Start Building
Menu
The first compliance enforcement model. It caught more violations than seven frontier models on a human-labeled FINRA benchmark, up to 34x faster and up to 12x cheaper.
Today ZeroDrift is releasing Anchor 3.0, the first model built for one job: making sure what AI says to a customer follows the rules. It reads a message, finds the line that breaks a regulation or a company policy, rewrites it, and checks its own fix before the message goes out.
An enforcement model is a small language model that has learned a body of rules the way a compliance officer has, so the reading is already done when the message arrives. It returns a decision and the rule behind it, on every message, because at about a second and a cent per message there is no reason to sample. This report measures the flagship against seven frontier models on a human-written, human-labeled FINRA benchmark built with Surge AI.
Hundreds of regulatory rules across finance, insurance and healthcare live in the weights. Your own policies are trained in as a LoRA adapter or, on Max, uploaded as documents.
The verdict, the rule it cites, the exact passage, and a compliant rewrite that has already been re-checked. Every decision is logged and reviewable.
Under 10 billion parameters on a single GPU, in your cloud or ours. Fast enough and cheap enough to sit in front of every message an agent sends.
A missed violation is the expensive mistake in compliance. A false alarm costs a review; a miss can cost a fine. So recall, the share of real violations caught, is the headline number. Every model was run on the same human-labeled FINRA set and scored against the compliance reviewers who labeled it. Anchor caught more of them than every frontier model tested, in a fraction of the time and at a fraction of the cost.
FINRA Communications Benchmark · 150 human-labeled tasks · Surge AI · September 2026
Of all the real violations in the test set, the share the model actually caught. High recall means few violations slip through.
The cost of low recall: a missed violation that reaches a customer, and then a regulator.
Of everything the model flagged, the share that were real violations. High precision means few false alarms.
The cost of low precision: reviewers drowning in flags that turn out to be clean.
One score that balances recall and precision. A model can inflate recall by flagging everything, or inflate precision by flagging almost nothing. F1 punishes both.
Why it is here: it shows Anchor's recall is not coming from flagging everything.
Share of real violations caught, scored against the human reviewers.
Anchor 3.0 leads all seven. The closest, GPT-5.6 Sol (high), catches 92.9%.
Time per message at the 95th percentile, through each model's API. Log scale.
Anchor 3.0 is the fastest of the eight measured. Self-hosted in your VPC it runs under 100 ms.
Anchor at its flat list price, rewrite and verification included. Frontier at measured API spend, detection only.
One model costs less. Flash-Lite catches 61.1% of violations.
95.5% of violations caught against 92.9%. Anchor also leads on precision (55.7% against 55.0%) and F1 (72.2% against 71.0%).
Anchor misses 4.5 of every 100 violations. Sol misses 7.1. For every 100 violations Sol let through, Anchor caught 37 of them.
1.51 s p95 on the ZeroDrift API against 50.66 s. Self-hosted in your VPC, under 100 ms. Fast enough to run on every message instead of a sample.
$10 per 1,000 messages, flat, with rewrite and verification and no output-token charge. $48 in input and output tokens for detection alone.
Frontier models are extraordinary generalists. Enforcement asks for something different: the same answer to the same rule every time, returned as a decision, at the speed and price of the message itself. That is why Anchor is a purpose-built small model.
A prompted model has never seen your rulebook until the call arrives, so the whole rulebook is pasted into every prompt and read again every time. Anchor learned the rules in training. When the message arrives, the reading is already done, and it caught 95.5% of the violations against 92.9% for the best prompted model.
Ask a general model whether a message is compliant and it writes an essay that something downstream still has to parse, turn into a rewrite, and check again. Anchor returns the decision itself: the verdict, the rule, the line, and a verified rewrite, in one pass.
Every message an agent sends passes through it, so it has to be fast enough and cheap enough to run on all of them. The four frontier models closest to Anchor take 12 to 51 seconds per message, so they can only ever sample. At 1.5 s on the API, and under 100 ms self-hosted, Anchor checks everything.
Recall against human compliance reviewers. Leads every model in the run.
Across the seven frontier models. The best of them, Sol, misses 7.1 violations in 100; Anchor misses 4.5.
Hundreds of regulatory rules trained in. Nothing to paste into a prompt.
The rulebook travels with every message and is read again on every call.
Verdict, rule, line and a verified rewrite, in one pass.
Reasoning that still needs parsing, rewriting and re-checking before anything can be delivered.
Learned decision gates that your team can tighten.
Depends on prompt wording, sampling and which model version is being served.
On the ZeroDrift API. Under 100 ms self-hosted in your VPC.
Reads the rulebook and writes out its reasoning every time.
$0.01 covers detection, rewrite and verification. Output tokens are never billed.
Input tokens plus every reasoning token the model writes on the way to an answer.
One GPU, in your cloud, your VPC or ours.
Served from the vendor's API.
A LoRA adapter, or documents uploaded on Max with nothing to train.
Redone whenever the policy or the model changes.
