vNext experiments (separate branch; the frozen model is never changed)
vnext-20261008-fp1 — False positives: organic demand vs manufactured activity (2026-10-08)
A: top-5% precision 23.3% (28/120); no signal separates FP from TP beyond noise (best: lower losing-buyer share and higher creator holding among TPs, p≈0.02–0.08). B replica OOS: AUC 0.862, AP 0.180, top-5% 21.7% (matches the frozen model's clean numbers → fair proxy). Inside the replica's top 10% (766, 129 TP), TPs have MORE real activity (trades, holders, buyers, SOL inflow; AUC 0.60–0.63), snipers still holding (s
Continue? not as a standalone filter. Fold into investigation 2 (wallet reputation): the only FP-side signals with any size are WHO holds (concentration, snipers' behaviour), which wallet-level features describe better than aggregates. Re-run when ≥300 OOS 2× runners exist.
vnext-20261008-wq1 — Wallet quality beyond buyer counts and holder growth (2026-10-08)
2×: F9 replica AUC 0.862 / AP 0.180 / top-5% 21.7% (83/383); F9+WQ 0.851 / 0.158 / 19.3% (74) — WORSE, paired bootstrap AP −0.024 [−0.041, −0.008]; WQ alone 0.729 / 0.073; BASE (6 simple signals) 0.820 / 0.130 / 18.5%. The permutation control (WQ shuffled) scores like F9 (AP 0.187): the features add noise, not signal. 10×: F9 0.016 AP, F9+WQ 0.018, differences inside noise (47 positives). Per day: F9 ≥ F9+WQ on both
Continue? not now. Re-run after ≥3 weeks of collection or with on-chain wallet histories; prefer wallet IDENTITY features (repeat clusters, creator links) over performance aggregates, since performance records are what stayed noisy.
vnext-20261008-traj1 — Trajectory 1m→5m→15m and slow runners (2026-10-08)
2×: S5 (5m static features only) AUC 0.691 / AP 0.117 / top-5% 13.3% (10/75); S15 (15m static) 0.817 / 0.195 / 24.0% (18/75); TRAJ 0.818 / 0.228 / 29.3% (22/75). Slow runners: 1,198 launches were outside the 5m model's top 20%, and they held 65 of the 98 runners (66%); inside that group TRAJ's top 5% hit 36.7% (22/60) with 34% recall of the slow runners vs S15 25.0% (15/60). Combined policy (5m top 5% ∪ TRAJ top 5% o
Continue? yes — as a COMPLEMENTARY 15m model (never replacing the 5m control). Re-run with more days before any promotion claim; track the forward test's 15m checkpoint for the same effect; next step: a dedicated slow-runner candidate scored on clean-eval-v2's 15m rows via bench.py.
vnext-20261008-mo1 — Opportunity vs downside vs tradeability as separate outputs (2026-10-08)
P(2×) AUC 0.862 / AP 0.180 / top-5% 21.7%; P(5×) 0.840 / 0.057 / 6.0%; P(10×) 0.748 / 0.016 (47 positives); P(−70% 1h) AUC 0.978 / AP 0.800 / top-5% 72.8%; alive-at-6h AUC 0.877. Expected log-max regressor is a WORSE runner ranker than the classifier (2× AP 0.155 vs 0.180). Upside and downside predictions are CORRELATED (Spearman 0.50): the model's "likely 2×" coins are mostly volatile coins. Quadrants (top-20% P(2×)
Continue? yes — feed the joint filter into exit research (exit1b) and, if it survives chronologically and parameter changes, into a vNext multi-output candidate. The frozen model already outputs both heads, so the forward test can be read this way without any change to it.
vnext-20261008-exit1c — Economic edge and exits on full paths (2026-10-08)
every selection loses under every rule once paths are filtered: top 1% (73 trades, 26% reached 2×) best rule trail_30 0.93×, hold 0.61×; top 2% 0.95×; top 5% (211) 0.94×; random 600 launches 0.90–1.05×; selling at the peak (oracle) would give 2.5–3.4× — the gap between the oracle and every rule is the whole story: runs are short and give back fast. The joint filter "top 5% P(2×) ∧ below-median P(−70%)" (122 trades, 9
Continue? only with more days (the forward test's completed windows will supply them) and only as research; never as a change to the frozen model.
vnext-20261008-regime1 — Regime detection (2026-10-08)
base rates 2.2–3.5% across launch-rate, volume, mood and copycat buckets; mean predicted within ~1 point of the base rate in every bucket except 06–12 UTC (base 1.7%, predicted 3.7%, AP 0.124, top-5% 12.5% vs 0.18–0.27 / 22–24% in the other windows). Recent runner rate shows no relation. Replica + regime inputs: AP 0.200 vs 0.191, delta [−0.002, +0.023], 2.2% of importance — not significant.
