jgrusewski bc9eaac89d fix(rl): B-6 — ISV-driven adaptive asymmetric Wiener-α (Bayesian shrinkage) ...

B-5 (asymmetric α with static α_slow=0.001) revealed the static parameter
problem: provably bounds cascades (avg_win peak $2k vs B-4's $40k) BUT
over-conservative in train (dckcc step 800: avg_l > avg_w → Kelly says
don't trade → model can't discover edges; wr_ema crashed to 0.145).

The fundamental tension: static α_slow can't satisfy BOTH
  - Train convergence: asymmetry must FADE so model learns from real data
  - Boundary safety: asymmetry must ENGAGE at every fold to prevent cascade

B-6 RESOLVES this via Bayesian shrinkage:

  trust(n)   = min(1, cum_dones / n_full_threshold)           [Phase 1]
  stability  = exp(-CV × cv_gain)                              [Phase 2]
  trust_eff  = trust(n) × stability
  α_slow_eff = α_slow_min + (α_fast − α_slow_min) × trust_eff

Phase 1 (data-quantity): trust grows with cum_dones; reset_session_state
zeroes cum_dones → asymmetry RESUMES at every boundary. Math: at n=0
α_slow_eff = α_slow_min = 0.001 (full skepticism). At n=n_full = 30k
trades: α_slow_eff = α_fast = 0.05 (full standard Wiener).

Phase 2 (data-quality): Welford CV of reward magnitude gates trust. Stable
signal (CV→0): stability=1, trust opens normally. Volatile signal (CV high):
stability→0, asymmetry persists. cv_gain=0 disables Phase 2.

Per-EMA asymmetry direction encodes Kelly safety semantics:
  avg_win:  slow-up (skeptical of wins),    fast-down
  avg_loss: fast-up (admit losses),         slow-down (slow forget)
  wr_ema:   slow-up (skeptical of high WR), fast-down

ISV slots (all signal-derived from cum_dones + Welford):
  721 RL_EMA_ALPHA_SLOW_MIN_INDEX           = 0.001
  722 RL_EMA_TRUST_FULL_THRESHOLD_INDEX     = 30000
  723 RL_EMA_CV_GAIN_INDEX                  = 1.0

Threshold calibration (n_full=30k):
- Train: ~1000 cluster steps for trust to fully open → asymmetry active
  during cold-start (first 30 steps, cascade prevention) then fades.
- Eval: 30 dones/step × 500 eval steps = 15k dones → trust climbs to 0.5
  by eval end → partial protection throughout eval.

Composes:
- B-3 Kelly fractional-trust (Kelly SIZING gated by cum_dones)
- B-6 EMA asymmetric-α (Kelly INPUTS biased conservative by cum_dones)
Both fade as data accumulates; both reset at boundary.

Spec: docs/superpowers/specs/2026-06-01-ema-asymmetric-trust-with-cv-gain.md

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

2026-06-01 01:13:27 +02:00

..

fixtures

fix(ml-core): deterministic GPU weight init via scoped_init_seed

2026-05-19 01:19:13 +02:00

adamw_invariants.rs

feat(ml-alpha): BCE multi-horizon + AdamW kernels

2026-05-16 21:57:26 +02:00

aux_heads.rs

feat(aux-heads): rewrite for A+B paired (BCE + conditional-Huber) (CB3+CB4)

2026-05-22 11:16:29 +02:00

aux_trunk.rs

feat(aux-trunk): smaller single-bucket CfC for aux supervision (Layer B3)

2026-05-22 09:03:00 +02:00

backward_finite_diff.rs

feat(ml-alpha): cfc_step_backward emits grad_x for upstream chain

2026-05-16 22:59:36 +02:00

bce_grad_finite_diff.rs

refactor(ml-alpha): consolidate BCE — Kendall σ is THE BCE, wire MHA into trainer

2026-05-18 14:59:12 +02:00

bucket_routing.rs

feat(per-horizon-cfc): add bucket_routing.rs (Controller A + transition dispatch)

2026-05-21 16:18:04 +02:00

bucket_transition_kernels.rs

Revert "fix(per-horizon-cfc): extend block-diagonal mask to heads_w1 (Smoke 2 fix)"

2026-05-21 23:48:15 +02:00

cfc_step_bit_equiv.rs

feat(ml-alpha): cfc_step kernel + invariant-based GPU validation

2026-05-16 21:50:20 +02:00

cfc_step_per_branch.rs

fix(per-horizon-cfc): ALPHA — skip reorder, channels_in_bucket lookup, Adam moment masking

