Algorithmic accountability at AssureLocker

Every signal is cited to the evidence behind it, and our agentic AI writes a plain-English risk narrative where each claim links to its source — and says so plainly when the evidence doesn’t cover a question. The logic is inspectable; the read is traceable; the decision stays yours.

Bias in decision systems is almost never designed in — it’s inherited from training data, defaults and the assumptions of whoever built the system, and it stays invisible until someone tests for it. We map our inputs for proxy effects and test them for systematic disparity — calibrated for India — and we document what we find.

And this isn’t a promise — it runs on every release: we test the scoring engine for identity-invariance (the result moves on the evidence, never on who or where the borrower is), monotonicity (stronger evidence never scores worse) and reproducibility; we adversarially check the narrative so it cites its evidence and won’t opine on creditworthiness; and we periodically test live deals for proxy-correlation across vertical, geography and size. Every run is recorded, and whatever a test flags opens a documented action we work and re-test.

• We name our proxy vectors rather than assume their innocence — location, clustering, scale and behavioural signals are exactly where proxy effects hide, so those are where we look first.
• We test for systematic disparity that our decision criteria don't justify, and we calibrate for India — its social hierarchies, its data-reliability gaps — instead of importing a framework designed for somewhere else.
• We separate “not enough data” from “elevated risk,” so a thin-file, first-time, tier-2/3 borrower is never quietly penalised for being new.

AssureFirst — our duplicate-financing check — is a deterministic fingerprint match, not an inference about a person or a firm. It carries no demographic surface to be biased by. Where we do use inferential signals, we test and document them.

Evidence is recorded in an immutable registry and post-quantum signed (ML-DSA-65), with holder-signed consent on every check — a tamper-evident trail your auditors can verify without taking our word for it. Every signal carries its evidence tier, from registry-verified through to self-declared, so nothing reaches your desk as an unexplained number.

Ask us for the Vendor Algorithmic Audit Pack: a per-signal record of function, inputs, explicit non-inputs, explainability and testing — mapped to FREE-AI, the SEBI white-box / black-box distinction and the DPDP Act. We make the duty easier to meet; we don’t pretend to discharge it — testing how you use our signals in your own pipeline stays yours.