AnantState
Explainable AI

Decisions you can inspect, explain and reproduce.

Every conclusion in AnantState comes with the facts behind it, how much the learned part changed the answer, and who approved the action. Re-run it on the same inputs and you get the same result, so a reviewer can check the work instead of taking it on trust.

In plain termsClick any recommendation and see what it rests on: the events, the forecast, the policy checks and the person who signed.

At a glance

  1. A chain back to the facts

    Each conclusion lists the evidence that supports it and the evidence that cuts against it.

  2. The learned part is measured

    You see how much it moved each answer, and you set how far it is trusted.

  3. Same inputs, same answer

    Any decision can be re-run and compared with the original.

  4. Every action has an owner

    Recommendations, approvals and overrides all carry a name and a reason.

What a decision record shows

An example from the OgMart payments domain. Every recommendation in the product has one of these behind it.

FieldExample
What changedCard-not-present declines at one store group rose for three days, well outside their usual weekly pattern
ForecastIf nothing changes, the shortfall keeps growing over the declared horizon; the record shows the range, not a single figure
Evidence forDecline rate, chargeback feed, terminal firmware version
Evidence againstNetwork status was normal. The two sources disagree, so the case is marked contested
Learned contributionThe learned part moved the forecast by a small, stated amount; the rest came from the declared baseline
Policy checksAction within spending limit; requires approval above the configured value
Recommended actionReview the terminal batch, with its cost and the cost of waiting beside it
Approved byNamed user and time, with any override reason

The same record can be pulled up months later to answer “why did we do that?” without anyone reconstructing it from memory.

One recommendation, and everything behind it

  1. What changedLeads to: Recommended action
  2. Forecast: A range, not one figureLeads to: Recommended action
  3. Evidence forLeads to: Recommended action
  4. Evidence against: Disagreement is keptLeads to: Recommended action
  5. Policy checks: Limits and approval rulesLeads to: Recommended action
  6. Recommended action: With its cost and the cost of waitingLeads to: Approved by
  7. Approved by: Named user, time, any override
Each recommendation carries what changed, the forecast, the evidence for and against, the policy checks and the person who approved it, so the answer to why we did that is a lookup.

Four properties you can test

“Explainable AI” is often a slide. We treat it as four properties a reviewer can check.

  • 1 · State is explicit

    If a conclusion depends on a quantity, that quantity is visible.

    Every entity has a defined set of state dimensions with declared bounds. Nothing important is implicit.

  • 2 · The derivation survives

    What supports it, what could, what is undecided, and how much the sources contradict each other.

    Contradiction is surfaced and kept, because a contested case is the one a person should look at.

  • 3 · Contribution is quantified

    How much did the learned part move the answer, and in which direction?

    It is a bounded, weighted correction to the business’s own dynamics. The trust weight is a configured number per domain, shown per dimension.

  • 4 · Every action is attributable

    Who decided this, and what did they see?

    Recommended actions, applied actions, overrides and gate refusals all carry an actor and a reason. Answering it is a query.

Reproducible does not mean correct

Re-running a decision and getting the same answer proves the computation is repeatable. It does not prove the answer is right. A conclusion can still rest on bad inputs, a wrong assumption or a forecast that misses.

That is why the platform also tests its forecasts against the simple guess that nothing changes, at the distance ahead you choose, and records where it loses. Traceability tells you how a decision was made; the model record tells you how often to trust it. You need both. See the evaluation methodology.

There is a learned component. It is a specialized predictive model that corrects a known behavioral baseline. It is bounded, its influence is visible per dimension, and its error is reported against a baseline you can check.

How this differs from a language model

QuestionAnantState
Does it generate text?No. Conclusions are computed from state, evidence and rules, so there are no prompts and no hallucinated prose.
Is there a token meter?No. Inference cost does not depend on how much you ask.
Does phrasing change the answer?No. The domain definition and the evidence determine it.
Does the same input give the same answer?Yes. The same state and evidence produce the same conclusion.
Does it depend on a third-party model?No. Nothing retires or reprices underneath you.
Nothing here is generated. A conclusion is derived from explicit state and evidence, so it can be traced back to its inputs, reproduced and audited.

Your rules, your thresholds, your actions, your authority

Most platforms ship an inference engine and ask you to accept its judgment. Here, the business defines the reasoning.

The business definesMeaning
The ontologyThe vocabulary of entities, state dimensions and relationships
The dynamicsHow state moves, in the business’s own terms
Thresholds and bandsWhat counts as elevated, for whom, and where the lines sit
The interventionsWhat actions exist, what they cost, what they may affect
The decision horizonHow far ahead the platform must be right
The action frameworkWhich actions need a human and which may be automatic
The escalation policyWhat happens when evidence conflicts or confidence is low

The platform contributes the forward position, the evidence and the discipline. The authority stays with you. Where you want automation you grant it explicitly, per action class, with limits, never as a global switch. See next best action.

What to ask any vendor that says “AI”

Six questions that separate the categories

  1. Is any of this generated text, or is it all derived? If generated, what stops it inventing something?
  2. Show me the chain from a conclusion back to its inputs.
  3. Will the same inputs produce the same conclusion tomorrow?
  4. How much of the answer came from the learned component, and how do I change that weighting?
  5. What happens to my reasoning when the vendor retires a model version?
  6. Where is the token meter? If there is one, you are buying inference.

We will answer all six on your data.

Why this entity: the drivers, in the open

Each conclusion shows the drivers behind it and the events that triggered it. Re-run it and the same drivers appear.

One OgMart entity opened: ranked candidate actions with cost and return, a notice that the evidence is conflicting so autonomous action is held, the top drivers, and the trajectory against the events that triggered it
One entity in depth: what to do, what it costs, why the platform thinks so, and a flag where the evidence disagrees.

Ask the six questions of us first

A technical session is the fastest way to see a derived conclusion traced back to its inputs, on a domain you recognize.