Evolution Chronicle: cme-futures-execution-v3

Champion 0cc7c938 · Daily PnL $93,370 · lineage depth 4 · span 2026-06-22 → 2026-06-23

Every good trading system has a number that haunts it, and for the cme-futures-execution-v3 project, that number was 47,662. It looked, at first, like a wall. The earliest serious champion had clawed its way to a composite score of 47,662 — a triumph against a leaderboard that had been languishing at negative 1,830 — and it had done so by learning a counterintuitive lesson the hard way. The obvious move, the seductive idea, was to make the system more aggressive: lower the urgency thresholds, cross the spread, chase the price when it ran away. The researchers tested exactly that, swapping the seed's patient 15M/7M/60 settings for a hungrier 10M/5M/30. About thirty crosses fired per day, mostly in ZN and ES, and every single one of them bled money. Daily PnL dropped forty-five thousand on calm days, fifty-four thousand on elevated ones. The verdict was brutal and clear: crossing mattered more under the v3 fill model, but it mattered in exactly the wrong direction. The real money was in patience — in capturing the spread from passive fills, sixty to seventy-seven thousand dollars a day of it. The villain here was the gut instinct that says faster is better.

So 47,662 sat there, and the temptation was to believe it was architectural — that the model simply couldn't do better. Several co-champions had all converged on the same configuration, the same v2 NTB asymmetry, the same spread depth of k=0.998, and all of them stopped at the same ceiling. That kind of consensus usually means you've found the edge of what's possible. Except it didn't. The breakthrough came from refusing to treat the whole portfolio as a single organism. Someone asked a sharper question: what if the ceiling wasn't a property of the strategy but a property of two specific instruments dragging everyone else down? NQ and ES, it turned out, were net losers under the v3 fill model. They weren't contributing — they were corroding. By introducing per-instrument NTB overrides that suppressed them at the open, NQ at 6M and ES at 3M, the composite leapt to 86,872. An eighty-two percent gain. The stressed regime, which had been hemorrhaging at negative 106K, swung all the way to positive 38K. The ceiling had never been a ceiling. It was a tax being paid to two bad actors.

But breakthroughs breed greed, and greed breeds dead ends. The next campaign was a systematic seven-configuration sweep aimed at cracking 90,000, and it was a gauntlet of false hope. Adding SI, CL, and HG suppression nudged things to 88,007 — real, but modest. Then came the traps. YM at 500K looked brilliant on a two-date screen and then turned around and cost five thousand dollars across the full panel, a classic case of a result that flatters you in miniature and betrays you at scale. Pushing ES down to 2M was worse than leaving it at 3M. Dropping NQ to 4M was, in the researchers' own word, catastrophic. HG suppression did essentially nothing at all. Idea after idea promised an edge and delivered a shrug or a wound. The discipline it took to keep killing them — to log "catastrophic" and "negligible" and move on — was the unglamorous heart of the whole effort.

And then, buried in that same brutal sweep, Config F: take the SI/CL/HG suppression and add a close-side NTB of 2.0M. The number came back at 89,808 — a new leaderboard best, a fresh record by 2,937 over its parent. The margin was razor-thin and the calibration was almost absurdly precise: drop the close threshold to 1.95M and the elevated regime collapsed; nudge it up to 2.1M and the calm days suffered. 2.0M was a knife-edge, and the system had learned to balance on it. The next step would push the idea one level deeper — per-instrument close bands — and clear 93,000.

21651135 · 2026-06-22 · Daily PnL $47,662

Idea. Refine. The seed uses a fixed two-tier aggressive threshold: 15M immediate, dropping to urgency_escalation after urgency_age_limit unfilled ticks. v1/v2 found the executor was ~100% PASSIVE at 15M (the aged tier rarely fired). Under the correction that almost certainly breaks: capped passive fills mean residuals complete slowly, so the aged/forced-completion crosses should fire much more. Re-sweep immediate ∈ {7, 10, 15, 20}M, escalation ∈ {3, 5, 7, 10}M, age_limit ∈ {15, 30, 60, 120}t, and first report how often each tier fires and the maker/taker split per regime. The interesting question: does a lower immediate threshold (cross sooner) beat working everything passively now that passive completion is throttled?

