Mass deletion of the "v7 gem" reward terms identified in the Phase 1
inventory as category errors — each one rewarded an outcome-adjacent
behavior instead of encoding the underlying physics, and each
empirically hurt validation metrics more than it helped training
stability.
Deleted from experience_kernels.cu and all plumbing (Rust configs,
launch args, hyperopt logs, TOML entries):
* order_credit_weight - reward redundant with compute_tx_cost
(order_type_idx already differentiates
fills by order type)
* risk_efficiency_weight - reward double-counted drawdown penalty
asymmetrically (only on winners)
* urgency_credit_weight - reward was vol-normalized unrealized P&L,
pure rename of core return
* commitment_lambda - triple-counted churn + tx_cost
* w_dsr - kernel wrote DSR EMA but no longer added
to reward (dead); removed the EMA
bookkeeping too
* dsr_eta - kernel arg for the deleted DSR EMA
* position_entropy_weight - rewarded action-bucket diversity
regardless of outcome; histogram buffer
+ zero-init removed too
* exit_timing_weight - already inactive (used raw_next future
price, comment-deleted earlier)
* ofi_reward_weight - dead plumbing; OFI already passed as
feature through state[OFI_START..]
* opportunity_cost_scale - penalized flat when Q-gap wide;
redundant with Q-values themselves
Kernel arg count: experience_env_step_batch shrank from ~55 to ~45 args.
Rust-side config surface reduced correspondingly.
Results on E1 smoke test (20-epoch):
BEFORE any Phase 2 work:
Val Sharpe -120 to -150, MaxDD 10-15%, Sharpe_raw -0.39
AFTER reward_noise + Kelly (both envs) + urgency + this sweep:
Val Sharpe -17 to -22 (7× better)
Val MaxDD 0.27% (40× better)
Val Sharpe_raw ~-0.09 (4× better)
Training Sharpe_raw ~0 (stabilized from ±20 swings)
Final q_gap 0.1712 (highest yet, collapse mechanism fine)
The extreme train-Sharpe swings (+17 one epoch, -13 next) were not
learning dynamics — they were shaping-term noise. Core reward (P&L +
drawdown + churn + holding + tx_cost + Kelly physics cap) gives training
metrics that actually reflect what the model does.
Inventory doc (docs/superpowers/specs/2026-04-21-phase1-reward-inventory.md)
extended with a "better-form taxonomy" section: every deleted gem has
a correct layer it belongs to (physics, feature, diagnostic, gradient-
level regularization — not reward). Kelly cap and Q-target smoothing
are already relocated; others are scheduled per the taxonomy's P1/P2/P3
priority list.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Phase 2 second relocation: the kelly_sizing_weight reward penalty
(experience_kernels.cu:1727-1746 — penalized deviation from Kelly-optimal
sizing) is deleted. Kelly is now a physics constraint in trade_physics.cuh:
the environment refuses to let the agent over-lever, not the reward
scoring the agent for matching a formula.
New helper in trade_physics.cuh (shared device function, reusable by
the forthcoming unified env kernel):
kelly_position_cap(win_count, loss_count, sum_wins, sum_losses,
max_position, safety_multiplier)
Applied in experience_kernels.cu between margin cap and execute_trade,
with health-coupled safety multiplier:
safety = 0.5 + 0.5 × health
- health=1 (healthy): full Kelly — trust the learned policy
- health=0 (collapsing): half Kelly — constrain when decisions less
reliable
Cold-start warmup (critical — otherwise balanced priors yield kelly_f=0
until real trades accumulate, starving Q-learning):
maturity = min(1.0, total_trades / 10)
effective_kelly = maturity × kelly_f + (1 - maturity) × 0.5
Early on (0 trades): cap dominated by 50% floor.
As real trades accumulate (10+): pure data-driven Kelly.
Validation env (backtest_env_kernel.cu) does NOT yet get the Kelly cap —
that requires extending its portfolio state or adding a separate
kelly_stats buffer, which naturally belongs in the Phase 3 unified env
kernel refactor. The current asymmetry is a KNOWN temporary — training
is constrained, validation is not — and will be resolved when both
kernels share the same env_step() device function.
Also completes removal of kelly_sizing_weight from all plumbing:
- experience_kernels.cu: kernel arg deleted
- gpu_experience_collector.rs: launch arg, config field, default
- training_loop.rs: hyperparam propagation
- config.rs: field, default, intensity clamp (with tombstone)
- hyperopt/adapters/dqn.rs: log reference
- config/training/*.toml (6 entries across 4 files): orphan configs
(none were wired to a profile parser field)
Verification:
- cargo check -p ml --lib clean
- E1 smoke test passes: final epoch q_gap=0.1109, health=0.51 (warmup
floor of 0.5 gives early exploration enough room; floor of 0.25
was too tight and failed at q_gap=0.0496)
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Phase 2 kick-off from the env-unification design. First relocation: the
reward_noise_scale field that perturbed training rewards is deleted, and
its regularization effect moves to the correct layer — Q-target label
smoothing in c51_loss_kernel.cu — now health-coupled rather than fixed.
