How BankPulse uses AI, and how we keep it honest
BankPulse is built and run primarily by AI, working from a written set of rules its own operator (not a marketing team) wrote down and keeps binding on every change. This page explains those rules plainly, and backs them with our real, current numbers -- not a generic AI-ethics statement.
The three rules everything else follows from
- Fail closed, not open. If a fact cannot be positively proven -- is this rule still in force, is this number current, does this link still point to the real source -- we do not assert it. We show “under verification” instead of a confident guess.
- No fact is trusted forever. Status, numbers, and links are re-checked on a recurring schedule. Something verified yesterday is not assumed still true today without re-checking.
- Diagnose automatically, gate the cure. Our systems can find and report problems on their own at any time. But anything that changes what we assert as fact to a banker -- a status, a number, a date -- either passes an automatic fail-closed check or is queued for the founder to personally approve. The system never quietly overrides a fact-changing decision by itself.
Human oversight, specifically
A panel of independent AI models cross-checks material claims against each other before anything is treated as settled (models disagreeing is itself treated as a signal, not averaged away). Findings that only affect formatting, links, or already-agreed facts can be auto-applied and logged. Findings that would change what is asserted as regulatory fact are routed to a founder-approval queue and do not go live on their own.
What our numbers say today
These are the real, current tallies from our own review ledger (not a marketing estimate):
- 5,664 items put through the review process
- 5,346 passed the multi-model consensus check
- 307 quarantined (blocked from going live) pending further check
- 7 revised after review before publication
Tamper-evident audit ledger
Our fact-check ledger is notarised into a hash-chained log -- each entry cryptographically links to the one before it, so a past record cannot be silently edited without breaking the chain. As of the last recorded block: 2 block(s), covering 46 verified ledger entries, with 9 open (not-yet-resolved) findings honestly still on record as of that block.
Known limits, honestly
This is not a certified AI-governance program and we do not claim one. We do not run a published, third-party-audited evaluation benchmark yet. What is real today: a written, binding internal charter; a working fail-closed publishing gate; a tamper-evident audit ledger; and a public correction log that records every fix our own process has caught, without cherry-picking. Building toward a fuller, externally-verifiable governance program is on our roadmap, not yet complete.