection running near 29% in the mine workforce by 2001. Silica dust independently raises TB risk, which is exactly why the Trust compensates the two diseases under one scheme rather than two.

Questions it must answer on demand

Who is eligible but not yet contacted? Which regions carry the highest unresolved caseload? Which claimants are stuck on missing documents or missed medical exams? Where should the next outreach trip go?

Data quality as a control, not an afterthought

Duplicate detection, missing-document alerts, mine-history validation and medical-status checks run continuously. Because a bad match here doesn't just skew a dashboard, it delays or misdirects a real payment.

Spatial NNA: informal settlements and job access

Migrant workers in mining regions like North West and Free State typically settle near mine shafts in informal or shack settlements, not the formal towns in Section 06's site map above. Research on migrant labour markets shows physical proximity carries real weight: workers living closer to already-employed neighbours hear about vacancies and recruitment notice boards sooner, and get referred faster. Spatial NNA is the tool that tests this. Run against settlement GPS points (not yet in a public dataset, the gap a future field survey or drone/satellite imagery pass would close), it would prove or disprove whether tight settlement clustering near a shaft head gives job seekers a measurable time-to-employment advantage, the same clustering logic already proven significant on the Trust's own 17 real sites above.

Chain migration: predicting where a claimant worked

A migrant's most useful "nearest neighbours" are often social, not physical: kinship, home language, village of origin. Migrant mining history runs on chain migration, a worker from a specific district travels to a specific mine because a relative or fellow villager is already there. That is a k-NN problem: encode a claimant's home country, language, village and education level as features, and their nearest social neighbours in the existing employment record predict which mine or shaft they most likely worked at, even when the claimant's own mine number is missing or illegible. The same feature-distance method used for identity resolution below, pointed at a different question.

Fixing broken records with k-NN

Exact database joins fail on a misspelled name or a partial ID. k-NN treats each record as a point in feature space (birth year, name, mine number, region) and merges whichever records sit closest together, not just the ones that match exactly. Run on 8 synthetic demonstration records built to mimic Section 01's example, the model correctly clustered the three-partial-record case at distance 0.02 to 0.05, tight, while showing why naive tokenization still needs fuzzy name-matching in production: a shortened first name ("S" vs "Sipho") pushed one true match further out at distance 0.53. That's the argument for edit-distance matching, not simplistic hashing, in the real build.

Finding dependants through kinship proximity

Dependant Silicosis and TB claims exist because a worker died and the family never knew. "Proximity" here means shared surname, shared rural local authority, or the same TEBA recruitment batch. When a deceased, uncompensated miner's record surfaces, the same nearest-neighbor logic scans current identity and mobile registries for the closest social matches, relatives sharing lineage and geography, giving field teams an algorithmic starting point instead of a cold call to an entire village.

Controls stack

MFA on every account, role-based access mapped to the table above, conditional access by device and location, encryption at rest and in transit, and audit trails on every read of medical or identity data. Because POPIA treats health and biometric data as special personal information requiring explicit safeguards.

Disaster recovery

Cloud hosting doesn't remove the need for backup discipline: automated, geographically separated backups of claimant, medical and payment data, a defined recovery-time and recovery-point objective, and a tested restore. Not just a scheduled job nobody has verified.