Introducing the Reasoning Graph: An AI Twin That Can Show Its Work

The problem with "an AI that thinks like you"

Every AI twin product makes the same promise: it answers the way you would. Almost all of them keep it the same way, by pasting a description of you into a system prompt. Your job title, your expertise, a few values, a tone. Then a large language model does what large language models do: it produces a plausible, well-structured answer that could have come from any competent consultant.

We know, because we audited our own product against that standard. We asked one of our test twins how its owner handles a product team that keeps changing requirements mid-sprint. It said the topic was "not really my lane" and that it focused on pricing and onboarding. Those were the focus areas in its configuration. It had answered from its settings, not from its owner's judgement.

That is the gap between describing a person and modelling how they decide. The reasoning graph closes it.

What the reasoning graph is

It is one graph per person, built only from things you actually said or wrote.

Evidence is the raw material: your answers to calibration scenarios, corrections you make when a twin gets you wrong, things you tell your own twin about how you would decide, and views you state in documents you wrote. Every piece is stored verbatim and dated. Nothing a visitor says to your twin ever enters it.

Principles are what Imora distils from the evidence. Not "Alex is cautious", which is a personality label, but "Alex ships the reversible version now and validates in parallel; slows down when the decision is irreversible or trust is at stake." A rule, with the conditions it applies under, the exceptions, and links back to the exact answers that support it.

Links and themes give it structure. Related ideas connect, the concepts you keep returning to surface on their own, and when two principles pull in different directions the tension stays visible instead of being averaged into a single number.

Growth is built in. Every calibration session adds answers aimed at the gaps. Nothing is forgotten, and you can replay how it grew.

What your twin does with it

When someone asks your twin a question, it reasons from your principles and the evidence behind them, not from a description of you. Where your evidence is strong it commits. Where it is thin it says so, and still gives your most likely answer rather than inventing a preference or refusing.

Then it shows its work. Under every reply there is a "Why this answer" panel with the principles and evidence it used. Anyone talking to your twin sees whose reasoning graph it answers from and which lens it is wearing.

Twins are lenses, not silos

The graph belongs to you, not to a twin. A board twin, a mentor twin and a pricing twin are three lenses over the same judgement, each with its own audience, focus and tone. Configuration shapes how a twin presents. It never limits what you have a view on, and it never outranks your evidence.

For teams this matters more. Five senior people get five graphs and can run ten subject-specific twins between them, and every answer is traceable to exactly one person's reasoning.

What we measured

We ran 36 business scenarios across engineering, product, commercial, leadership, security, incidents and strategy through the old pipeline and the new one, on the same model, and had a stronger model judge every answer against the owner's own calibration answers.

The old pipeline produced generic consultant answers in half of the scenarios and, with real chat history attached, disclaimed the topic or slipped into the third person in half of the samples. The new pipeline answers in the first person every time, never disclaims a topic, and scores higher on "is this the answer the owner would most likely give" with only five pieces of evidence behind it. Those five pieces are the point: the graph compounds. Ten answers give a sketch. A hundred give a map nobody else can reproduce.

Why this is hard to copy

Not because the code is secret. Because the graph is made of your decisions, accumulated over time, with provenance. A competitor can copy a prompt in an afternoon. They cannot copy six months of your calibrated judgement, and neither can a model that has never met you.

Try it

Every Imora account now has a My Reasoning page. Run a calibration session, watch the first principles appear, then ask your twin something you have never been asked and open "Why this answer". If it gets you wrong, correct it. The correction outranks everything else.