Breaking

Probably Raises $9M, Led by a16z, to Curb AI Errors With Validation Layers

June 16, 2026 at 09:20 EDT

AI Agents
Funding & M&A

Probably, a San Francisco startup that aims to suppress AI "hallucinations" (fabricated facts) using deterministic validation layers, has raised $9 million in a seed round led by Andreessen Horowitz. As reported by TechCrunch on June 16, 2026, the company was founded by Peter Elias. It wraps LLM outputs in a "deterministic validator" to catch factual errors before they reach users.

June 16, 2026 · Probably

$9M to Build an AI That Won't Hallucinate

Probably raised a Seed round led by a16z to wrap a small, local LLM in a deterministic validator — every answer is checked against the data, rejected if it doesn't match, and revised until it does.

$9M

Seed round, led by a16z

99.99%

Targeted accuracy, on par with a deterministic system

Billions

Rows analyzed in natural language, locally

Model class: weaker, but harnessed

Frontier model

full capability

Probably's model

4 classes weaker, runs local

A weaker model still produces correct results — the better the harness, the lower the token cost.

How the validator loop works

LLM answers

reasoning only, in the cloud

→

Validator checks

deterministic, down to the number

→

Pass or revise

failures sent back; artifacts reproducible

Data stays on the device or network. Math is delegated to a local optimization engine — the "data science mech suit."

Built in

Reports only what the data supports
CSV, JSON, Parquet, Snowflake, BigQuery, Postgres, MySQL
No row or size limits; pay-as-you-go or subscription

Still ahead

Public Preview 0.1 — data analysis only
Real-world evaluation and benchmarks pending
Accounting, healthcare and PDFs listed as future plans

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.