┌─ dead-signals v0.1.4 ────────────────────────────────────────────── │ evidence ledger · entry 001 · 2026-10-09 └─────────────────────────────────────────────────────────────────── $ autopsy run --target "solana-memecoin-early-entry" --strict
20 variants of the "buy a fresh memecoin, sell at 5–30×" playbook, run against the real trade tape.
corpus: launches tracked ......... 1,194,068 coins harvested .......... 5,654 trades (tx-signed) ....... 43,600,000 candles (1s) ............. 15,600,000 labels ................... 741 winners / 17,397 controls api calls spent .......... 4,871,466 [ ok ] corpus loaded
id axis fires verdict ── ──────────────────────── ───── ────────────────────────── H01 sustained volume velocity 139 WEAK 96% EV in top-5% trades H02 dev holds 44 WEAK dev-conc 64% H04 top-10 concentration 5 FATAL n=5, tail-only H05 social/caller 19 WEAK 93% EV in one week H06 funder clusters 119 DEAD WEAK-DEAD H07 holder acceleration 33 WEAK H09 — 1 DEAD H10 sync-buy avoidance 234 DEAD volume proxy (lift 1.24) H11 leaderboard presence 65 DEAD PIT leak + recount H12 first caller event 590 DEAD mass-fire, all-tail H13 social velocity 6 DEAD H14 CMC sniper% 0 N/A data-gap (5/206) H15 launchpad base-rate 139 WEAK mirrors H01 H16 SOL regime gate 139 WEAK same signal set H17 multi-feature 42 DEAD 85% P&L one dev wallet H18 whale/caller entry 10 DEAD recount −43% H19 — 17 DEAD H20 misc axis 27 WEAK 97% P&L one dev ────────────────────────────── 0 / 20 strategies survive
The naive results looked incredible — which is exactly the trap. A whale-entry variant showed +352.4 SOL on ten trades. A leaderboard feature showed +73.2 SOL. Stop there and you have a strategy and a Twitter thread. Instead we ran four audits.
Fresh code re-derived every result from raw files — no shared functions, no shared assumptions. Sampled exit prices matched the real tape 10/10. The numbers didn't survive: n=10 is an anecdote, and in one group a single trade carried 42% of the entire EV.
For every feature: was it knowable at decision time? The leaderboards table stores fetch-time state — snapshot_date is NULL on every row. "Was this wallet a leaderboard wallet when it bought" silently became "is it one now". The alpha was the future leaking backward.
Surviving signals fired almost only in the top volume quintile — where the base win rate is already 23.7% vs 6.0% overall. Volume-matched controls collapsed the lift to ~1.24×. Two dev wallets accounted for ~12,500 SOL of aggregate P&L. A "population pattern" made of two actors.
Exits assumed zero price impact — infinite liquidity, the oldest lie in backtesting. Fixed with constant-product math + a 2% sandwich tax. The audit also caught a bare except swallowing a NameError on every time-stop exit: 539 of 1,618 simulated trades had silently vanished — selectively deleting exactly the ugly exits that drag EV down.
exhibit A — simulated EV, naive vs post-audit (SOL) H18 +352.4 ███████████████████████████░░░░░░░ +200.8 −43.0% recount H11 +73.2 ██████░░░░░░░░░░░░░░░░░░░░░░░░░░░░ +0.19 −99.7% PIT leak H17 +50.2 ████░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░ ≈0 dead dev-conc exhibit B — 1,618 simulated trades (each cell ≈ 16) ■■■■■■■■■■■■■■■■■■■■■■■■■ ■■■■■■■■■■■■■■■■■■■■■■■■■ ■■■■■■■■■■■■■■■■■■□□□□□□□ □□□□□□□□□□□□□□□□□□□□□□□□□ ■ recorded 1,079 □ deleted by bare except · 539 · 33.3%
Over 50% of launches get sniped in block 0 (median first buy: 0 seconds). ~17% of all transactions are wash trades, and wash trading raises graduation odds up to 10× — so the label "graduated" is partly manufactured, and every feature trained on it inherits the manufacturing.
[ survived ] regime-60s scorer ev/oos ......... +0.087 SOL/trade n=1,192 win rate ....... 53.4% caveat ......... regime-dependent, thin margin, not a launch signal # reported because "everything died" also needs an exception check
VERDICT: NO-GO · 20/20
what survived is not a strategy — it's the pipeline:
recount → point-in-time → selection → execution.
if a backtest has no PIT check and no independent
recount, it isn't a backtest. it's a bedtime story.
next time: write the audits before the hypotheses.
dead-signals — published after the thing dies, receipts attached.
figures verified against source files · 182/182 citations checked, 0 fabricated.
$