Verification models
Judges under two billion parameters, trained on verdicts rather than prose. They score grounding, citations, and constraints on device.
Deep-tech research lab · Grounding verification
Intelligence
you can verify.
Fahrenheit Research builds grounding verifiers: small judge models that sit beside a domain model on your hardware and prove every answer before it ships.
One category. One job. The verdict.
Pass -- / 12 · one subject
Sweep
Live inspection
Click to run the next check
Click to run the next check
Lens 06 · Binary · The problem
Every domain can generate. Almost none can check.
Fluent answers are everywhere. Proof is not. The gap between the two is where adoption stalls, budgets leak, and value pools.
001Fluent answers shipped without proof of groundingrisk
002Hallucinations found by customers, not by systemslate
003Every check routed to a hosted frontier model∞ $
004Data that cannot leave the device for reviewcrit
005Human review as the only quality gateslow
006No local model whose whole job is the verdictgap
Observed across the fieldVerifier-poor domains show no measurable progress
Lens 02 · Quadtree · What the lab builds
One lab. Two model lines.
We build the author and the examiner: small language models that know the domain, thin today and industry-specific next, and verification models that hold every answer to ground truth.
Thin Language Models
The category we coined. Frontier-grade judgment for one domain, on a laptop.
Edge-native runtime
Weights and checks that run on local silicon, with no cloud round trip.
The Verifier Boundary
Our thesis: automatic verification, not data or compute, decides the winners.
Deep data architecture
Pipelines that surface signal, never blind scraped.
Custom model engagements
Select projects where we build the domain-specific model and its verifier for your stack.
Lens 04 · Halftone · Model 001
Rankine, the grounding verifier
The first Fahrenheit model. A small judge that reads an answer against its evidence and returns a verdict on device, in milliseconds, with no cloud in the loop. It does not generate. It checks.
In training · lab hardware
FR-Rankine
Pair it with any small language model or FR model: the model writes, Rankine grounds every claim in evidence before the answer ships.
Class
Verifier
Job
Grounding
Run
On-device
Pairs with
Any SLM or FR model
001
Claim-level verdicts
Every claim is scored against evidence; a weak sentence cannot hide in a fluent paragraph.
002
Evidence spans
Each verdict cites the exact passage, so review takes seconds instead of a re-read.
003
Calibrated abstention
Thin evidence returns "unverifiable" instead of a guess.
Verdicts in milliseconds, fully offline. Named for the Rankine scale: absolute measurement, in Fahrenheit degrees.
Lens 03 · Circle pack · The thesis
Progress follows verification.
Our meta-analysis across five model lines: task classes with automatic verifiers gain roughly twelve points per quarter. Classes without them show no computable trajectory at all.
Lens 05 · Trace · Research
The lab publishes first, then ships what survives review.
Pre-registered protocols, public predictions, and results that survive hostile review. Every thread feeds the verifier line.
The Verifier Boundary
A pre-registered protocol and meta-analysis on why automatic verification decides domain AI.
Grounding verification
Verdicts with evidence spans, built for retrieval-heavy work.
Thin Language Models
Sub-2B domain models that run locally.
Compressed intelligence
10× smaller, under 3% accuracy loss.
Lens 07 · Dither · Principles
What we refuse. What we own.
One philosophy, drawn in two tones.
Refuse
- ✕We do not license breakthroughs.
- ✕We do not outsource thinking.
- ✕We do not rent intelligence.
Own
- ✓Every model we specialize.
- ✓Every dataset we build.
- ✓Every product we ship.
Lens 11 · Quantize · Contact
Work with the lab.
Researchers, enterprises, and builders who need answers they can prove: reach the lab directly.
Custom engagements: your domain, your data, your hardware.
Lens 12 · Moire · Before you reach out
Questions we get asked
A small model whose only job is to check another model's answer: grounding, citations, constraints. It returns a verdict with evidence, not more prose.
Our meta-analysis across five model lines finds that task classes with automatic verifiers improve roughly twelve points per quarter, while verifier-poor classes show no computable trajectory. Scale generates. Verification compounds.
The first Fahrenheit model: a grounding verifier that runs on device and pairs with any small language model or FR model. It is in training now.
Yes. The lab takes on select engagements to build custom domain-specific models, from data architecture to a deployed model with its verifier. Write to research@f-r.co.