AnchorForge
An epistemic benchmark. You run the v5 protocol against any AI. You bring the raw output here. The site reads it, pings every anchor, and writes a permanent record to a public GitHub log.
Not fact-checking. Not debunking. Physics. The model earns its anchors or it doesn’t.
Cert vs. Site-log.
AnchorForge has two public ledgers. They don’t mix.
anchor-forge-protocol
SHA-256 hash + OpenTimestamps + 2 trusted human verifiers. Slow. Expensive in human time. The real leaderboard.
View certified runs →anchorforge-site-log
Fast read by Claude Sonnet 4.6 + HEAD-ping on every anchor. Not a certification. Not a replacement. Just a tamper-evident pre-check.
View site log →Five steps.
Run it.
Login required. Free = 1 gate per calendar week. AnchorForge Pro = 1 per day.
By hitting Gate you agree the full raw output will be written to the public anchorforge-site-log repo, forever, under your GitHub handle.
Gate complete.
View commit on GitHub →Free or Pro.
- GitHub login
- 1 gate per calendar week
- Public ledger commit
- Sonnet 4.6 read
- Everything in Free
- 1 gate per day
- Priority queue
- Cancel anytime
Ranked #1 among non-retrieval models. By our own data.
The site gate uses Claude Sonnet 4.6 (claude-sonnet-4-6). Our benchmark shows Sonar Pro at 95.2% — that’s the #1 overall score. But Sonar Pro uses live web retrieval. A judge that can look up the answer in real time isn’t a blind judge.
Strip retrieval-augmented models and Sonnet 4.6 ranks first: 88.6% citation survival, 2.9% authority misuse — tightest non-retrieval numbers in the field. It adjudicates under the same constraints as every model it judges.