> ## Documentation Index
> Fetch the complete documentation index at: https://docs.platform.nora.my/llms.txt
> Use this file to discover all available pages before exploring further.

# Analysis sessions

> Interactive reasoning over the graph — hypothesize, test, iterate.

The graph isn't just a static verification tool. You can **analyze** it — ask questions, run counterfactuals, and get answers grounded in the causal structure.

## Opening an analysis session

From a graph, click **New analysis**. A side panel opens with a chat-style interface. Ask a question in plain language:

> "If we cut latency in half, what would happen to churn rate?"

The session engine walks the graph and returns an answer with:

* A **reasoning chain** showing which edges were followed.
* **Assumptions** it made (e.g., "assuming linear response").
* **Sensitivity** — how the answer changes if you tweak inputs.
* **Confidence** — a rating based on edge strengths and data.

## What analysis can answer

* **Predictive**: "What happens to Y if X changes?"
* **Explanatory**: "Why did Y go up last quarter?" (given input observations)
* **Diagnostic**: "What upstream variable is most likely responsible for this anomaly in Y?"
* **Counterfactual**: "What would Y have been if X had stayed constant?"
* **Do-operator**: "If we intervene on X to force it to a value, what changes downstream?" (Different from observation — accounts for the fact that intervention breaks other causal links.)

## Session state

Each session has persistent state:

* **Observations** — variable values you tell it.
* **Interventions** — variables you're pretending to control.
* **History** — every prior question and answer.

State propagates: answers reflect everything you've told the session so far.

Reset with **New session**, or fork off from an existing session to explore a variant scenario.

## Attaching data

Some analyses need real numbers, not just structure. Attach a dataset to the session:

* Sensitivity of a prediction to input distributions.
* Fitting edge strengths from observations.
* Grounding counterfactuals in historical data.

The dataset is a CSV or a query against a connected database. Nora joins it to variables by name — same name in both places means "these are the same thing."

## Saving analyses

Save any session as a **named analysis**. Team members can open it, see the reasoning chain, and continue from your point. Great for documenting a decision:

> "Q3 pricing decision" — saved analysis showing why we set the price at \$19.

Saved analyses show up in the graph's sidebar and are exportable as PDF.

## Sharing findings

* **Export as memo** — one-page PDF with the question, answer, reasoning chain, and evidence.
* **Copy citation** — a link to the specific session state (view-only for others).
* **Post to Slack** — shares a preview card with the key finding.

## Difference from an Agent chat

An Agent chat can also answer domain questions — but it uses retrieval and inference, not causal walking. Analysis sessions:

* Trust the graph structure over anything the model wants to say.
* Show explicit reasoning steps you can audit.
* Are reproducible — the same session state always yields the same answer.

Use Agent chat for open-ended questions. Use analysis sessions when the question maps to your graph and you need auditable reasoning.