Models answer with confidence and a frozen brain. Paste any claim and we check it against the live present — and hand back the current truth with an Ed25519 signature you can verify yourself.
Before an autonomous agent acts — answers a user, files a report, makes a call — it can run the claim past live data and get back a verdict it can prove. That's the difference between a model that sounds right and a system that checked. One call, keyless:
POST https://dynamicfeed.ai/v1/batch with {"calls":[{"tool":"reality_check","args":{"claim":"…"}}]} — returns the verdict, the live value, a confidence, and a signature. Wire it into LangChain, CrewAI, the OpenAI Agents SDK & more as a pre-answer guard. Verdicts are signed but not gospel — we prove the live value, not the wisdom of the claim.
The whole guard is ~8 lines, keyless, stdlib-only — drop it in front of any answer your agent gives about the present:
import json, urllib.request
def guard(claim, base="https://dynamicfeed.ai"):
"""True = safe to assert; False = stale/wrong (answer from the live note instead)."""
req = urllib.request.Request(base + "/v1/batch",
data=json.dumps({"calls":[{"tool":"reality_check","args":{"claim":claim}}]}).encode(),
headers={"Content-Type":"application/json"})
d = json.load(urllib.request.urlopen(req, timeout=25))["results"][0]["data"]
return d.get("verdict") != "stale_or_wrong", d.get("note","")
ok, live = guard("the latest stable Python is 3.11")
# ok == False · live == "You said python 3.11; the latest stable is 3.14.6. Out of date."
Keyless, signed, one line. Connect it — or see what your model's already wrong about.