Train an Adapter

You have imported and activated a policy. Extraction gives Anchor the letter of the policy. Training Studio trains a policy-specific adapter on ZeroDrift’s base so enforcement can apply that policy’s judgment, not only the extracted rule list. Anchor handles regulations. Your adapter handles your policies. Train when the policy is nuanced, when extracted rules alone over-escalate, or when you want a single model path for regulations plus your policy. You do not have to train for custom rules to run — Activate Rules is enough for extracted rules. Untrained custom rules run on Max (anchor_3_0_max). Training lets Flagship (anchor_3_0) check that import. At launch, adapters are available only for Anchor 3.0 (Flagship). Max evaluates uploaded policies at inference and has no training charge. See Choose a model and Pricing.

Terms

Term	Meaning
Policy	The imported document and the rules extracted from it.
Adapter	The policy-specific model trained from that import. After a successful training poll, scoped enforcement can use it.
Training run	One Training Studio job for an import, identified by `training_run_id`.

Prerequisites

Full-access API key (read-only returns 403).
An import whose rules are activated (pending review returns 409).
The original document still within retention — 7 days by default (410 if expired).
No other training run in progress for this import (409).
Training enabled for your environment (503 if not).

Set IMPORT_ID to the activated import id from Import a Policy.

Flow

train → poll until succeeded|failed → enforce with validation_scope.imports

Start, poll, enforce

Each sample starts training, polls to a terminal status, then enforces with validation_scope.imports. Omit the POST body to use Training Studio defaults, or set cost_cap_usd and examples_per_rule. Do not enforce against a still-training run. ZeroDrift promotes the adapter only after a successful status poll. Until then, scoped enforcement follows its configured fallback path. A failed retrain does not replace an earlier working adapter.

IMPORT_ID="${IMPORT_ID:?set IMPORT_ID to an activated import}"

curl -fsS -X POST "https://api.zerodrift.ai/api/policies/import/$IMPORT_ID/train" \
  -H "x-api-key: $ZERODRIFT_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"cost_cap_usd": 25, "examples_per_rule": 8}'

# Poll until training_status or status is succeeded or failed
while true; do
  RESP=$(curl -fsS "https://api.zerodrift.ai/api/policies/import/$IMPORT_ID/train" \
    -H "x-api-key: $ZERODRIFT_API_KEY")
  STATUS=$(printf '%s' "$RESP" | jq -r '.training_status // .status')
  echo "$STATUS"
  case "$STATUS" in
    succeeded|failed) break ;;
    queued) ;;
    *) printf '%s\n' "$RESP" >&2; exit 1 ;;
  esac
  sleep 10
done

printf '%s\n' "$RESP"
test "$STATUS" = "succeeded"

curl -X POST "https://api.zerodrift.ai/api/v3/content/validate" \
  -H "x-api-key: $ZERODRIFT_API_KEY" \
  -H "Content-Type: application/json" \
  -d "{
    \"content\": \"Our fund guarantees 20% returns with zero risk!\",
    \"model_engine\": \"anchor_3_0\",
    \"mode\": \"sync\",
    \"validation_scope\": {\"imports\": [\"$IMPORT_ID\"]}
  }"

import os
import time

import requests

API = "https://api.zerodrift.ai"
HEADERS = {
    "x-api-key": os.environ["ZERODRIFT_API_KEY"],
    "Content-Type": "application/json",
}
IMPORT_ID = os.environ["IMPORT_ID"]
URL = f"{API}/api/policies/import/{IMPORT_ID}/train"

started = requests.post(
    URL,
    headers=HEADERS,
    json={"cost_cap_usd": 25, "examples_per_rule": 8},
    timeout=30,
)
started.raise_for_status()
print(started.json())

while True:
    training = requests.get(URL, headers=HEADERS, timeout=30)
    training.raise_for_status()
    body = training.json()
    status = body.get("training_status") or body.get("status")
    print(status, body.get("stage"))
    if status in ("succeeded", "failed"):
        break
    time.sleep(10)

if (body.get("training_status") or body.get("status")) == "failed":
    raise SystemExit(body.get("error", "Training failed"))

enforced = requests.post(
    f"{API}/api/v3/content/validate",
    headers=HEADERS,
    json={
        "content": "Our fund guarantees 20% returns with zero risk!",
        "model_engine": "anchor_3_0",
        "mode": "sync",
        "validation_scope": {"imports": [IMPORT_ID]},
    },
    timeout=30,
)
enforced.raise_for_status()
job = enforced.json()
if job.get("mode") == "async":
    poll = job["poll"]["url"]
    while True:
        job = requests.get(f"{API}{poll}", headers=HEADERS, timeout=30).json()
        if job.get("status") in ("completed", "failed"):
            break
        time.sleep(2)
print(job.get("status"), (job.get("result") or job).get("overall_status"), job.get("served_by_adapter"))