Before:
- reward += pseudo_noise × max(|reward| × 0.05, 0.01) (in env reward path)
- Q-target label smoothing = fixed LABEL_SMOOTHING_EPS = 0.01
After:
- Reward untouched by noise. Core reward = actual outcome + aligned penalties.
- Q-target label smoothing eps_eff = 0.02 × (1 − health) read from ISV[12]
- health=1 (healthy): eps_eff=0, sharp targets preserved
- health=0.5: eps_eff=0.01, matches old fixed behavior at mid-health
- health=0 (collapsing): eps_eff=0.02, maximum regularization prevents
overcommitment to the collapsed distribution
Why health-coupled:
Same insight as the distillation SAXPY fix — every fixed kernel scalar is
a temporal-coupling candidate when we have the ISV pinned buffer available.
Regularization strength should scale INVERSELY with network health: it's
most needed exactly when things are falling apart.
Files touched:
- c51_loss_kernel.cu: LABEL_SMOOTHING_EPS const replaced with
LABEL_SMOOTHING_BASE + in-kernel health read from isv_signals[12]
- experience_kernels.cu: deleted reward noise block + kernel arg
- gpu_experience_collector.rs: dropped launch .arg + config field + default
- training_loop.rs: dropped hyperparam propagation
- config.rs: deleted field + intensity clamp + default (with tombstone)
- hyperopt/adapters/dqn.rs: dropped log reference
- config/training/*.toml (4 files): dropped orphan reward_noise_scale
entries (none were being parsed — the profile parser had no field)
Verification:
- `cargo check -p ml --lib` clean
- E1 smoke test passes: final q_gap=0.1190, health=0.52 (health-coupled
smoothing at ~mid-health matches old fixed behavior, collapse-prevention
mechanism intact)
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Atom warm-start: bitonic sort rewards → quantile positions → write to
atom_positions_buf as initialization. Existing SGD optimizer refines.
Atoms start where reward mass actually is instead of uniform [-50,+50].
Robust PopArt: median/IQR normalization from sorted rewards replaces
Welford mean/var. More robust for bimodal distribution (many ±0.1
micro-rewards + few ±5.0 trade exits). Conditional: popart_robust=true.
Both reuse the same bitonic sort (~14ms per epoch, amortized).
gather_quantiles kernel extracts positions. extract_median_iqr reads
Q25/median/Q75 from sorted buffer.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Counterfactual: Hold(1)/Flat(3) direction mirror maps to self — trivial.
Now falls through to magnitude CF for Hold/Flat instead of wasting a
replay buffer slot on same-action same-reward experiences.
reward_noise_scale: 0.05→0.0 (dense micro-rewards are already noisy,
adding 5% label noise destroys the per-bar signal).
Added micro-reward weight config fields (price_confirm_weight=0.5,
book_aggression_weight=0.3, hold_quality_weight=0.2, micro_reward_temp=3.0,
holding_cost_rate=0.0001) — defined in TOML, kernel plumbing next session.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Removed the legacy non-branching Sequential q_network and target_network
from DQNAgent. These were never used (branching+dueling always active)
but allocated VRAM and ran noise resets every step. -190 lines.
Config tuning for dense micro-reward system:
- n_steps: 5→1 (TD(0), micro-rewards cancel over n>1)
- tau: 0.007→0.01 (faster target tracking for TD(0))
- c51_alpha_max: 0.5→1.0 (full C51, PopArt handles normalization)
- curiosity_weight: 0.1→0.0 (dense micro-reward replaces curiosity)
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Three signal-killing issues fixed:
1. Rank normalization DISABLED — was double-normalizing with PopArt,
destroying magnitude difference between micro-rewards (±0.1) and
trade exits (±5.0). PopArt alone preserves relative magnitude.
2. n_steps 5→1 (TD(0)) — dense micro-rewards alternate ±0.1 each bar.
With n=5, they cancel out over 5 bars. TD(0) preserves the per-bar
signal that the temporal pipeline needs to learn from.
3. per_alpha 0.6→0.3 — lower = more uniform PER sampling. Dense micro-
rewards have tiny TD-errors (easy to predict), so high alpha ignores
them. Lower alpha ensures micro-reward experiences get sampled.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Linear scaling rule (2x batch → 2x LR) doesn't hold for DQN+PER with
major architectural changes (Hold action, OFI embed, wider attention).