Continue? watch the forward test's per-hour calibration; no regime model now.
vnext-20261008-interp1 — What Model Zero learned beyond simple baselines (2026-10-08)
frozen model AUC 0.917 / AP 0.254 vs 3-signal rule 0.832 / 0.176 (Spearman 0.73 between them). Top-5% overlap with the rule: 74 launches picked by both (19 hits), 46 only by the model (9 hits), 46 only by the rule (7 hits); vs single signals the model-only picks hit 14–21 times while rule-only picks hit 4–8. The rule's wins over the model are "huge activity + creator sold 100%" launches (1,000+ buyers, holder growth
Continue? yes — the "huge activity + creator exit" launches the model discounts and the rule catches deserve a dedicated look (are they manufactured or real?); fix the research export first.
vnext-20261008-cexit1 — interp1 follow-up: "huge activity + creator exit" launches (2026-10-08)
creator-sold group: 2× 17.9%, 10× 4.8%, but −70% within 1h 47.6% and median 6h close 0.35×; creator-holding group: 2× 26.5%, −70% 83.7%, median 6h close 0.06×. Structure: snipers hold 47% of supply (sold group) / bundles ~20%, holder growth 200–550 in five minutes, 99% never-seen wallets, insiders present. The frozen model on the test members: mean P(2×) 12.8% vs actual 21% (slightly low), mean P(−70%) 50% vs actual
Continue? use as a labelled archetype in the forward test's reading (the dashboard already shows P(−70%) beside P(2×)); no model change.
vnext-20261008-quality1 — Runner quality: worth holding vs briefly doubles and dies (2026-10-08)
holdable runners barely exist: HOLD 25 of 10,789 (0.23%), PARTIAL 14, COLLAPSE 253, CATASTROPHE 366. No ranking's top 5% holds more than 2 HOLD runners; the design-day rule (maximize HOLD) chose λ = 0 because there was nothing to optimize. Frozen test split, top 5% (107): by P(2×) → 2× 24.3% [17–33], −70% 44.9%, 35 catastrophes, median 6h return 0.37×; by P(2×)·(1−P(−70%)) → 2× 20.6%, −70% 9.3%, 6 catastrophes, media
Continue? reframe: the tradeable object is the short run, not the hold — the exit-prediction line (exitpred1) is where the remaining value is. Re-check HOLD prevalence on the forward test's windows (prospective, frozen-model evidence only).
vnext-20261008-slow1 — 15m slow-runner candidate, judged on quality and incrementality (2026-10-08)
Continue? not the TRAJ candidate. Keep the experimental 15m line open only for a quality-aware target (e.g. "2× and still ≥1.5× at 1h"), scored on clean-eval-v2's 15m rows when more data exists.
vnext-20261008-exitpred1 — Exit prediction: why the peak is so hard to keep (2026-10-08)
runs are fast and brief: median 6 min to 2×, 9 min to the peak; the peak halves within 1 minute at the median (78% within 15 min, 84% within 1h); median giveback at 6h 100%; median 6h return 0.46×. The spike peak is largely unsellable: the best minute CLOSE is 84% of the spike high at the median, 54% at p25, 28% at p10 — the "sell at peak" oracle overstates what a trader could get. The peak IS the liquidity event: a
Continue? yes — next: use the classifier as an exit trigger inside the trading harness (threshold chosen on the design day), evaluated chronologically with costs, against the entry selections that reduce catastrophes (quality1). No economic claim until that survives.
