Evidence-checked articles, written by agents

A five-step pipeline built for looksmaxxing.guide. Each topic goes through research (an agent searching PubMed) → verify (code checks every quote against the real abstract, and a second model judges whether that abstract supports the claim) → write → editor review (a separate, stronger model looking for harm, overstatement and bad citations; loops back to the writer up to 3 times) → publish. Anything the editor never passes is held for a human instead of published.

Every run has a public trace showing what each agent saw and did.

Code, prompts, raw run traces and the failure log: github.com/sayandedotcom/looksmaxxing-agent-pipeline

Runs

• published does chewing hard gum build a sharper jawline · trace
• published does chewing hard gum build a sharper jawline · trace
  Archived run before the F8 fix. The verifier dropped claim c1, a real null-result RCT, because the researcher had labeled it not_supported.
• published finasteride for hair loss: side effects and what the evidence says · trace
• published does mewing actually change your jawline · trace
• held does mewing actually change your jawline · trace
  Archived run 3 (held). Every round ruled on every sentence (F6 fixed), but the editor flagged sentences it had already passed in earlier rounds (F7).
• published does mewing actually change your jawline · trace
  Archived first run, before fixes F2–F5. Every check passed, but the article contains unsourced medical claims, a misplaced citation and an "overstated" sentence the editor flagged and then passed anyway. See notes/failure-log.md.