AI Architecture

IO's intelligence is not the property of any single model — it's the property of how models are composed. We don't fine-tune a secret model or chase a private benchmark. We orchestrate the best open-weight models in the world and make the orchestration itself verifiable.

Most providers give you one model and one answer. IO gives you a pipeline.

The AI architecture treats a prompt the way a build system treats source code: it is classified, routed, optionally fanned out across specialists, and reduced to a single answer — while every stage emits a signed receipt. Orchestration doesn't sit beside the proof layer. It produces it.

The Orchestration Flow

Execution Modes

One prompt, four ways to dispatch it. The mode is the unit of orchestration.

Intent Router

A lightweight classifier (7–14B) reads the prompt and decides, in under three seconds, what kind of work it is — general reasoning, code, math, multilingual, or long-context — and which execution mode applies.

The router never sees who you are. IP addresses and headers are stripped at the edge before the prompt reaches it. Its output is a routing decision and a confidence score — never a stored profile.

Broadcast — Parallel Fan-Out

In Broadcast and Diffuse, the prompt is dispatched to all open-weight models at once. Because the models genuinely disagree, the disagreement is itself signal: where they converge, confidence is high; where they diverge, you see the seam instead of a laundered consensus.

Each branch runs in an isolated context and returns its own answer, receipt, and self-destruct timer.

Diffuse — Synthesis with Attribution

Diffuse adds a reducer pass. A synthesis model reads the parallel answers and composes a single response — but unlike a black-box ensemble, it preserves provenance. An attribution panel names which model contributed which insight.

Synthesis is additive (merge complementary reasoning), not a majority vote — so a lone correct answer isn't outvoted by three plausible wrong ones. The reducer's own pass is itself receipted.

Fork & Merge — Branching Reasoning

A conversation is a tree, not a line. Fork splits a thread into independent branches — each with its own context window, receipt chain, and timer — so you can explore competing approaches in parallel and merge the winning path back.

These are git semantics for thought: cheap branches, explicit merges, and no history left behind, because there is no history to leave.

Redaction Pipeline

Before any model observes input, a redaction pass strips wallets, keys, emails, and identifiers — substituting typed placeholders ([REDACTED]) that preserve grammatical structure so the model still reasons correctly.

The receipt records which classes were redacted without recording the values. Privacy becomes an auditable event, not a promise.

The Receipt Spine

Every stage above — route, broadcast branch, synthesis, redaction — is bound into the response's io_receipt_v1. The AI architecture and the verification architecture are the same architecture:

• Misattribution is detectable — the receipt names the model and policy that produced the answer.
• Redaction is auditable — the receipt proves which classes were stripped.
• The chain is tamper-evident — receipts link into a Merkle chain; breaking one link invalidates the whole.

Intelligence and attestation are produced in a single pass. You never have to choose between a good answer and a provable one.

Next Steps