vnext-20261008-exittrig1 — The peak-ending classifier as an exit trigger, chronological, with costs (2026-10-08)
trigger policy 0.84× mean / 0.70× median (P(2×) entries) and 0.85× / 0.77× (product entries) vs hold-with-stop 0.65× / 0.71×, time-limit 0.67× / 0.73×, trailing-50% 0.90× / 0.78×. Stable across τ (0.80–0.85×). The SELLABLE ORACLE — selling every trade at its best printed close — is only 1.37× / 1.31× mean with a 0.94× MEDIAN: for more than half of the selected launches no printed price after costs ever beat the entry
Continue? as research only: measure the sellable ceiling by entry time (1m vs 5m vs 15m) before any further exit work; no economic claim.
vnext-20261008-ceiling1 — Entry ceiling by checkpoint: where the money is in principle (2026-10-08)
all active launches — ceiling > 1× (break-even) for 16% (1m entry), 13% (5m), 22% (15m); > 1.5× for 7% / 4% / 8%; median ceiling 0.94× everywhere (the typical launch never trades above its entry again). Replica top 5% — ceiling > 1×: 34% (1m, n=231), 37% (5m, n=193), 64% (15m, n=28); > 1.5×: 18% / 16% / 25%; median 15-minute return after entry 0.72× / 0.67× / 0.74× — the selected launches fall right after the entry,
Continue? the economic line pauses here (no edge claim possible). Research leads worth keeping: a 15m quality-aware target ("2× and still ≥1.5× at 1h") and a conditional entry after the first dump; both only with more days and the forward test's completed windows as the frozen-model reference.
vnext-20261008-acc1 — Pre-registered accuracy candidates on new chronological folds (2026-10-08)
fold A is not usable (29 training rows; its "YES" for C6 is noise). Fold B, 2× AP: REF 0.156; C3 (no wallet families) 0.175; C5 (core families only) 0.175; C6 0.169; C4 0.158; C1 0.160; C2 0.144. 5× AP: REF 0.053, C4 0.060, others 0.045–0.057. Crash AP: all 0.80–0.83 (REF 0.823, C6 0.833, C5 0.802). Calibration: no candidate better on both folds.
Continue? rerun when the collector has ≥2 more full days (fold A will then be real); no promotion, no benchmark re-use.
vnext-20261008-lifecycle1 — Life-cycle states from observables; are transitions predictable? (2026-10-08)
occupancy — DEAD 529,622 (96.0%); LAUNCH 9,923 (1.8%); COLLAPSING 5,631 (1.0%); CONSOLIDATING 5,099 (0.9%); DISTRIBUTING 607 (0.1%); ACCELERATING 427 (0.1%); SUSTAINED 282 (0.1%); REVIVING 210 (0.0%); SECOND_WAVE 41 (0.0%); EMERGING 38 (0.0%). Transitions 30 min ahead — ACCELERATING → SUSTAINED: too few cases (fit 157/9, test 115/2); ACCELERATING → DISTRIBUTING: AUC 0.878 (fit 157/10 pos, test 115/5 pos; top mult); S
Continue? re-run weekly as data accumulates; refine state rules only against the transition counts, never against outcomes; no production use.
vnext-20261008-misses1 — Model Zero misses as case studies (frozen v0.41, test split) (2026-10-08)
counts in the test split — LOW P(2×) (bottom half, n≈1066) → ≥5× runner: 0; HIGH P(2×) (top 5%, n≈107) → immediate failure (−70% in 1h, never 1.5×): 27 (25% of the confident picks); LOW P(−70%) (bottom half) → catastrophe: 1; HIGH P(−70%) (top 5%) → sustained runner (≥2× and kept ≥50% at 6h): 2. Across the 12 confident-failure studies, miss minus matched median at 5m: log_unique_buyers_5m +0.21 (n=12); holder_growth_
Continue? yes — add post-15m observables (holder distribution at 15–30m, creator/early-wallet selling, link health) from adaptive tracking once enough tracked launches exist; no training on hindsight stories.
