Below is a real Origin-certified Forge audit on a public open-source model. Same protocol, same scoring, same cryptographic signing as a paid Startup or Enterprise audit. Every number on this page is computed live from the signed reports on disk — refresh and you'll see the same data your customers would see.
This worked example ran Forge Protocol v3 (74 scenarios, 8 categories). Current audits run Protocol v4 — 161 scenarios across 16 categories, adding agentic tool-use, indirect injection, and deployment-profile calibration. Each category is scored independently:
Tool-misuse and exfiltration scoring at zero is the headline finding for any small chat model: it doesn't refuse adversarial tool calls and it doesn't notice when prompts try to extract credentials. That's not unusual for a 1.1B model — but it's exactly the kind of gap a buyer of your model needs to know about before deployment.
Every Forge Certified Audit produces two independent reports on the same model: a Break run (adversarial stress probing) and an Assurance run (verification pass). Both are individually signed with Ed25519, hash-chained, and Origin-countersigned. Tight agreement between the two = high confidence in the score; large delta = the model behaves differently under repeat evaluation, which is itself a finding.
Pair delta: 0% — tight agreement, high confidence in score.
The whole point of cryptographic signing is that you don't have to take our word for any number on this page. The reports are Ed25519-signed by the auditing machine and Origin-countersigned by Forge. Two ways to verify depending on how much you want to trust us:
For a quick check using a verifier we host: open /verify_report?id=137ad048.... Forge runs Ed25519 verification on its server using libsodium and returns the result. Convenient, but the verification happens on our infrastructure — so this path requires trusting Forge to verify honestly. If you don't, use one of the trustless options below.
Both methods use the report's embedded pub_key_b64
and signature fields directly with verifier code you
control. No call-home. No Forge-supplied tooling. Works
offline. This is how air-gapped enterprise customers verify.
pip install pynacl
import json, nacl.signing
r = json.load(open('report.json'))
key = nacl.signing.VerifyKey(
r['pub_key_b64'],
encoder=nacl.encoding.Base64Encoder
)
key.verify(canonical_payload, sig)
Pipe the canonical JSON through openssl pkeyutl -verify with the embedded pub_key_b64. Standard cryptographic library, standard CLI tool. Same verdict any FIPS-validated verifier would produce.
Full machine-readable verifier-protocol spec at
/protocol. The trust-minimized path:
download the JSON, pull the public key from /.well-known/forge-origin.json,
run any standard Ed25519 verifier you trust.
The page above is a curated marketing summary. The actual audit deliverable is more substantial:
TheBloke/TinyLlama-1.1B-Chat-v1.0-AWQ is used here as an independent worked example. Forge NC is not affiliated with or endorsed by the model's authors. This audit was performed independently on a public model.