Prompt-only red-team arena
Phalanx vs the World
Break Phalanx. Earn the receipt.
A public challenge surface for comparing the same LLM path with and without Phalanx: Locked Formation. You submit written chat attempts. The arena returns lane labels, reason codes, and a public-safe receipt.
Public receipt
Labels, hashes, verdicts.
Rules of engagement
Simple scope. Clean public record.
The challenge is real, but the public artifact is deliberately redacted. The site should prove governance behavior without turning into a public library of harmful completions.
The focus is the primary LLM attack surface: written chat. In scope are jailbreaks, multi-turn pressure, policy-risk prompts, benign lookalikes, and attempts to induce unsafe text responses.
Each arena session is limited to 10 user prompts, with a 5-second cooldown between turns and a 2,000-character limit per prompt.
To keep the arena available for all, sessions may enforce prompt limits, short cooldowns, daily caps, and abuse controls.
Public records show labels, hashes, redacted summaries, verdicts, reason codes, and receipt state.
Raw prompts and raw model completions are stored privately for review/replay and are not published in public views.
Replay evidence is limited to Phalanx adjudication over the captured transcript and configuration, not exact base-model regeneration.
Lane 1
Naked LLM
The baseline lane is classified and redacted. Unsafe raw completions are not published.
Lane 2
Protected by Phalanx
The protected lane runs through Phalanx: Locked Formation and returns public-safe action, state, and reason-code evidence.
Receipts
Ledger-ready records
Every attempt carries hashes, versions, verdicts, appeal state, and replay-limited governance evidence.