Question 1

Is this a causal inference library?

Accepted Answer

No. A causal inference library estimates effects from data — doubly robust estimators, causal forests, propensity methods. This does not estimate from data at all. What it does is build the model those libraries need and hand it over with the caveats attached.

Question 2

Then what is it?

Accepted Answer

A system for making a decision defensible before it is taken and reconstructable afterwards. It turns what an organization already knows into an inspectable model, tells you whether your question is answerable at all, and keeps a record of what was decided and what happened.

Question 3

Does it estimate effects from observational data?

Accepted Answer

No, and this is the distinction the whole product rests on. A heterogeneous treatment effect learned from observations carries sampling error; an exact elimination result is arithmetic on a stated model. Those are different claims, and this product keeps them in different response shapes rather than behind one field.

Question 4

Is that range a confidence interval?

Accepted Answer

No, and the response says so rather than leaving it to the reader. The band is a sensitivity envelope: the same effect recomputed with each calibrated row moved to the ends of its interval. It names its own basis — sensitivity_envelope, ensemble_envelope, or combined — and it never means sampling error, because an elicited model has no sample to draw error from.

Question 5

What happens when it cannot answer?

Accepted Answer

It says so, names the route it tried, and gives the reason. An unidentified question returns identification not established with a sentence explaining why — usually that an unmeasured common cause of the treatment and the outcome would defeat any adjustment, and that there is no front-door route because the treatment has no child leading to the outcome.

Question 6

How do I know the answer is not just a plausible guess?

Accepted Answer

You do not, from this product alone, and any answer that said otherwise would be selling validation. What the refutation battery does is try to break the result three ways and report each attempt individually — including inconclusive, which is never counted as a pass. A passing battery means the answer survived those attacks. Attacks a model cannot detect are exactly the attacks a model cannot detect.

Question 7

What if there is a confounder we did not measure?

Accepted Answer

Then adjustment cannot help, no front-door route may exist, and the honest answer is that the question is not identified from observational data without further assumptions. A bigger model does not fix this, and an estimator does not either — it is a property of what was not measured, not of the method.

Question 8

Where does the causal graph come from?

Accepted Answer

From content you already have: a description of the process, a document, a table, a stored dataset, or a language model reading a document. Every edge records its source, and edges the people involved disagree about can be flagged contested so they get reviewed rather than silently asserted.

Question 9

How large a model does it handle?

Accepted Answer

Up to 24 variables, with a cap of 2,048 cells in any one conditional table. Exceeding either is refused rather than degraded — with a message naming what made the question too wide, so the caller can narrow it rather than guess at a timeout.

Question 10

My outcome is continuous. Is that a problem?

Accepted Answer

Yes, and it is a modeling limitation rather than a presentational one. Variables here are discrete, so a continuous outcome reaches the model through binning, which is lossy and labeled on the row rather than hidden. Where binning would destroy the question, this is the wrong tool.

Question 11

Where does the AI come in?

Accepted Answer

In exactly two places, and neither is where the number is computed. A language model can read a document to propose structure, and it can propose probabilities where no counts exist. Both proposals are tagged as elicited — they are claims, not measurements — and the arithmetic that produces an answer calls no model at all.

Question 12

What exactly is kept about a decision?

Accepted Answer

Four things, stored separately: the question as it was asked, the analysis it was approved against, what was assigned versus who was actually reached, and the measured outcome reconciled against the design-time claim. The question is stored rather than a fingerprint of it, because a digest cannot be replayed.

Question 13

Can I find out whether an old answer still holds?

Accepted Answer

Yes for effects. Re-asking a stored question classifies the result as identical, drifted, or reidentified — the last meaning the question now identifies differently, which is a different failure from a number moving, because a drift can be a model update while a reidentification means the question is no longer answerable the same way.

Question 14

What leaves our environment?

Accepted Answer

Nothing, in a self-hosted deployment — which is the deployment this product is built for. The service is one API process and one Postgres instance that you run. Where a hosted model is used to read documents, the document text and the resulting proposals are recorded on the model as elicited, so it is visible which parts of a model were informed by one.

Question 15

Do we need a data scientist to run this?

Accepted Answer

Not to build the model, and yes to judge the answer. Eliciting the structure and reading the control advice does not require one; deciding whether a given adjustment set is substantively right, and whether the refusal on a question you care about is worth working around, does.

Question 16

How is it licensed and priced?

Accepted Answer

Described in a conversation rather than on a page, because the answer depends on deployment and on how many decisions are in scope. What is fixed is the shape: self-hosted, no per-query metering that would make a refusal cost the same as an answer, and no charge for the record.

Question 17

How do I know this is not for us?

Accepted Answer

Four tests, and failing any of them means this is the wrong purchase. There has to be a decision rather than a question; the structure has to be knowable from documents, experts or data; the outcome has to be discrete or plausibly binned; and somebody has to be held to the number afterwards.

Question 18

What is decision intelligence?

Accepted Answer

The practice of improving how an organization makes decisions by combining data, models and human judgment, and by recording how each decision was made. AnantCausal covers the causal-modeling part: it builds a readable model, reports whether your question can be answered from it, stress-tests the answer and keeps the decision on the record.

Question 19

How is this different from A/B testing?

Accepted Answer

An A/B test measures what happened when you changed something for a randomly chosen group. It is the best evidence you can get, but it is slow and many decisions cannot be tested first. A causal model works out what is likely to happen if you act, before you spend, and helps decide which interventions are worth testing. It does not replace experiments.

Question 20

Can this replace our dashboards or BI tools?

Accepted Answer

No, and it is not meant to. Dashboards show what happened. AnantCausal answers a different question, what to change, using a model of how the system responds. Teams typically keep their dashboards and use the causal model for the decisions where the lever is unclear.

Questions the rest of the site should have raised.