const API = "https://api.zerodrift.ai";
const HEADERS = {
  "x-api-key": process.env.ZERODRIFT_API_KEY!,
  "Content-Type": "application/json",
};
const IMPORT_ID = process.env.IMPORT_ID!;
const URL = `${API}/api/policies/import/${IMPORT_ID}/train`;

const started = await fetch(URL, {
  method: "POST",
  headers: HEADERS,
  body: JSON.stringify({ cost_cap_usd: 25, examples_per_rule: 8 }),
});
if (!started.ok) throw new Error(`train ${started.status}`);
console.log(await started.json());

let body: {
  status?: string;
  training_status?: string;
  stage?: string;
  error?: string | null;
};
for (;;) {
  const r = await fetch(URL, { headers: HEADERS });
  if (!r.ok) throw new Error(`status ${r.status}`);
  body = await r.json();
  const status = body.training_status ?? body.status;
  console.log(status, body.stage);
  if (status === "succeeded" || status === "failed") break;
  await new Promise((res) => setTimeout(res, 10_000));
}

if ((body.training_status ?? body.status) === "failed") {
  throw new Error(body.error ?? "Training failed");
}

const enforced = await fetch(`${API}/api/v3/content/validate`, {
  method: "POST",
  headers: HEADERS,
  body: JSON.stringify({
    content: "Our fund guarantees 20% returns with zero risk!",
    model_engine: "anchor_3_0",
    mode: "sync",
    validation_scope: { imports: [IMPORT_ID] },
  }),
});
if (!enforced.ok) throw new Error(`enforce ${enforced.status}`);
let job = await enforced.json();
if (job.mode === "async") {
  const pollUrl = job.poll.url;
  for (;;) {
    const polled = await fetch(`${API}${pollUrl}`, { headers: HEADERS });
    job = await polled.json();
    if (["completed", "failed"].includes(job.status)) break;
    await new Promise((res) => setTimeout(res, 2000));
  }
}
const verdict = job.result ?? job;
console.log(job.status, verdict.overall_status, job.served_by_adapter);

Errors

Status	Cause	Fix
`400`	Invalid training options	Check `cost_cap_usd` and `examples_per_rule`.
`403`	Read-only key, or import belongs to another customer	Use a full-access key for the owning account.
`404`	Import not found, or no run started (GET)	Confirm `IMPORT_ID`; POST before you poll.
`409`	Import not activated, or a run already in progress	Activate rules; wait for the current run.
`410`	Original document expired	Re-import, activate, then train (Data Retention).
`422`	Document too short or rejected	Provide a fuller policy document.
`502`	Training Studio request failed	Retry; contact support if it persists.
`503`	Training not configured	Contact support@zerodrift.ai.

Gotchas

One training run per import at a time. A second POST while a run is active returns 409.
Retrain after you materially change the policy (re-sync or re-import, activate, train again).
Adapters apply when enforcement includes that import in validation_scope.imports.
Live GET responses report training_status (queued, succeeded, failed). Some responses also include status.
A failed retrain does not replace an earlier working adapter.

FAQ

Do I have to train before custom rules run?

No. Activate Rules puts extracted rules into enforcement immediately. Untrained custom rules run on anchor_3_0_max. Training Studio adds a policy-specific adapter so Flagship (anchor_3_0) can check that import. See Enforce Content.

Which models can I train an adapter for?

At launch, Training Studio adapters are Flagship-only. Max has no training step — upload the policy and it is evaluated at inference. Training and inference rates are on Pricing. Model choice is on Choose a model; quality and latency are on the Anchor 3.0 benchmark.

Can I train a document I synced from Notion or Drive?

Yes. A connected-source import is the same import_id as a file upload. Activate it, then call the train endpoints. See MCP Connectors.

Why did training return 410?

Training needs the original imported file or text. After default content retention (7 days unless your contract says otherwise), that document is deleted. Import again, activate the new rules, then train.

Why did GET training return 404?

No training run has been started for that import. POST first, then poll.

Import a Policy

Get an activated IMPORT_ID before you train.

Train Policy Adapter

POST /api/policies/import/{import_id}/train

Pricing

Flagship training rates and per-enforcement custom-rule billing.

Getting Started

Integration Guides

Tools

Resources

Terms

Prerequisites

Flow

Start, poll, enforce

Errors

Gotchas

FAQ

Next

Import a Policy

Train Policy Adapter

Pricing

​Terms

​Prerequisites

​Flow

​Start, poll, enforce

​Errors

​Gotchas

​FAQ

​Next

Import a Policy

Train Policy Adapter

Pricing

Terms

Prerequisites

Flow

Start, poll, enforce

Errors

Gotchas

FAQ

Next