2026-05-21 20:43:51 +02:00

controller_adaptive_floors.rs

feat(rl): adaptive controller floors — 12 controllers, all signal-driven

2026-05-30 15:41:51 +02:00

dqn_toy.rs

feat(rl): R6 — replace LobEnv with kernel-driven GPU-pure env step

2026-05-23 10:39:25 +02:00

encoder_gradient.rs

feat(rl): expose forward_encoder() for RL head consumption (Phase B)

2026-05-22 22:28:19 +02:00

eval_diag_emission.rs

fix(rl): B-6 — ISV-driven adaptive asymmetric Wiener-α (Bayesian shrinkage)

2026-06-01 01:13:27 +02:00

forward_graph_capture.rs

feat(ml-alpha): CUDA Graph capture of forward_only (P5)

2026-05-19 17:32:28 +02:00

forward_step_golden.rs

feat(per-horizon-cfc): forward_step_into uses Phase 2 fused dispatch

2026-05-21 19:10:22 +02:00

frd_head.rs

test(ml-alpha): migrate integration tests to post-Phase-4 trainer API

2026-05-30 11:52:36 +02:00

gpu_log_ring_invariants.rs

refactor(per-horizon): N_HORIZONS 5→3 — remaining ml-alpha tests

2026-05-22 01:45:45 +02:00

heads_bit_equiv.rs

test(ml-alpha): migrate integration tests to post-Phase-4 trainer API

2026-05-30 11:52:36 +02:00

heads_block_diagonal.rs

feat(per-horizon-cfc): add heads_block_diagonal_fwd.cu (compact ragged w_skip)

2026-05-21 16:04:29 +02:00

integrated_trainer_smoke.rs

fix(rl): R9 local-smoke prep — two test-fixture bugs

2026-05-23 13:36:03 +02:00

iqn_oracle.rs

perf(rl): fuse adamw_increment into adamw_step + tau into iqn_forward

2026-05-26 16:46:10 +02:00

isv_bootstrap.rs

test(ml-alpha): migrate integration tests to post-Phase-4 trainer API

2026-05-30 11:52:36 +02:00

isv_bus.rs

…

multi_horizon_loader.rs

refactor(rl): FRD horizons + range_σ are ISV-driven, not literals

2026-05-24 19:23:39 +02:00

multi_resolution_pipeline.rs

test(ml-alpha): synthetic multi-resolution pipeline test

2026-05-22 20:46:55 +02:00

noisy_oracle.rs

test(rl): GPU oracle tests for IQN + NoisyNet + PRNG advance fix

2026-05-25 22:37:40 +02:00

output_smoothness_grad_finite_diff.rs

refactor(per-horizon): N_HORIZONS 5→3 — remaining ml-alpha tests

2026-05-22 01:45:45 +02:00

perception_forward_golden.rs

fix(ml-core): deterministic GPU weight init via scoped_init_seed

2026-05-19 01:19:13 +02:00

perception_overfit.rs

feat(aux-supervision): wire BCE + conditional-Huber actual calls (CB5)

2026-05-22 11:32:19 +02:00

pinned_slots.rs

feat(ml-alpha): mapped-pinned ingress slots for snapshot + fill

2026-05-16 21:42:28 +02:00

ppo_toy.rs

feat(rl): R6 — replace LobEnv with kernel-driven GPU-pure env step

2026-05-23 10:39:25 +02:00

projection_bit_equiv.rs

feat(ml-alpha): projection kernel (128->8 + layer-norm)

2026-05-16 21:52:57 +02:00

r3_ema_advantage.rs

test(ml-alpha): migrate integration tests to post-Phase-4 trainer API

2026-05-30 11:52:36 +02:00

r4_action_kernels.rs

feat(rl): R7d — PER wired + off-policy DQN with stop-grad on encoder

2026-05-23 12:59:42 +02:00

r5_controllers_and_soft_update.rs

test(ml-alpha): migrate integration tests to post-Phase-4 trainer API

2026-05-30 11:52:36 +02:00

risk_stack_invariants.rs

feat(rl): IQN τ tail-recency consumer + G24/G25 invariants (F5)

2026-05-31 12:59:48 +02:00

signal_variance_kernel.rs

feat(rl): adaptive controller floors — 12 controllers, all signal-driven

2026-05-30 15:41:51 +02:00

smoothness_lambda_controller_invariants.rs

refactor(per-horizon): rewrite smoothness controller test for N_HORIZONS=3

2026-05-22 01:07:33 +02:00

snap_feature_bit_equiv.rs

Revert "arch(ml-alpha): restore count-delta redundancy at feature slot [26]"

2026-05-18 23:16:38 +02:00

trade_management_kernels.rs

feat(rl): adaptive risk-management stack — 5 layers, all ISV-driven

2026-05-30 20:52:28 +02:00