Banking & Financial Services
You are not asked whether the control exists. You are asked to show that it operated.
Payment instruction changes, credit decisions, fee calculations, transaction monitoring, model governance, third-party risk. Every one is a written rule applied to a flow, under an authority that expects the rule to be demonstrable rather than asserted. Read this page as an architecture rather than a single product: it describes the Runink TIDE platform working with Atlas, an assessment platform from Logical Leap.
Where the decision happens
- Records you already keepSpread across the systems you already run.
- One finding, with its evidenceDrawn together into a single proposed action.
- A person approves itThe drafted action waits; approving is what sends it, under their name.
- The work goes outAnd the reason stays on the record.
Where it goes wrong
-
A break that grows inside the range you always clear
Each month's movement sits within what has been cleared before, so no single month escalates. What is unusual is the three-month sequence, and nobody is reading the sequence.
-
The verification rule on a payment destination
A change to where money goes is governed by a rule. That rule is enforced in the system, performed by a person following a procedure, or neither — and which one it is has a habit of surfacing after the event rather than before it.
-
Contracts nobody reads again after signature
Supplier agreements carry obligations on service levels, sub-contracting, data handling and notification. Reading the contract against the relationship you actually have is a comparison at a volume nobody has staffed.
-
Rate cards applied across a portfolio
Tiers, thresholds and product terms are a reconciliation between what the terms say and what was charged. It gets done periodically, for the same reason as everything else here: volume.
Who owns this
The supervisor's question and the internal one are the same question, asked by different people.
-
Compliance and risk
The answer moves from "controls were operating effectively during the period" to a list: the exceptions found, when each was found, what was decided and by whom — with the ones that could not be judged named as that, rather than folded in as passes.
-
Internal audit and control testing
The design's answer to sampling is that the population tested is the population and the exceptions come back as named items. Reading your stream is Atlas's work; judging each finding it sends is TIDE's.
-
Finance
Where the tier applied and the tier earned differ, and where the accrual differs from what the agreement earns, is stated with the clause it was read from.
-
Operations
A departure from the plan is noticed while a remedy is still available, instead of becoming a recovery action later.
-
Security and IT
The model doing the reading runs where TIDE runs, never at an outside AI service. The records, the files and the credentials stay on the machines your plan names.
What the arrangement changes
- A finding submitted for judgement is graded on a different credential from the one it was submitted with. The two are separate doors: a verdict arriving at the findings door is discarded, and the submitting platform's own secret is refused by name at the verdict door, with the refusal saying why. A submitter cannot grade its own work, and that is decided by what credential was presented rather than by a flag in the message — a flag is something a sender sets, a credential is something a sender either holds or does not.
- Where a finding claims a rate, the counts behind it travel with it and the arithmetic is recomputed rather than accepted. A figure the evidence does not reproduce is a dissent with the arithmetic named. A count divided by nothing is unable to judge, never a rate of zero. Two readings of the same quantity that disagree are unable to judge, never averaged into a third figure neither party observed. A claim carrying a number never reaches a model at all, because asking a model to opine on a number nobody measured produces a fabrication wearing a number.
- A claim in prose is put to the model as one question — does this evidence support this claim — and the model is never told what the submitter concluded. There is no parameter through which it could be told. Where no model is reachable, the answer is unable to judge with that stated as the reason, never agreement by default.
- The record covers what was allowed, what was refused and what failed — the time, the person by verified identity, the action and the outcome — with each entry hashed against the one before it, so a later change to the record shows.
- Specific paths fail closed rather than open: an unconfigured ingest refuses rather than accepting anything, an undeclared identity is refused, and where no secret has been issued the findings door accepts nothing rather than accepting everyone — so a configuration nobody finished cannot quietly become an open one.
How you will know it worked
None of these numbers are ours. They are yours, and most already sit in a report somebody runs monthly. Write down where you stand today before anything starts: the baseline stops being recoverable the moment things improve. The right-hand column says which way each figure should move, and by what mechanism.
| The measure | Your figure today | What moves it |
|---|---|---|
| Time to close an audit or regulatory finding | Your issue register — raised date to closure date, for findings closed in the last four quarters. Split those that waited on evidence from those that waited on a decision. | Down on the evidence-bound half, because the record a closure rests on is written as the work happens rather than reassembled afterwards. |
| Findings open past their due date | The ageing view of that register at last month end: count, original due date, and how many times each item has been extended. | Down, because a control checked on every item can be re-evidenced when the remediation lands instead of waiting for the next testing cycle. |
| Person-days to assemble evidence for one request | Hours booked against your last supervisory request or audit sample, plus the informal time the business spent finding files. Compare your last two cycles. | Down, and the fall is in the finding rather than the reviewing: the reviewer gets the record made at the time of the decision, not a pack built later to describe it. |
| Reconciliation breaks and their ageing | Your month-end break report: open items, value bands, days outstanding, and how many were cleared with no cause recorded. | Fewer aged items, and fewer cleared with nothing written down, because a run of movements that each sit inside the usual range can be raised on the shape of the sequence rather than on any single month. The machine says the sequence is unusual against the available history, cites the series, and files the item for a person, who knows what changed operationally that month and names the cause. |
| Control-testing coverage and exception rate | Your testing plan for the last cycle: per control, the population in scope against the population actually tested, and the exceptions found. | Coverage toward the whole population, and the exception rate becomes a count of named items rather than an estimate drawn from a sample. |
| Time from a control failing to somebody knowing | Your last handful of control failures: from the timestamp of the failing event in the source system to the date it first appears in your issue register. | Down, because that gap is the review interval, and judging each item as it arrives removes the wait rather than shortening it. |
Two things that make the above possible
-
The reasoning goes to no outside model service
The records, the contracts and the working-out stay on the machines TIDE runs on. The judging model is the cluster's own inference plane: there is no other endpoint, no API key, no vendor SDK and no fallback, and no code path that would take an external model plane. So no customer record, no payment instruction and no draft finding is sent anywhere to be read.
- What you would notice
- The length of your own third-party risk and model governance review. What is reviewed is software you run, so there is no external model service to assess and no data-transfer clause to negotiate. When a supervisor asks how a decision was reached, the answer is a record of what was decided, what it cited and who approved it, hash-chained and held by you.
-
Open-web research with no vendor account behind it
Public material — filings, court and registry records, press — is found through a public search endpoint queried from the machines your plan names, and the pages behind the results are then fetched and read by a browser on those same machines rather than summarised for you by a paid research service.
- What you would notice
- Adverse-media and counterparty research, where the question is more sensitive than the answer. The search engine sees the query, as it would from any browser. What does not exist is an account: no API key, no vendor contract, no per-question bill, so no supplier accumulates a history of the names your bank has been asking about, filed under your bank and retained on their terms. That is the part that matters when the name is market-sensitive or the file later becomes evidence.
Who you would be working with
Runink is founder-led. The person in the first meeting is the person who designed the thing being discussed — his track record, what he does here, and the means to check any of it are on the company page.
See whether it fits
Bring one control and the systems it is supposed to live in — the verification rule on payment destinations, a supplier agreement, a rate card. A short session is usually enough to see whether the written rule and the applied rule still match, and whether this arrangement is the right shape for you.
The long version of the arrangement on this page, written jointly with Logical Leap: what a finding contains, the checks it passes before anything reaches a model, how the arithmetic is recomputed, and how a second judgement is formed on a separate credential. Start at the chapter on what judging means. No case studies, no customer names and no return-on-investment figures.