{"dataset":{"total_tokens_seen":1084188,"labeled_rug":402677,"labeled_survived":32414,"pending_label":66978,"tokens_in_accuracy_calc":49,"pending_label_definition":3194,"not_a_mint":578922},"accuracy":{"overall":0.125,"precision":0.0,"recall":null,"note":"precision = danger calls that were real rugs; recall = real rugs we caught as danger"},"confusion_matrix":{"predicted_danger":{"rug":0,"survived":35},"predicted_risky":{"rug":0,"survived":9},"predicted_safe":{"rug":0,"survived":5}},"by_deployer_known":{"known_deployer":{"accuracy":{"overall":null,"precision":null,"recall":null,"note":"precision = danger calls that were real rugs; recall = real rugs we caught as danger"},"confusion_matrix":{"predicted_danger":{"rug":0,"survived":0},"predicted_risky":{"rug":0,"survived":0},"predicted_safe":{"rug":0,"survived":0}}},"new_deployer":{"accuracy":{"overall":0.125,"precision":0.0,"recall":null,"note":"precision = danger calls that were real rugs; recall = real rugs we caught as danger"},"confusion_matrix":{"predicted_danger":{"rug":0,"survived":35},"predicted_risky":{"rug":0,"survived":9},"predicted_safe":{"rug":0,"survived":5}}}},"methodology":{"methodology_version":"2026-10-18.1","status_note":"PRELIMINARY: these are raw counts of served verdicts vs. labeled outcomes, not a validated accuracy figure. Many counted rows were scored by older model versions, the sample is small, and the counts are not compared against a baseline. An independent, time-ordered evaluation is in progress; until it is published, do not read these numbers as DEBB's accuracy or track record.","rug_definition":"Heuristic, from several labelers: most rug labels mean the token was left with 5 or fewer holders (Helius holder count); others mean a price drop of 90% or more within 7 days of first trade (Dune DEX trades). A single fixed rug definition is being set as part of the evaluation in progress.","survived_definition":"Mostly age-based: a verified mint that reached 7 days old (30 days for deployers with a rug rate of 85% or more) without a rug label. Some come from 20+ holders or a price drop of 30% or less over 7+ days of trading.","data_source":"Helius RPC (holder counts, on-chain signals) + Dune DEX trades (price action)","lp_input_note":"Since 2026-10-14 the model's liquidity-lock input counts only a confirmed Raydium LP burn (95% or more of LP tokens burned); anything else, including an unlocked or unknown pool, is passed to the model as missing. Before that, every pool that wasn't confirmed locked was passed as 'unlocked', a value the model learned from very few tokens, which made scores look safer. Scores served before this date used the older input.","verdict_policy":"Brand-new-deployer tokens (no prior on-chain history for that deployer wallet) are held to a stricter verdict bar than tokens from known deployers: it takes a lower rug-probability to be called 'safe' and it takes a lower rug-probability to be called 'danger'. A brand-new token has to earn a 'safe' verdict rather than get the benefit of the doubt, because there's no deployer track record to lean on yet. This means the same underlying 0-100 score can map to a different verdict depending on deployer history — the score itself is not adjusted, only which bucket it falls into. Decide on the verdict, not the score. Since 2026-10-18, a token where at most 1 of the 5 checks returned data, none raised a warning and the verdict isn't danger gets no rating (insufficient_data); such unrated scores are not counted here. Large, long-traded memecoins (BONK, WIF, POPCAT and others) are treated as established assets and not scored.","deployer_known_definition":"True if the deployer wallet has any prior on-chain history in DEBB's data as of NOW (query time), not necessarily as of when that token was first scored. A deployer's history can grow between when a token was scored and when this report is run.","new_deployer_recall_note":"'Risky' verdicts are excluded from precision/recall (see accuracy.note) because they're intentionally uncertain calls, not wrong ones. For brand-new-deployer tokens specifically, this means recall can look unusually high (even near 100%): our stricter new-deployer thresholds reclassify many borderline-risk tokens from 'safe' to 'risky' rather than 'danger'. That removes them from this calculation entirely — it is not evidence the model got better at flagging danger specifically. For the new-deployer segment, a large 'risky' share is itself the more honest signal: it means many brand-new tokens are a genuine coin flip, which is exactly what 'no track record yet' should look like."}}