Arithmetic, not inference
A score is the sum of the weights of the rules that matched, clamped and banded. Anyone can recompute it by hand from the record — which is the point.
Risk · AI Risk SystemsDecisions are rules,
Decisions are rules,
not predictions.
No model decides whether your money moves. A risk score at BINK is arithmetic over rules a person wrote and anyone can read. There is a language assistant on the platform — it helps analysts understand a case, and it has no authority over a single decision.
DecidesActs on the transaction
- What it is
- A rule engine
- How it works
- Matched weights, summed
- Authority
- Decides, and can act
- Output
- A score, a band, reasons
ExplainsWrites for a person
- What it is
- A language assistant
- How it works
- Reads context, writes prose
- Authority
- None — it cannot act
- Output
- A summary, for a person
The two never meet. The decision services do not import the assistant, so this is a structural separation rather than a policy.
MethodDeterministic rules
ModelNone in the decision path
AssistantExplains, cannot act
JudgementBelongs to an analyst
The lifecycle
Event to judgement.
Where the machine stops and a person starts. Everything up to the decision is deterministic; everything that requires judgement is handed to someone accountable for it.
- EventArrivesA payment, transfer or conversion is about to run.
- FeaturesReadValues computed from the account’s own history.
- RulesAppliedEach rule tests a field against a threshold.
- DecisionRecordedA score, a band and the reasons behind them.
- ReviewBy a personWhere a case is opened, an analyst decides.
The choice
Why rules, deliberately.
A model would be easier to market and harder to defend. In a regulated money flow, the ability to say exactly why a payment was refused is worth more than a marginally better hit rate that nobody can account for.
BINK publishes no accuracy, precision or detection-rate figure. There is no model to measure, and an unverifiable number on a risk page helps nobody.
- It can be explained
- A decision returns the rules that matched, the field each one read and the value it saw. There is no interpretation step between the evidence and the answer.
- It can be corrected
- When a decision is wrong, the fix is a rule — changed by an analyst, effective on the next transaction. There is nothing to retrain and no waiting for a release.
- It can be replayed
- The features are stored with the decision, so the same evaluation can be reconstructed exactly, months later, without needing the model that produced it.
- It cannot drift
- The engine behaves the same way today as it did last quarter unless somebody changed a rule — and that change is a record with an author.
The assistant
What the language model may do.
BINK does run a language model. Being precise about its scope is the difference between a useful tool and an unaccountable one.
- What it can do
- Summarise a case, describe the risk factors on an account, and draft an analysis for a person to read. It is given the same records an analyst can already see.
- What it cannot do
- Score a transaction, change a decision, place or release a hold, or alter a rule. It is not connected to the decision path in any way.
- Who can use it
- Access is by role. The analyst, compliance, treasury and merchant views are separate permissions, and a user sees only what their role allows.
- Availability
- It requires a configured language-model provider. Where none is configured, the assistant is simply unavailable — there is no fallback that fabricates an answer.
- What it records
- Conversations and generated reports are stored, attributable and auditable, in the same way any other analyst action is.
What this gives you
Answerable, every time.
The properties that matter when a customer, an auditor or a regulator asks why.
Reasons, not confidences
Each contributing rule returns a category, a severity and the evidence it acted on. There is no probability to interpret and no ranking to second-guess.
Change is a record
Rules are data. Adding, retuning or disabling one is an attributable change, so the question “why did this start declining last week” has an answer.
People make the calls
Where a decision needs judgement it becomes a case, assigned to an analyst, moving through a state machine that refuses transitions it does not allow.
Role-scoped assistance
The assistant’s views are permissioned separately — a merchant-support view and a compliance view are different roles, not the same tool with a different prompt.
No autonomous action
Nothing generated by a language model moves money, blocks a payment or releases a hold. Its output is written for a person to act on.
Stated plainly
What BINK does not claim.
- What is rules
- Everything that decides. Scoring, banding, the action chosen and whether it is applied are all deterministic rule evaluation — no model, no training, no inference at any point in the decision path.
- What is AI
- Everything that explains. A language model with role-scoped personas summarises cases, describes risk factors on an account and drafts analyses for a person to read. That is its entire scope.
- Terminology
- BINK does not call its risk engine artificial intelligence or machine learning, because it is neither. Describing a rule engine as a model would misrepresent both how it works and how it can be challenged.
- No model versions
- There are no trained models, so there is nothing to version, retrain or backtest. What is versioned is the rule set.
- No adaptive scoring
- The engine does not learn from outcomes or adjust its own weights. Any change to how it scores is made by a person and recorded as such.
- No published latency
- BINK does not publish a scoring latency, throughput or accuracy figure. None is measured, and an unmeasured number on a risk page is worse than no number.
- Separation
- The decision services import no language-model service. The separation is structural, not a policy that could be relaxed by configuration.
- Where AI is used
- Explaining cases, drafting analyses and answering questions about records a user is already entitled to see. That is the whole scope.
Integration
What this looks like to build against.
Determinism is not only a governance property — it changes what your own code has to do about a risk answer.
- No score to interpret
- You receive a decision and the reasons behind it, not a probability to threshold on your side. There is no confidence value for an integrator to second-guess.
- Stable by default
- The engine behaves the same way tomorrow as today unless a person changed a rule. Nothing adjusts itself between releases.
- Reproducible on demand
- Because the features are stored beside the decision, the same evaluation can be re-derived later without the system that produced it being in the same state.
- The assistant is separate
- Anything generated by the language model reaches a person, never an endpoint. There is no path by which its output changes a decision or moves money.
In practice
What explainability buys.
Explain a refusal to a customer
Give a support agent the rules that fired and what each one saw, rather than an opaque score they have to apologise for.
Answer a regulator
Show the decision, the evidence it rested on and the authored change history of the rule that produced it.
Brief an analyst faster
Let the assistant summarise a case and its risk factors from records the analyst already has access to, so the reading is quicker and the deciding stays human.
Change behaviour deliberately
Adjust a rule and know exactly what will differ, instead of retraining and hoping the difference is the one you wanted.
Related
Where this sits.
Risk and security
Every surface.
Risk decisions
Defensible beats clever.
Every refusal has a reason, every reason has evidence, and every rule has an author.