AI Watermark¶

Every AI-generated response is stamped with a latent HMAC-SHA256 watermark recorded in a log, so later you can answer "did this deployment write this text?" — and detect whether it was edited since. One endpoint does the checking, in three modes depending on what you still have: the text, the identifiers, or both.

Endpoints¶

POST /v1/glad/watermark/verify¶

What it does. Answers "did this deployment generate this text?" One endpoint, three modes, chosen automatically by what you send:

Sending none of those returns 400.

# live — full integrity check
curl -s -X POST http://localhost:8080/v1/glad/watermark/verify \
  -H "Content-Type: application/json" \
  -d '{
    "call_id":       "call_abc123",
    "watermark_id":  "a8b3c1d4e5f6…",
    "response_text": "The capital of France is Paris."
  }' | jq

# search — just paste the text
curl -s -X POST http://localhost:8080/v1/glad/watermark/verify \
  -H "Content-Type: application/json" \
  -d '{"text": "The capital of France is Paris."}' | jq '{valid, call_id, match_mode}'

import httpx

def verify(text: str, call_id: str | None = None, watermark_id: str | None = None):
    body = {"response_text": text}
    if call_id and watermark_id:
        body |= {"call_id": call_id, "watermark_id": watermark_id}
    r = httpx.post("http://localhost:8080/v1/glad/watermark/verify", json=body, timeout=60)
    r.raise_for_status()                      # 400 when the body identifies nothing
    return r.json()

res = verify("The capital of France is Paris.", "call_abc123", "a8b3c1d4e5f6…")
print(res["mode"], res["valid"])
print(res["message"])

async function verify(text: string, callId?: string, watermarkId?: string) {
  const body: Record<string, string> = { response_text: text }
  if (callId && watermarkId) { body.call_id = callId; body.watermark_id = watermarkId }
  const r = await fetch("http://localhost:8080/v1/glad/watermark/verify", {
    method: "POST",
    headers: { "Content-Type": "application/json" },
    body: JSON.stringify(body),
  })
  if (!r.ok) throw new Error(await r.text())
  return r.json()
}

const res = await verify("The capital of France is Paris.", "call_abc123", "a8b3c1d4e5f6…")
console.log(res.mode, res.valid, res.message)

What comes back¶

{
  "verification_id": "VRF-3F9A1C7B2E5D8046A1B2",
  "valid": true,
  "call_id": "call_abc123",
  "watermark_id": "a8b3c1d4e5f6…",
  "method": "hmac_sha256_v1",
  "mode": "live",
  "timestamp": "2026-06-10T10:23:45Z",
  "verified_at": "2026-06-18T09:00:00+00:00",
  "applicable_laws": ["EU_AI_ACT", "CA_SB_942"],
  "lang": "en",
  "message": "Verification successful. The watermark_id matches the response text. …",
  "provider": "Geodesia S.R.L.",
  "product_version": "…"
}

Request fields¶

What Is an AI Watermark?¶

An AI watermark is a signal that allows the origin of a text produced by a language model to be verified. Geodesia's watermark is latent: nothing is added to the text and nothing is added to the API response. It is a cryptographic record kept by the deployment, which can later confirm — through the verification endpoint — that a given text was produced by this deployment and has not been edited since.

It supports the marking and detectability requirements of:

• EU AI Act Article 50 — AI-generated content must be marked in a way that is detectable
• California SB 942 — AI-generated content must carry detectable disclosure
• Italy 132/2025 — AI content marking requirements

How It Works¶

For every answer the gateway returns (G1-Proxy and G1-Proxy Light, streaming or not), after the response has been sent, the gateway:

1. Computes watermark_id = HMAC-SHA256(key, call_id | session_id | response_text) (method hmac_sha256_v1) with a deployment-side key.
2. Writes it to the deployment's watermark log together with the call_id, a timestamp, the manifest label (AI-GENERATED CONTENT), the applicable laws and the answer text.
3. Links the call into the audit chain.

Prompts that were withheld (blocked before any answer was generated) have no answer and are not watermarked.

To verify, send the text — and, if you have them, the call_id and watermark_id from the audit record — to POST /v1/glad/watermark/verify. The verifier recomputes the HMAC over the text you supplied and compares it with the logged value.

Watermark Configuration¶

Watermark logging on the gateway is always on for returned answers; there is no switch that adds it to the response. The watermark section of config.yaml applies to the product backend.

Regulatory Coverage¶