Evidence
What is verified, and what is not.
A company that publishes only the things it has proven has told you nothing, because you cannot tell what it left out. So the unproven list is first.
Last reviewed 23 October 2026 · this page is updated when the answer changes, not on a schedule
Not established
No broker export has been reconciled against a broker's own total
Riskora reads eight market segments and five broker column layouts, and every fixture proving that was written by us. They show the engine handles those shapes consistently. They do not show that any of them matches what Zerodha, Angel One, Groww, Upstox or Dhan actually emits, and they cannot cross-check a figure against a broker's own statement. Closing this needs real permissioned exports. Until then the formats are synthetically exercised, not certified compatible.
The Angel One layout's date order is an assumption
That layout is read as month-first, and its clock treated as a placeholder rather than a trade time. Both are assumptions about the layout, recorded as such, and both would be settled by one real export.
No academy has used the cohort layer
It is built and tested end to end, and sold to nobody. There are no case studies, no retention figures and no measured learning outcomes, because none exist. See Riskora for academies for the full position.
Multi-tenant isolation has not been tested in a staging environment
Cross-tenant reads, consent withdrawal, deletion and idempotency need two organisations in a staging environment with synthetic data. That environment does not exist yet, and seeding production to manufacture a passing test would be worse than having no test.
There is no minimum-cohort suppression
A cohort of any size returns aggregates. A small group can be identifying by inference even with no names attached, so this is a limitation to plan around rather than a protection we provide.
No user research, and no instructor evaluation
Nobody unfamiliar with the product has been asked what they think it is. No instructor has reconciled a Riskora report against an example they already understood. Neither can be substituted with internal opinion.
Performance is measured on a desktop, not a phone
The behavioural pass was quadratic and took 501ms on a 9,990-position book; it now takes 25ms for the same book with identical results. That is one desktop. There is no p50 or p95 from a mid-range mobile device, and no progress or cancellation UI for a long parse.
The operating entity is being reconciled
Two legal identities appear across this site's footer and its policy pages. Which one is the contracting and data-controlling entity is being established from business records with appropriate review, and every page will be made consistent once it is. We would rather say that than quietly pick one.
Verified, and how you can check it
These are testable claims rather than assurances. Where a route exists for you to check one yourself, it is given, because that is worth more than our word for it.
| Claim | How it was established | How you can check |
|---|---|---|
| Your tradebook never leaves the browser | A test uploads a file whose instrument names are markers no arithmetic can produce, exercises the report, the cost block, the hold view, the share card, a real PDF download, the email opt-in, a refusal and the sample, then searches every request's URL, headers and body. The detector is itself checked against nine deliberate leak routes, including a Blob body that a naive wrapper misses. | Open your browser's Network tab and run a report. |
| The score can be recomputed by hand | Every threshold, weight and formula is published rather than described. | The methodology page, against your own report. |
| Unknown is never reported as zero | An export without trade times withholds the score and every timing finding, with the reason stated, rather than reporting them as clean. This had been wrong: a date-only file once printed a red high-risk verdict from a score the engine had refused to produce. | Upload an export with dates but no times and read what it declines to say. |
| The PDF and the page are the same report | Both render from one versioned report object, and a test compares the downloaded PDF's text against the page for every shared figure. | Download the PDF and compare it with the report above it. |
| An ambiguous date is asked about, not assumed | A file whose dates could be read two ways produces no report until you choose, and each choice is labelled with the date range it produces. | Upload a file written in a form like 04/03/2024. |
| Every supported layout enforces the same safeguards | One contract suite runs the same cases against each layout: duplicate and conflicting execution IDs, cross-account matching, invalid contract economics, date ambiguity and timestamp honesty. This exists because one layout once bypassed all of them. | Not externally checkable. Stated because it was a real defect. |
How this was reviewed
A readiness audit was carried out on 22 October 2026 against the deployed product, followed by three verification passes as the findings were worked through. These were commissioned by Riskora and carried out with AI assistance rather than by an external auditing firm, and that distinction matters: they are an adversarial read of the work, not a certification of it.
What they found is more useful than the fact that they happened:
Sixteen findings in the first pass
Including a chronology error that flagged trades as reactions to losses that had not yet been realised, duplicate execution IDs counted twice, a contract multiplier ignored, and a second score presented on a credit score's range. All were reproduced before any were fixed.
Three more in the second pass, after the first fixes
Conflicting execution IDs resolved silently, cross-account trades paired, and invalid contract economics filtered out of their own validation.
One in the third, which mattered most
A single broker layout had its own code path and enforced none of the safeguards the others did. Every automated check passed throughout, because every one of them went through the other path. A safeguard proven on one layout is not a safeguard, and the suite now runs the same cases against every layout.
The automated checks
There are 2,304 of them and they all pass. That number is the weakest evidence on this page, and it is last for that reason: the third review's finding above is exactly what a large green number fails to catch. 255 of the 2,304 have been rerun independently of us. The rest are ours, run locally.
They are useful for one thing, which is that each defect above now has a test that reproduces it, and each of those tests was confirmed to fail against the code before the fix.
What this page is not
It is not a security audit, a penetration test, a regulatory approval or a certification of any kind. It is a list of what has been checked and what has not, kept current, so that you can decide how much of it you believe. If you find something on this site that overstates what Riskora does, write to support@riskora.in and it will be corrected here.