The model needs stability to learn the new action space, not speed.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Linear scaling rule: 2x batch → 2x LR to maintain effective update
magnitude. Tau increased to compensate for fewer steps/epoch (target
network tracks faster). H100 VRAM: 41GB free at B=8192, B=16384 adds
~4GB — easily fits.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Default was hardcoded as 64 in train_baseline_rl.rs and evaluate_baseline.rs,
overriding the config default of 32. Added num_quantiles=32 to production
config so it's explicit.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Auto-sizer inflated replay buffer to 15.8M entries on H100 (45% of 80GB VRAM).
The PER prefix scan inside the CUDA graph processes ALL 15.8M entries every step,
dominating step time. With buffer_size=500K, the scan processes 500K entries instead.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
At 16384, each GEMM saturates 97% of 132 SMs — no room for multi-stream
parallelism. At 8192, GEMMs use ~50% SMs, allowing branch streams and
aux parallelism to fill idle SMs. 2x more steps/epoch but each ~2x faster.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Removed: enforce_hold(), min_hold_bars from config/kernels/backtest.
Added: holding_cost_rate (inventory penalty), churn_threshold_bars +
churn_penalty_scale (graduated flip penalty) to reward in
experience_env_step and backtest kernels.
The model learns optimal hold timing from cost signals:
- Per-trade tx cost prevents churning (existing)
- Inventory penalty makes large positions expensive to hold
- Churn penalty graduates cost for rapid flips
- Temporal attention learns when holding cost > expected profit
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
min_hold=10 allowed ~1200 trades/day — costs destroyed the edge.
min_hold=50 (~6.5 min) limits to ~230 trades/day max.
Adaptive extension now scales 0-2× base (was hardcoded 0-8 bars).
With base=50: range is 50-150 bars depending on ISV stability.
Losing positions (negative reward_ema) get base only (50 bars).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The warp-parallel (32,1,1) change to c51_loss_reduce caused the same
hang as isv_feature_gate: changing block_dim inside graph_mega alters
CUDA Graph node structure on Hopper's TMA scheduler. Must stay (1,1,1).
Also: batch_size 8192→16384, gpu_n_episodes 1024→4096, num_atoms 51→52
to align h100.toml with dqn-production.toml.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
mbp10_data_dir and trades_data_dir uncommented in localdev config.
Doc comments updated: 20 microstructure features always appended.
OFI is no longer optional — it's core to the model's feature set.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
NoisyLinear was completely dead after cuBLAS migration — neither training
nor experience collection applied ANY exploration noise. With identical
Q-values across actions, Boltzmann selection was uniform, making the
dueling mean-subtraction cancel ALL advantage gradients. The advantage
heads could NEVER learn (Q-gap permanently 0.0000).
Fix: add_advantage_noise CUDA kernel injects per-action Gaussian noise
into Q-values AFTER compute_expected_q, BEFORE action selection.
- Philox PRNG + Box-Muller for GPU-native Gaussian sampling
- Per-sample, per-action noise (breaks within-batch symmetry)
- Only during experience collection (not training targets)
- noise_sigma=0.1 (configurable, TOML + hyperopt tunable)
The noise creates non-uniform Boltzmann selection → different actions
have different frequencies → advantage gradient survives the dueling
mean-subtraction → Q-gap can grow.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Q-gap = max_a Q(s,a) - mean_a Q(s,a): measures action preference per state.
Q-var = Var_s[Q(s,a)]: measures state differentiation across the batch.
Both are 0.0000 through all 10 smoketest epochs despite Sharpe 8+ and
Q-values growing to 0.012. This proves the good metrics are from
reward-induced mechanical bias, NOT learned Q-values.
Also: c51_warmup_epochs=0 (C51 from step 1, bypass MSE dead zone),
smoketest lr=1e-4 (50× higher, makes 40 steps representative).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
With per-component clipping removed, raw gradient norms are ~4000.
The old gradient_clip_norm=1.0 (calibrated for pre-clipped norms ~7)
discarded 99.97% of gradient magnitude every step, making the effective
learning rate 570× too small. Q-values grew 9× slower than old run.
10000 lets normal gradients (~4000) pass through unclipped. Only fires
on genuine divergence spikes (>2.5× normal). Adam's internal adaptive
step sizing handles the rest.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Fundamental fix: C51 atom spacing (dz=0.6) was larger than the Q-value
range (0.02), making the distributional loss unable to resolve rewards.
The network was blind to its own learning signal.