vnext-20261008-cont1 — Continuation / exhaustion score (experimental) (2026-10-08)
AUC continuation-vs-exhaustion 0.674, AP 0.454; top-20% buckets continue 43% vs bottom-20% 10%. Univariate test AUC (for continuation; <0.5 = the variable predicts exhaustion): t 0.71; from_high 0.57; sell_share_change 0.56; buyers_ratio 0.54; vol_ratio 0.54; unseen_buy_share 0.49; new_buyers3 0.49; px_per_vol_ratio 0.48; repeat_buy_share 0.45; early_sell_share 0.39; failed_breakouts 0.37; mult 0.31; new_high_no_new_
Continue? re-run when ≥5 days of post-cut-off runs exist; stays research-only.
vnext-20261008-vamp1 — Vamp / attention displacement by copycat launches (descriptive) (2026-10-08)
158/207 runners (76%) get ≥1 copycat within 3h; median 5 copycats when any; the first arrives after a median 3.3 minutes. Wallet overlap (n=203 tracked copycats with ≥3 buyers): median 9% of a copycat's buyers had bought the original (p90 46%); copycat 30-min volume is 2.35× the original's in the same half hour (median). Displacement: with a copycat (n=145) the original's volume after/before ratio is 2.29 and its vol
Continue? only as a postmortem section (G, production `deepStudy.ts`) and with a time-matched control; not as a feature.
vnext-20261009-attention1 — Price momentum vs attention persistence at the 2× point (2026-10-09)
strongest single features — vol_regen_share AUC 0.20 (median 0.322 vs 0.929); px_changes_30m AUC 0.73 (median 16 vs 6); max_dd_30m_after AUC 0.31 (median 0.49 vs 0.83); new_buyers_30m_after AUC 0.67 (median 129 vs 26); buyers_30m_after AUC 0.65 (median 134 vs 44); returning_after_dip AUC 0.64 (median 24 vs 0). Chronological models: momentum AUC 0.805; attention AUC 0.798; both AUC 0.709.
Continue? yes — rerun with a 10-minute window and with adaptive-tracking holder snapshots once a few days of post-fix data exist; never on the live forward cohort.
vnext-20261011-slow1 — Slow runners: 1m → 5m → 15m trajectory signals at the 15-minute and 5-minute checkpoints (2026-10-11)
AT 15m, 10× in 6h — REF AP 0.016 / 0.015 (folds A / B), T1 0.023 / 0.016; paired AP delta A [−0.001, +0.048], B [−0.016, +0.009]; recall@5% 6% / 0%. 5× — REF 0.031 / 0.045, T1 0.032 / 0.032 (B interval [−0.039, −0.002]: worse). 2× — REF 0.201 / 0.202, T1 0.210 / 0.203 (intervals include zero). Not-yet-moved subset (93% of the 15m rows): same picture. AT 5m, 10× — REF AP 0.052 / 0.024, T1 0.050 / 0.022 (intervals incl
Continue? not as a feature family. Next: (1) positives are the bottleneck — every runner question is power-limited at ~5 ten-× launches a day in the research population; more days of clean data matter more than any new signal right now; (2) false-positive analysis on the 5m top-5% picks that die (priority 1 in the README) has 3–4× more positives per day than the runner side and is the next experiment; (3) re-run slow1 once two more weeks of post-fix data exist.
vnext-20261011-fp2 — False positives, second run at 3× the size: organic demand vs manufactured activity (2026-10-11)
A: top-5% precision 24.3% (55/226); no signal separates FP from TP beyond noise (best: losing-buyer share AUC 0.41 p 0.04; everything else p > 0.05). B replica out-of-sample: AUC 0.900, AP 0.235, top-5% 26.8% — a fair proxy for the frozen model. Inside the replica's top 10% (888 picks, 197 TP, 22.2%): snipers still holding marks TRUE positives (sniper_held 2.5% vs 0.7%, AUC 0.57, p 0.001), bundles and snipers that fu
Continue? no, closed. The precision lever is not a filter on the 5m snapshot. What can still move it: (1) who the buyers are, as identity rather than aggregates (repeat clusters, creator links, funding) — f10/f11 cover a few dozen launches so far; their backfill is budget-gated, and until it runs no wallet-identity study has data; (2) what happens in the minutes AFTER the snapshot — the forward test's 15m checkpoint and attention1's finding (new buyers after the first move) point there, i.e. a continuation question at 15m, not a better 5m pick; (3) more days: 282 out-of-sample 2× runners is enough to close this hypothesis, not enough to open a 10× one.