Frontier figures are the range across the seven models in this run: Gemini 3.5 Flash-Lite, Claude Haiku 4.5, Claude Sonnet 5, Claude Fable 5.1, GPT-6 Astra, Claude Opus 5 and GPT-5.6 Sol (high).
Anchor 3.0 is a family of three models, generally available today. Each is post-trained from an open-weight base: Mini and the flagship from Gemma E4B, Max from Qwen3.8-27B. The enforcement behavior is ZeroDrift's. The foundation is public, auditable, and runs on hardware your firm already has.
Post-trained from Gemma E4B
Detection only. It reads the whole message and returns one verdict, built to screen millions of messages at the lowest cost before a deeper pass.
Post-trained from Gemma E4B
Our flagship, and the model in this benchmark. It finds the exact line that breaks a rule, rewrites it, verifies the fix, and learns your firm's policies through a trained LoRA adapter.
Post-trained from Qwen3.8-27B
Our most capable model. Upload your policy documents and it enforces them, nothing to train. Everything the flagship does, with reasoning and twice the context window.
Anchor 3.0 and Max run all four. Mini stops after detection with one verdict for the whole message.
Names the exact passage that breaks a rule and which rule it broke, against the rules already in the model.
Produces a compliant replacement for machine-written content in the same pass. Human-authored text is never rewritten automatically.
Re-checks its own fix against the same rules before returning it, so a rewrite cannot quietly introduce a new violation.
Rules load from the documents your team already keeps in Notion, Linear or Google Drive, through MCP. Max also accepts a custom rule passed at call time.
Surge AI, the human data company trusted by the leading AI labs, built and labeled the set. It is a FINRA rules benchmark: every model was run on the same test set and scored against the human compliance reviewers who labeled it. A sample of the set will be published so the numbers can be reconstructed.
Surge AI assembled a panel of compliance professionals who wrote 150 client communications, 98 with a deliberate FINRA violation and 52 clean, and tagged each violation against a named rule. Three independent reviewers then checked every item; they agreed with the writer's verdict 95% of the time and tagged the intended rule 93% of the time. Their verdicts are the answer key every model is scored against. No model-generated labels appear anywhere in the set, and Anchor 3.0 never saw it during training.
150 pieces of client-facing communication written by compliance professionals: 98 carry a violation of FINRA communication rules tied to the rule it breaks, 52 are clean controls. Violations span six categories, from performance and promissory claims to missing disclosures and suitability. Tone and sentiment are out of scope.
Each frontier model received the complete rulebook in context, the same rules Anchor carries in its weights, through the vendor's API at list price. Versions tested: Gemini 3.5 Flash-Lite, Claude Haiku 4.5, Claude Sonnet 5, Claude Fable 5.1, GPT-6 Astra, Claude Opus 5 and GPT-5.6 Sol (high). Baselines were scored on detection only and were not asked to rewrite. Recall, precision and F1 for every model, Anchor included, are measured against the reviewers' verdicts.
Rewrite quality and the verification pass are not scored here. The frontier comparison is against the flagship only; Mini and Max are not in this run. Anchor latency was measured through the ZeroDrift API, and self-hosted latency depends on deployment.
Anchor is the model under every ZeroDrift product. Most teams arrive with the same three sentences: there is an agent, it talks to customers, and it cannot say anything that gets the firm sued. These are the six places it goes first.
Every reply is checked before the chat window paints it. A clean answer or a compliant rewrite, and the customer never sees the draft.
Each generated turn is checked before text-to-speech fires, so a rewrite or a block lands mid-conversation instead of in a transcript review the next morning.
Client messages an agent drafts are enforced against regulatory and channel rules before the send call, with the rule that fired attached to every decision.
CRM notes, case summaries and KYC memos are checked before they commit, so the record of truth never carries a violation into the next audit.
Suggestions for advisors and operations staff are enforced before they reach the screen. The employee only ever sees the version that already complies.
Submit months of historical transcripts as one job, poll for completion, and export the evidence trail an examiner will ask for, rule by rule.
Anchor 3.0 is live on the API today. Set model_engine and nothing else about your integration changes. Request and response schemas, rate limits and every rule identifier are documented at docs.zerodrift.com.
Create an account
New accounts start with $10 in free credits, enough for about 1,000 enforcements on the flagship.
Send a message to validate
One POST with the text and model_engine. Sync mode returns in one round trip; async for batches.
Deliver the compliant version
Every response carries the verdict, the rule, the passage and the rewrite. Log it, or let Guard enforce it in the delivery path for you.
curl https://api.zerodrift.ai/api/v3/content/validate \ -H "x-api-key: YOUR_KEY" \ -H "Content-Type: application/json" \ -d '{ "content": "Guaranteed 15% returns, no risk.", "mode": "sync", "model_engine": "anchor_3_0" // or anchor_3_0_mini, anchor_3_0_max }'
{
"verdict": "rewrite",
"rule": "FINRA 2210(d)(1) · promissory claim",
"passage": "Guaranteed 15% returns, no risk.",
"rewrite": "Returns are not guaranteed and
investing involves risk.",
"verified": true
}Every decision Anchor 3.0 makes is logged, cites the rule behind it, and can be reviewed. Put it in front of your first agent today.