jgrusewski 8394e17224 feat: wire config weights to kernel + revert C51 alpha + atom warm-start spec ...

Config weights wired end-to-end (5 files): price_confirm_weight,
book_aggression_weight, hold_quality_weight, micro_reward_temp now
parsed from [reward] TOML section → DQNHyperparameters → GpuExperienceConfig
→ kernel args. No more hardcoded magic numbers in micro-reward formula.

Reverted c51_alpha_max 1.0→0.5: full C51 collapsed atoms to 3% util
at epoch 30 (death spiral). MSE floor prevents atom collapse.

PopArt warmup 100→10: 100 batches = ~5 epochs unnormalized → unstable.

Added bitonic sort integration spec: 4 uses (atom warm-start, robust
PopArt, experience curriculum, top-K PER).

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

2026-04-20 08:51:42 +02:00

..

2026-03-10-zero-cpu-hotpath-design.md

…

2026-03-11-dqn-factored-actions-design.md

…

2026-03-13-gpu-test-workflow-design.md

…

2026-03-14-kanidm-sso-design.md

…

2026-03-15-fused-cuda-training-design.md

…

2026-03-21-h100-epoch-optimization-design.md

…

2026-03-21-h100-epoch-optimization-phase3-design.md

…

2026-03-21-training-profiles-design.md

…

2026-03-22-gpu-composite-reward-design.md

…

2026-03-22-gpu-segment-tree-design.md

…

2026-03-22-two-phase-hyperopt-design.md

…

2026-03-23-9-exposure-action-space-design.md

…

2026-03-23-dual-distributional-c51-iqn-design.md

…

2026-03-23-profitability-roadmap.md

…

2026-03-23-trading-realism-conformance.md

…

2026-03-25-boundary-trade-counting-design.md

…

2026-03-25-trade-lifecycle-fix-design.md

…

2026-03-26-eval-cleanup-design.md

…

2026-03-27-bf16-performance-design.md

…

2026-03-27-bug-fixes-design.md

…

2026-03-27-code-hygiene-design.md

…

2026-03-27-full-bf16-rewrite-design.md

…

2026-03-31-feature-precompute-pipeline-design.md

…

2026-04-01-h100-gpu-optimization-design.md

docs: H100 GPU optimization spec — 9 optimizations for 2min→<10s epochs

2026-04-01 21:36:40 +02:00

2026-04-04-c51-loss-accumulator-fix-design.md

docs: spec + plan for C51 loss accumulator fix (separate MSE/C51 buffers)

2026-04-04 11:49:13 +02:00

2026-04-04-cuda-kernel-safety-design.md

docs: spec for CUDA kernel safety fix (__restrict__ aliasing NaN bug)

2026-04-04 11:07:47 +02:00

2026-04-04-feature-extraction-pipeline-fix-design.md

docs: spec for feature extraction pipeline fix (strict cache, symbol scoping)

2026-04-04 09:28:42 +02:00

2026-04-04-gpu-test-fixes-design.md

docs: spec for fixing 14 GPU test failures + OFI pipeline issue

2026-04-04 00:57:49 +02:00

2026-04-04-q-value-explosion-investigation.md

docs: spec for Q-value explosion / NaN investigation

2026-04-04 23:00:16 +02:00

2026-04-04-training-step-performance-design.md

docs: spec for training step performance + Q-value explosion investigation

2026-04-04 17:30:18 +02:00

2026-04-05-gpu-regime-and-training-hang-fix.md

docs: spec for GPU regime + H100 training hang fix

2026-04-05 21:25:05 +02:00

2026-04-06-gpu-native-validation-design.md

spec: GPU-native validation + H100 performance optimizations

2026-04-06 22:48:07 +02:00

2026-04-06-h100-maximum-performance.md

docs: H100 maximum performance spec and implementation plan

2026-04-06 11:15:24 +02:00

2026-04-06-per-prefix-sum-design.md

fix: promote entire backtest pipeline from bf16 to f32

2026-04-07 13:58:11 +02:00

2026-04-06-per-tree-fix-and-zero-sync-readback.md

docs: H100 maximum performance spec and implementation plan

2026-04-06 11:15:24 +02:00

2026-04-07-f32-backtest-pipeline-design.md

spec: f32 backtest pipeline — eliminate bf16 financial arithmetic

2026-04-07 13:43:04 +02:00

2026-04-07-reward-v7-design.md

feat(v8): wire 5 dead CUDA kernels + config fixes

2026-04-08 01:26:09 +02:00

2026-04-07-reward-v8-master-plan.md

docs: reward v8 master plan — 18-item comprehensive training overhaul

2026-04-07 23:14:40 +02:00

2026-04-07-reward-v71-fixes-design.md

docs: reward v7.1 spec — 6 critical fixes for exposure degeneracy and training quality

2026-04-07 22:21:11 +02:00

2026-04-08-advantage-standardization-design.md

fix(critical): disable mirror universe — only negates 4/42 features but inverts direction

2026-04-08 20:05:22 +02:00

2026-04-08-h100-epoch-optimization-v2.md

docs: H100 epoch optimization v2 spec — 47s → 19s (2.5x speedup)

2026-04-08 01:49:01 +02:00

2026-04-08-h100-graph-optimizations.md

docs: H100 graph-level optimization spec — 4 deferred fixes (~13s/epoch)

2026-04-08 08:48:36 +02:00

2026-04-08-hierarchical-4branch-design.md

docs: Phase 2 hierarchical 4-branch DQN spec — direction(3) × magnitude(3) × order(3) × urgency(3)

