Making a consequential number traceable.
Models increasingly propose the numbers that decisions rest on. An enterprise needs to know which parts of such a number were measured, which were assumed, and what the answer becomes when the assumptions move.
AnantCausal is an AnantHQ product for causal decision modeling and decision records.
What we are building, and why
Our thesis
Evidence and assumption must not average into one number.
A count fitted from ten thousand rows and a value asserted in a workshop are different kinds of claim. Once they are blended into a single score, nothing downstream can separate them again, and the guess borrows the credibility of the count. Provenance therefore attaches to each probability row rather than to the model, because the average of three kinds of support is the one thing none of them is.
What we are building
An inspectable model, and the record of what was decided with it.
The graph, the probabilities and the values are three separate layers, kept separate so that replacing an elicited row with a fitted one is a visible event rather than a silent improvement. Above them sits the record: what was approved against which stored analysis, what was assigned versus what was reached, and the measured outcome reconciled against the design-time claim.
Why this matters
A number nobody can trace back is not a number anyone can defend.
The question that decides whether a model is usable is rarely whether it is accurate. It is whether, eighteen months later, somebody can say what it rested on and whether that still holds. Answering no is a result here rather than an error — an unanswerable question returns the reason and the failed routes, and never a number shaped like an answer.
Our discipline
A product that asks customers to show their assumptions should be held to the same standard. Each of these is enforced somewhere in the product, and each names the enforcement so it can be checked.
Every number carries its origin
Provenance attaches to a probability row, not to a model. This is why a language model's proposals are tagged rather than blended in, and why a table built from your own counts never carries an elicited tag: the counts are the evidence, and asking a model to invent a probability when the numbers are already in the request replaces evidence with a guess.
A refusal is a result
When the structure does not identify a question, the answer is no with the reason attached and the failed routes named. It is stored, replayable and reportable, and it never arrives accompanied by a number shaped like an answer. Four of the industry pages on this site publish a refusal, because a demonstration with none is a demonstration of a different product.
The record outlives the model
What was decided, against which stored analysis, what was assigned versus what was reached, and what was measured against what was claimed — kept separately, so that revising the model does not revise the history of a decision already taken.
Limitations are stated where they are hit
Every capability on this site carries the specific thing it does not establish, in the same block rather than in a footnote. The evidence page publishes its worst case beside its mean and both of its denominators. The manufacturing vertical states that a missed edge means the effect was computed without adjusting for a variable it should have been adjusted for.
The last two commitments are checkable rather than asserted: the evidence page carries the benchmark's worst case beside its mean and both of its denominators, and the manufacturing vertical states what a missed edge did to the answer on that page.
How it is built
A claim about deployment and dependency footprint, which is the kind of thing a security reviewer needs in order to plan rather than take on trust.
- Self-hosted: a service and a database that run on your hardware, with no telemetry to us
- Ten dependencies and 47 installed packages, of which 14 carry a compiled extension
- No queue broker, no vector store, and no third-party analytics on the application origin
- The model transport can be pointed at a model running locally, and the shipped benchmark was run that way
AnantHQ, Inc. · AnantCausal