Anchor reads every message your AI writes, catches anything that breaks a regulation or your company policy, and fixes it before it goes out. A small model trained for that one job, for a cent a message, in a fraction of the time a frontier model takes.
At a glance
Model:
Anchor 3.0, in mini, flagship and max sizes
Design:
A small language model trained on the regulations, post-trained from Gemma E4B, with a deterministic rules engine beside it
Runtime:
A verdict, the rule it applied, the exact passage and a verified rewrite
Works on:
Agents, email, chat, voice, marketing, documents, APIs
Modality:
Text
Runs:
Our cloud or your VPC, one L40S GPU
Speed:
1.5 s per message on the API, under 100 ms self-hosted, fast enough to check every message instead of a sample
Accuracy:
Catches 95.5% of violations on a human-labeled FINRA set, scored against human compliance reviewers. The best frontier model: 92.9%
Price:
$0.01 per message, flat. Rewrite and verification included, output tokens never billed
Enforcement
Every message in.
Every message out, compliant.
Messages from your AI flow through Anchor. Clean ones go straight out. Anything that breaks a rule is rewritten, verified and delivered compliant.
Messages IN
DELIVERED COMPLIANT
ANCHOR 3.0 · 95.5% RECALL
What comes back
The verdict, the rule,
the passage and the fix.
Every response names what it found and what it changed, so compliance, AI and security teams can read it, and it is logged as evidence.
Original AI Output
Composed {
"This fund is a safe way to increase returns and is a strong fit for your portfolio."
}
Anchor Verdict
Verdict:
REWRITE
Rule Cited:
FINRA Rule 2210
Policy Match:
Promissory performance language
Reason:
The output implies safety, guaranteed suitability, and future performance.
ZeroDrift Compliant Rewrite
Sent {
// Rewritten
"This fund may be appropriate for certain investors depending on their objectives, risk tolerance, and time horizon. Past performance does not guarantee future results."
}
Why its own model
General models reason. Enforcement models decide.
Frontier models are generalists handed a rulebook in the prompt. Anchor learned the rulebook, so it answers the same way every time, in one pass, at the speed of the message.
The rules live in the model.
Regulations are learned in training, so nothing is pasted into a prompt and every message is read the same way.
It returns a decision.
The verdict, the rule, the exact passage and a compliant rewrite, in one pass.
Every fix is verified.
Anchor re-checks its own rewrite against the same rules before it ships, so a fix cannot introduce a new violation.
Your policies, enforced too.
Train your rules into Anchor 3.0 as an adapter, or upload policy documents on Anchor 3.0-max with nothing to train.
Benchmark
Measured against seven
frontier models.
Scored against Claude, GPT and Gemini models on a human-labeled FINRA set built with Surge AI. Same rulebook for every model.
95.5%
RECALL
Share of real violations caught, scored against human compliance reviewers. The best frontier model catches 92.9%.
34x
FASTER
Up to 34x faster than the four frontier models closest to it. 1.5 s per message on the API, under 100 ms self-hosted in your VPC.
37%
FEWER MISSES
Than the closest frontier model. For every 100 violations it let through, Anchor caught 37 of them.
12x
CHEAPER
Up to 12x cheaper than those same four models. $0.01 flat per enforcement with rewrite and verification, and no output-token charge.
FINRA Communications Benchmark, 150 human-labeled tasks, Surge AI, September 2026. Results are self-reported and an independent run is in progress.
Models
One rulebook,
three sizes.
Three sizes, the same rules in every one. Pick by how much you send and how deep each check needs to go.
CLASSIFY AND ENFORCE
//A
Anchor 3.0-mini
→
Classifies the whole message and returns one verdict.
Built to screen very high volumes.
Base model:
Gemma E4B
Unit of work:
Whole message
Rewrites:
No
Custom policies:
No
Runs on:
L40S
$0.002
PER
ENFORCEMENT
REWRITE AND ENFORCE
//B
Anchor 3.0
→
Finds the exact passage that breaks a rule, rewrites it and verifies the fix. Learns your policies as a trained adapter.
Base model:
Gemma E4B
Unit of work:
Per line
Rewrites:
Yes, verified
Custom policies:
Trained LoRA adapter
Runs on:
L40S
$0.01
PER
ENFORCEMENT
ENFORCED REASONING
//C
Anchor 3.0-max
→
Reasons through the most sensitive documents. Upload policy documents and they are enforced at call time, nothing to train.
Base model:
Qwen3.8-27B
Unit of work:
Per line, with reasoning
Rewrites:
Yes, verified
Custom policies:
Uploaded documents
Runs on:
H100
$0.05
PER
ENFORCEMENT
One enforcement is one message up to 2,000 tokens, with rewrite and verification included. Output tokens are never billed. New accounts start with $10 in free credits.
Regulations
Every regulation,
built into the weights.
Thirteen regulation areas are learned in training and ready on day one. Activate the ones you need in Policy and add your own rules on top.
→
US Securities
SEC · FINRA · REG BI
→
US Banking
and Lending
CFPB · UDAAP · FAIR LENDING
→
US Insurance
NAIC · STATE AI BULLETINS
→
US Futures
and FX
CFTC · NFA
→
US Real Estate
and Fair Housing
FAIR HOUSING ACT · RESPA
→
US Marketing
Channels
TCPA · CAN-SPAM · FTC
→
EU Promotions
MIFID II · PRIIPS · GDPR CONSENT
→
UK Promotions
FCA · CONSUMER DUTY · COBS 4
→
APAC and
Canada Promotions
CSA · CIRO · MAS · ASIC · SFC
→
Digital Assets
CRYPTO PROMOTIONS · RISK WARNINGS
→
Cyber and Privacy
GDPR · CCPA · HIPAA · GLBA
→
AI Governance
EU AI ACT · NIST AI RMF
→
Core Conduct
HONESTY · DISCLOSURE · ESCALATION
→
Your policies
TRAINED INTO ANCHOR 3.0 AS AN ADAPTER, OR UPLOADED ON ANCHOR 3.0-MAX