AnantState
Resources · Demo guide

Demo guide: 100 stores, 15 departments, one live company.

OgMart is a fictional 100-store, 1–2 billion USD mini-Walmart running as a single live operation across merchandising, supply chain, distribution, inventory, stores, e-commerce, customers, pricing, payments, finance, treasury, workforce, facilities, IT and compliance.

Book a working sessionRead the evaluation methodology

It is fictional, and we label it as fictional on every screen. The mechanics are real; the company is not. We do not present synthetic data as customer data, and we ask that you do not either when you share screenshots internally.

Not a tour. A working session.

We stand up the reference world, then spend most of the time on your domain: what the entities are, what state moves, how far ahead you need to be right, and what the model record says when we evaluate against a naive baseline. Bring one domain, the one your team already argues about weekly.

The route through the world, in order

What to look at, and what to look for.

  1. The queue · 30 seconds

    Entities ranked by value at stake, with band, trend and time at risk. Look for: is the ordering defensible? Could you hand this to a team lead and have them work top-down?

  2. One entity, in depth · 2 minutes

    A single entity’s state trajectory against the events acting on it, plus its evidence panel: belief, plausibility, uncertainty and conflict. Look for: does conflict ever go high, and does the platform treat that as interesting rather than smoothing it away?

  3. A priced decision · 2 minutes

    An intervention with its cost, its modeled effect on the primary dimension, and the cost of waiting beside it. Look for: is the arithmetic inspectable? Can you see the assumptions? Does “do nothing” now have a price?

  4. Counterfactual replay · 2 minutes

    Run an entity’s real history again with an intervention applied at a chosen point. Two trajectories, side by side. Look for: does this replace a post-incident argument with a measurement?

  5. The what-if, in your browser · 2 minutes

    Fork the live state, change the lever, watch it recompute locally. Look for: no round trip, no spinner. Then ask the hard question: what if it is wrong here?

  6. The model record · the one that matters

    Per-horizon error against persistence, the furthest horizon at which the model still wins, the verdict, and the corpus provenance. Look for the horizons where it loses. If a vendor will not show you this, that is your answer.

The OgMart payments domain as a ranked list of entities, each with its value pressure, value at stake and value band, ordered by pressureFictional reference world: OgMart
The queue: every entity ranked, with the money at stake beside it, so a team can work it top-down.

Judge it, do not admire it

SignalWhat it means
The queue ordering is defensible without explanationRanking is doing real work
Conflict goes high on some casesThe evidence layer is not decorative
An intervention’s cost changes the recommendationPricing is wired into the decision, not displayed
The model record contains a failureThe evaluation is real
The what-if is instant and localSame engine, not a demo build
Someone says “we would not use it that way”They are translating it to their world, which is the best sign there is

The five questions to ask us

Bring these. We will answer them live, on the reference world and then on your data.

  1. Show me the model record. Which horizons beat persistence, and which do not?
  2. Which horizons were not measured?
  3. What is the corpus provenance? How much training data is unattributed?
  4. Show me a refused promotion and the recorded reason.
  5. What is the furthest horizon you would stand behind for my domain, on my data?

Practical notes

  • A session runs about 45 minutes, most of it on your domain.
  • Bring a sample of real events if you can. A month is plenty, and it turns the session from a demo into an evaluation.
  • Bring the person who owns the decision, not only the person who owns the data. The most revealing question is “how far ahead do you need to be right?”, and it is usually answered by the operator, not the analyst.
  • Screenshots are welcome. Please keep the fictional-world labeling in frame.

One entity, in depth

Step two of the route: the trajectory against its events, the drivers, and the evidence panel.

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 itFictional reference world: OgMart
One entity in depth: what to do, what it costs, why the platform thinks so, and a flag where the evidence disagrees.

A priced decision

Step three: the effect of acting against the cost of waiting.

Impact analysis for an OgMart entity: the forward path with and without the action, and the net value of actingFictional reference world: OgMart
Impact analysis prices the action against doing nothing, and says plainly when acting costs more than it saves.

Counterfactual replay

Step four: the same entity run forward with and without action.

Counterfactual theater: two forward projections of one OgMart entity, one with no action and one with early action, with the dollars at stake in eachFictional reference world: OgMart
Counterfactual: the same entity run forward twice, once ignored and once acted on early.

The model comparison

Step six: look for the models that lose.

Prediction error by model on data the model never trained on, with a dashed baseline line: models that beat the baseline fall below it and those that do not rise above itFictional reference world: OgMart
Models are judged against a baseline on data they never trained on. The ones that lose are shown, not hidden.

Book the session

Name the domain and the decision you wish you could make earlier.

Book a working sessionRead the evaluation methodology

Continue exploring