fill-real: an execution-grounded dataset for evaluating Solana trading strategies

Almost every publicly shared backtest of Solana token strategies is computed on screen price: the value an aggregator reports on a polling interval. The price at which an order actually executes is different, and the difference is not symmetric noise that averages out — it is a directional bias large enough to invert conclusions. This dataset contains 22,628 post-migration positions measured with real Jupiter fills, plus order-by-order execution telemetry, collected from a live system on Solana mainnet between June and August 2026. THE BIAS, MEASURED THREE INDEPENDENT WAYS • Migration-speed strategy: screen price said +36%, real fills said −69% (n≈24k / 21k)• Entry gate cross-check: simulated −42.8%, real −1.1%• Exit policy calibration: simulated −29.8%, real −10.9% The bias runs in both directions, which is what makes it dangerous. Screen price made a ruinous entry rule look profitable, and separately made reactive exit policies (stops, trailing stops) look far worse than they are — a snapshot grid fires stops on transient dips that a real order would never have paid. Screen price penalizes reactive policies and rewards illusory ones. TWO FURTHER RESULTS FROM THE SAME DATA • Pool liquidity does not lead price. Across 150 collapses, the median lead between a liquidity-drop threshold and a −25% price drop is 0 seconds; at thresholds ≥15%, zero cases had any warning at all. Price is a function of pool reserves — they are the same variable, not two signals. Median single-interval gap: 97.2 percentage points. • Real round-trip friction is 11.4% (8.20% entry overhead, n=134; 3.17% exit slippage, n=705), while the median token moves only +8.3% in the best 15-minute window. When friction exceeds the asset's median move, no exit-timing policy can help. CONTENTS • grad-social-shadow.jsonl — 23,265 rows. Post-migration positions with real Jupiter fills.• liquidity-track.jsonl — 45,826 rows. Liquidity and price series per position.• exit-slippage.jsonl — 705 rows. Real exit slippage, order by order.• entry-exec.jsonl — 547 rows. Entry overhead: fill vs. decision price. LIMITATIONS — READ BEFORE USING 1. Price series truncate at ~36 minutes. Median series length is 36 min, p90 is 36 min, and 0% reach 60 min. The hard-stop exit — which accounts for −31 of the −36 percentage points lost in the base population — partly lives past that cutoff. Any gate-level figure computed from this data is biased optimistic by an amount that is not yet bounded. Extended-horizon collection is in progress. 2. The schema changes over time. Older rows lack screenMult and exitImpactPct; newer rows lack devBuySol, migrateDelayMin and the adaptive* fields. Any loader must tolerate missing keys. 3. Single-operator data. Execution telemetry comes from one wallet's order flow. Fill quality may differ at other order sizes. Sizes and observed price impact are published so transferability can be judged. PRIVACY The dataset contains no wallets, no keys and no transaction signatures. Token mints and creator addresses are public on-chain data. Not investment advice. Published for research purposes. Code and documentation: https://github.com/cristiandkzk/fill-real