# Neither — reconstruct what happened, not just related documents

> Your AI retrieves passages. Neither reconstructs what happened.

Relevant isn't the same as right — when ownership, sequence, contradictions, and gaps decide the answer, a confident wrong answer is worse than none. Neither is the **Company Brain** — a governed context graph humans use through an AI Chief of Staff and agents query through the Context API.

## What Neither does

RAG optimizes relevance. Neither optimizes reconstruction. Agent teams query the Context API. Operators use the AI Chief of Staff as the human on-ramp. Both surfaces read the same context graph — who blocked the renewal, which meeting changed the timeline, and what evidence contradicts the leading explanation.

## Retrieval vs reconstruction

**RAG retrieves matching passages. Neither keeps the thread.**

**Question:** Who owns the security review, and what's it blocking?

**Search / RAG:** Returns documents that match "security review" — not who owns it or what it blocks.

**Neither:** Maya Chen (Security Lead, Trident Corp) — last mentioned Feb 8, no reply to Ahmed's follow-up. Blocks Milestone 3 of the Trident contract, 5 days overdue.

Reconstruct a thread in ~60 seconds — nothing connected: https://www.neither.online/context-api-demo

## Why reconstruction beats retrieval

1. **Consistent answers** — every agent reads the same thread, not a different guess per workflow.
2. **Provable context**: citations, provenance, and tenant isolation you can show security and audit.
3. **Trustworthy action**: agents act on maintained people, decisions, and dependencies without stale passages driving the next step.

## How it works

1. Connect sources — email, chat, docs, meetings, CRM, tickets.
2. Maintain the context graph — people, projects, decisions, dependencies, evidence.
3. Agents call the Context API — structured context with citations, freshness, and workspace scope.

## Operating checks (honest evidence)

**Headline:** Tenant isolation — zero cases of one workspace's context leaking into another, across 126 adversarial tests.

We publish honest rates with visible sample sizes. These are operating checks on a production-like workspace, not a marketing scoreboard.