Changes to how scores are computed. Each score's methodology_ref tells you which version produced it.
2026-10-18: When at most 1 of the 5 checks returns data and nothing it found is a warning, DEBB now says "insufficient data" instead of giving a verdict. Large, long-traded memecoins (BONK, WIF, POPCAT and others) are treated as established assets and are no longer scored like brand-new tokens.
Details: the model itself and its thresholds are unchanged. Over the API, such a token returns insufficient_data: true; its verdict and score are still the model's output, for older clients. A danger verdict, or any warning from the one check that ran, is always shown. A check that returned no data is now null and listed in unknown_signals, and every response leads with a headline such as "RISKY · 84/100 · new creator: Lower risk needs 88+". Decide on the verdict, not the score: a brand-new creator needs a higher score for Lower risk than a known one. BONK's entry in the established list had a wrong address until this date, so BONK was scored.
2026-10-14: New tokens whose liquidity lock we can't confirm are no longer treated as safe. Expect many more 'risky' verdicts on brand-new tokens. That's intended.
Details: the model's liquidity input now counts only a confirmed Raydium LP burn (95% or more of LP tokens burned). An unlocked or unknown pool is treated as missing data. Before, every pool that wasn't confirmed locked was passed to the model as "unlocked", which it had learned from very few tokens, and that made scores look safer than they should. Scores served before this date used the older input.