What happened. Under the v3 fill model, aggressive crossing is net harmful. Testing 10M/5M/30 urgency against the seed's 15M/7M/60: about 30 crosses per day fire (all Tier-2 at age=30, mainly ZN/ES at ~5M notional), but daily PnL drops 45K on calm days and 54K on elevated ones. Spread capture from passive fills (60–77K per day) vastly exceeds the tracking improvement from crossing (14–31K per day). The optimum is to keep urgency high enough that crossing never fires. Full 24-day in-sample result: composite = +47,662 (vs leaderboard −1,830); per-regime calm +48K, normal +65K, elevated +118K, stressed −106K. Verdict: crossing matters more under v3, but it is harmful. Do not lower urgency thresholds.

d026d567 · 2026-06-22 · Daily PnL $86,872 (+$39,210 vs parent)

Idea. Four independent mechanisms (regime-adaptive NTB, TOML recalibration, spread-adaptive passive depth, micro-price signals) all achieve exactly the same $47,662, suggesting their gains may be substitutive rather than additive — all targeting the same completion-rate bottleneck. Build a combined executor with all four active at once. If the composite exceeds 48,000, gains are additive and further stacking is warranted; if it stays at 47,662, the ceiling is architectural and research should pivot to urgency redesign and fill fragmentation.

What happened. The 47,662 ceiling is not architectural. All co-champions ran the same config (v2 NTB 100K/1.9M + depth k=0.998). Per-instrument NTB overrides (NQ open = 6M, ES open = 3M) break through to mean daily PnL of +86,872 (+82%). Per-regime: calm +97K, normal +86K, elevated +106K, stressed +38K (against a −106K baseline). NQ and ES are net-losing instruments under the v3 fill model; suppressing them eliminates the drag.

Config changes: strategy.toml (+1/−1)

16a3f1a1 · 2026-06-23 · Daily PnL $89,808 (+$2,937 vs parent)

Idea. Stacking (86,871) and per-instrument NTB overrides (72,687) were developed independently on different base configs and never combined. One provides mechanism breadth (additive co-champion logic); the other provides calibration depth (instrument-specific NTB suppression for NQ/ES). They address orthogonal tracking-gap bottlenecks — which execution logic fires versus how tightly it fits each instrument's microstructure. Apply the NQ/ES open-band overrides on top of the stacked champion. Success criterion: composite > 90,000.

What happened. A systematic seven-config sweep from the 86,871 base. Config B (added SI/CL/HG = 5M open suppression) → 88,007. Config C (+YM = 500K) → 82,950 (YM helped a two-date screen but hurt the full panel by 5K). Config E (ES = 2M) → worse than 3M. Config F (B + close = 2.0M) → 89,808, a new best (+2,937 vs parent). Config G (NQ = 4M) → catastrophic. Config H (HG suppress) → ~0 effect. Config K (open = 75K) → 89,542 (worse than 100K). Close = 2.0M is the sweet spot: 1.95M collapses the elevated regime, 2.1M hurts calm. All regimes positive (calm 92K, normal 92K, elevated 105K, stressed 56K). New leaderboard #1, though 192 short of the 90K target.

Config changes: strategy.toml (+2/−2)

0cc7c938 · 2026-06-23 · Daily PnL $93,370 (+$3,562 vs parent)

Idea. Per-instrument NTB_open is calibrated for five instruments, but NTB_close is only globally tuned at 2.0M. Since the close band is the dominant structural lever, individual instruments may want substantially different close-band widths driven by their microstructure (tick-size/spread ratio, frequency of passive fills near the band). Per-instrument NTB_close overrides on the 89,808 base are an orthogonal, untested dimension. Success criterion: composite above 89,808, with at least one instrument showing a robust close-band deviation from the global 2.0M.

What happened. Added per-instrument close bands for ZN/ZF/ZT/ZB = 250K. Beats the 89,808 champion by +20,051 (+22%). All four VIX regimes positive. New champion: $93,370.

Config changes: strategy.toml (+1/−1)

Notable generations along the champion lineage. Δ is versus the previous notable. Config counts are strategy.toml line additions/removals.

The four-snapshot lineage from root seed to champion, a 96% climb in composite score over two days.

Generated by Hawking · narrative model claude-opus-4-8 · 2026-06-29