design(foxhuntq): architectural pivot away from per-bar DQN to decoupled Belief Bus + Conformal DRL
After 16+ SP-runs producing WR pinned at ~0.435 — and a session-end smoke
showing SP22 H6 vNext K=3 head architecture moves WR to 0.458 only via
degenerate Hold-collapse (PF erodes 1.45 → 1.08) — pivot to a research-
honest architecture: distributional supervised alpha + meta-labeling gate +
Coverage-Gated Kelly execution, integrated through a novel GPU-native
publish-subscribe Belief Bus substrate.
This is a DESIGN doc only. No implementation yet. v1 → v4 evolution captured
in the doc itself; v4 is research-honest with explicit prior-work citations:
- Bellemare/Dabney distributional RL (already shipped in foxhunt SP5+)
- Lopez de Prado triple-barrier + purging + meta-labeling
- Vovk/Romano/Gibbs-Candès conformal prediction foundations
- Sun-Yu 2025 NeurIPS CPTC (change-point-aware CP)
- Gan et al. 2025 NeurIPS arXiv:2510.26026 (CP for infinite-horizon RL —
we PORT Algorithm 1 directly in Phase 7, not invent)
- Zhu-Zhu ICML 2025 AlphaQCM (QCM variance estimation, adopted)
- Berti-Kasneci 2025 TLOB (motivates MLP baseline)
Honest novelty narrowed to three claims after literature review:
1. Belief Bus substrate — GPU-native pub/sub bus with per-slot
distributional semantics + conformal coverage + causal DAG metadata.
Extends our existing 539-slot ISV pattern (already novel architecture
vs published trading systems). The substrate integration is not in
literature.
2. Application domain — imbalance-bar HFT futures + MBP-10 microstructure +
triple-barrier labels. Existing distributional CP + DRL papers use
daily stocks, general RL benchmarks, or alpha formula discovery.
3. Adaptive controllers + per-slot conformal coverage — every adaptive
quantity in the system (Kelly priors, reward caps, Adam β1, regime
probabilities) gets conformal coverage attached. Not seen in
literature.
Tiered success criteria recalibrated per CFTC 2014 E-mini HFT study
(median firms hit ~55% WR / PF 1.2-1.4):
- Minimum viable: WR ≥ 50% AND PF ≥ 1.4 → deploy
- Goal: WR ≥ 53% AND PF ≥ 1.7
- Stretch: WR ≥ 55% AND PF ≥ 2.0 (original v1 target — aggressive
top-quartile HFT)
Eight phases with explicit falsification gates:
Phase 0: Purged walk-forward + bar audit (Lopez de Prado hygiene)
Phase 1a: MLP baseline alpha (cheapest falsification)
Phase 1b: TLOB/Mamba2/Liquid encoders
Phase 1C (conditional): tick-resolution if bar fails
Phase 2: Multi-head IQN + QCM + class weights
Phase 3: Belief Bus substrate
Phase 4: CPTC calibration
Phase 5: Coverage-Gated Kelly execution (deployment trigger if viable)
Phase 6: Production wiring + 2-week shadow mode
Phase 7 (optional): Port arXiv:2510.26026 conformal-DRL Q-residual
Phase 7 specifically detailed with concrete Algorithm 1 port (~1100 LOC
total), tunable params (k=5-10 ours vs 1-5 paper, due to γ=0.99 vs 0.8),
and falsification gate (empirical coverage ≥ 88% + PF improvement ≥ 0.2).
Deferred indefinitely (research-grade risk too high):
- Neural SDE (training instability per Kidger 2021)
- Hawkes process bar replacement (O(N²) MLE prohibitive at HFT scale)
- Multi-asset portfolio
- Learned in-trade exit head
Total minimum-viable path: Phases 0-6 ~6-8 weeks engineering + 20 hours
L40S compute. Falsification gates at every step.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
2a2f16b944