2026-04-08 12:04:23 +02:00

2026-04-08-per-branch-loss-weighting-design.md

feat(magnitude): per-branch loss weighting — 0.2× C51 + 4.0× MSE for magnitude

2026-04-08 22:23:19 +02:00

2026-04-08-volume-bar-pipeline-design.md

docs: volume bar pipeline spec + implementation plan

2026-04-08 20:26:40 +02:00

2026-04-10-argo-training-pipeline-redesign.md

spec: Argo training pipeline redesign — one workflow, smart caching

2026-04-10 20:26:55 +02:00

2026-04-10-dead-code-cleanup-gpu-profiling.md

spec: dead code cleanup + GPU per-phase CUDA event profiling

2026-04-10 22:16:24 +02:00

2026-04-11-autonomous-branch-intelligence-design.md

spec: Autonomous Branch Intelligence (ABI) — fix DQN action collapse

2026-04-11 11:23:23 +02:00

2026-04-11-candle-varstore-elimination.md

spec: Candle GpuVarStore elimination — 40+ call sites, 15 files, blocks bottleneck

2026-04-11 13:30:12 +02:00

2026-04-13-deterministic-forward-design.md

docs: spec for deterministic eval via training graph replay

2026-04-13 09:26:14 +02:00

2026-04-13-iql-full-integration-design.md

docs: fix 6 issues in IQL integration spec from code review

2026-04-13 14:51:30 +02:00

2026-04-13-iql-h100-hardening-design.md

docs: IQL H100 hardening spec — drift readiness, p5 floor, buffer-relative staleness

2026-04-13 20:13:39 +02:00

2026-04-13-single-cublas-handle-design.md

docs: spec for single shared cuBLAS handle refactor

2026-04-13 11:29:05 +02:00

2026-04-14-adaptive-training-dynamics-design.md

spec: Adaptive Training Dynamics v2 — 4-layer design for breaking Sharpe plateau

2026-04-14 21:40:03 +02:00

2026-04-14-cross-pollinated-branch-intelligence-design.md

spec: Component 6 early termination — PLATEAU_EXHAUSTED after 12 attempts

2026-04-14 23:46:27 +02:00

2026-04-14-direction-conditioned-magnitude-design.md

spec: direction-conditioned magnitude head (Layer 3)

2026-04-14 22:14:49 +02:00

2026-04-15-oos-generalization-enhancement-design.md

spec: add G11-G15 — Q-anchoring, predictive coding, Sharpe reward, confidence replay, commitment

2026-04-15 20:46:05 +02:00

2026-04-15-oos-performance-enhancement-design.md

spec: OOS Performance Enhancement — 8 components for closing val/OOS gap

2026-04-15 18:44:18 +02:00

2026-04-15-supervised-architecture-transfer-design.md

spec: Component 13 + full train-vpb4w results + backtracking bug analysis

2026-04-15 08:52:17 +02:00

2026-04-16-homeostatic-regularization-design.md

spec: trade-level reward attribution + exploration risk budget + homeostatic regularization

2026-04-16 01:24:03 +02:00

2026-04-16-introspective-state-vector-design.md

spec: Phase 4 adaptive minimum hold time — prevent coin-flip exits

2026-04-16 23:51:45 +02:00

2026-04-16-learned-risk-management-design.md

feat: trade-level reward attribution + learned risk management spec

2026-04-16 01:29:11 +02:00

2026-04-16-self-improving-training-loop-design.md

spec: Self-Improving Training Loop — 9 enrichments, eval-as-training

2026-04-16 17:30:01 +02:00

2026-04-16-tick-microstructure-intelligence-design.md

spec: fix 4 critical issues from review — OFI_DIM cascade, deploy atomicity

2026-04-16 23:43:55 +02:00

2026-04-16-trade-level-reward-exploration-design.md

spec: trade-level reward attribution + exploration risk budget + homeostatic regularization

2026-04-16 01:24:03 +02:00

2026-04-16-trade-plan-head-design.md

spec: Plan Phase 2 — 6 pearls/gems/novels for adaptive planning

2026-04-17 09:01:53 +02:00

2026-04-16-training-environment-alignment-design.md

spec: adapt components instead of removing — DSR on trade P&L, rank non-zeros only

2026-04-16 21:44:35 +02:00

2026-04-17-unified-cublas-training-graph-design.md

docs: update unified graph spec with Phase 1 H100 measurements

2026-04-18 14:59:57 +02:00

2026-04-18-multi-stream-phase2-design.md

spec: Phase 2 multi-stream parallelism design

2026-04-18 16:03:45 +02:00

2026-04-18-true-single-graph-design.md

spec: true single-graph — zero CPU on hot path, zero ungraphed launches

2026-04-18 23:59:35 +02:00

2026-04-18-unified-single-graph-design.md

spec: add Phase 3 performance optimizations (8 items, prioritized)

2026-04-18 22:04:41 +02:00

2026-04-19-kernel-optimization-design.md

spec: kernel optimization — cuBLAS projections + batch-parallel scans

2026-04-19 12:10:49 +02:00

2026-04-19-ofi-momentum-dense-reward-design.md

feat: target_dim 4→6 + spec v5 with pearls (bar duration, book CoM, retrospective hold)

2026-04-19 23:47:04 +02:00

2026-04-20-bitonic-sort-integration-design.md

feat: wire config weights to kernel + revert C51 alpha + atom warm-start spec

2026-04-20 08:51:42 +02:00

2026-04-20-data-pipeline-gaps-design.md

cleanup: remove legacy Sequential q_network + tune config for dense reward

2026-04-20 08:17:34 +02:00