Adaptive v_range via device buffer (same pattern as tau_buf):
- v_range_buf[2] on GPU, all C51/MSE/CQL/expected_q kernels read via
pointer (graph captures stable address, reads value at replay time)
- Cold start: v_range=±1.0 → dz=2/51=0.039 → 25× atom resolution
- Monotonic expansion based on Q-stats every 50 training steps
- Never shrinks (avoids invalidating Bellman targets in replay buffer)
Kernel changes (5 files):
- c51_loss_kernel.cu, mse_loss_kernel.cu, mse_grad_kernel.cu,
cql_grad_kernel.cu, compute_expected_q: scalar v_min/v_max → pointer
Also fixed:
- PopArt warmup 100→1 (start normalizing from step 1)
- PopArt variance floor 0.0001→1e-8 (allow sigma < 0.01)
- Backtest evaluator: allocate own v_range_buf for compute_expected_q
- num_atoms 51→52 in all TOMLs (TF32 4-element alignment)
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
v_min/v_max ±15 with gamma=0.99 only covers 15% of theoretical Q range
(Q_max = 1/(1-0.99) = 100 for unit-variance PopArt rewards). C51 top atom
saturates on sustained winners, degrading distributional learning. ±50
covers 95th percentile.
PopArt GPU buffers (popart_mean, popart_var, popart_count) were never
zeroed between walk-forward folds — fold 2's reward normalization was
contaminated by fold 1's statistics. Now reset alongside Adam state in
reset_adam_state().
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Added OwnedGpuLinear to ml-core (self-contained linear layer with owned
weight/bias CudaSlice, forward_with_slices bypasses GpuVarStore lookup)
- Removed GpuVarStore from all 12 ml-dqn modules: Sequential, branching,
distributional_dueling, dueling, network, rmsnorm, residual, curiosity,
attention, iql, quantile_regression, regime_conditional
- Removed get_q_network_vars/vars/store accessors from DQN, branching,
distributional_dueling, dueling, network, quantile_regression
- Replaced GpuLinear+GpuVarStore with OwnedGpuLinear in all cold-path modules
- Made load_from_safetensors a no-op (flat buffer path handles checkpoints)
- Made sync_to_varmap a no-op (flat buffer is source of truth)
- Moved branching->VarStore conversion to gpu_weights::branching_to_varstore
- Re-enabled bottleneck_dim=16 in config defaults + all 3 TOML configs
- Removed flatten_online_weights bottleneck skip workaround
- Zero GpuVarStore references in crates/ml-dqn/src + crates/ml/src/trainers/dqn
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
bottleneck_dim=16 causes w_s1 size mismatch between VarStore (full state_dim)
and flat params_buf (reduced bottleneck_concat_dim). EMA target sync crashes
with CUDA_ERROR_ILLEGAL_ADDRESS.
Root cause: GpuVarStore allocates weights at full state_dim but GpuDqnTrainer's
flat buffer uses bottleneck-reduced dimensions. These two weight sources are
fundamentally incompatible when bottleneck is active.
Fix: eliminate GpuVarStore entirely (next task), then re-enable bottleneck.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Task 3: Enable PopArt reward normalization (default true, all 3 TOML configs,
wiring confirmed in fused_training.rs submit_forward_ops_main()).
Task 4: Couple PopArt variance with tau reset — add prev_popart_var to
FusedTrainingCtx, read_popart_var() to GpuDqnTrainer, read_popart_variance()
+ should_reset_tau() to FusedTrainingCtx, tau reset injected at epoch boundary
in training_loop.rs after log_phase_timing().
Task 5: Add best_sharpe/best_epoch/best_val_loss reset to reset_for_fold() so
each walk-forward fold competes independently. Also wire v8 reward fields
(popart_enabled, micro_reward_scale, td_lambda, max_trace_length,
hindsight_fraction, hindsight_lookahead) through training_profile.rs apply_to().
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- iqn_lambda restored to 0.25 (was 0.0 for NaN isolation)
- STEP_DIAG per-step logging removed (was temporary)
- FOXHUNT_GRAD_DIAG env var removed from Argo template
IQN NaN root cause fixed in 8cbabcef5 (f32-as-bf16 type mismatch).
Ready for H100 baseline deployment.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Graph recapture ruled out as NaN cause (NaN persists without graph
invalidation). Now testing: does NaN disappear when IQN is disabled?
If yes → IQN backward/Adam produces NaN on H100.
If no → NaN is in the main C51/MSE pipeline.
Also reverts graph invalidation hack from previous diagnostic commit.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The f32→bf16 cast in cast_d_logits_to_bf16() reads
b*(b0+b1+b2+b3)*na + 32*4 elements, but d_adv_logits_buf and
d_adv_logits_bf16 were allocated with only +32*3 padding (96 vs 128).
The 32 extra f32 reads went past the allocation into adjacent GPU
memory. On H100 with 80GB and different memory layout, this grabbed
NaN/garbage values. The NaN propagated through the cuBLAS backward
pass into grad_buf, killing training at step ~720 (non-deterministic
depending on what lives in adjacent memory).
Fix: allocate +32*4 padding in d_adv_logits_buf, d_adv_logits_bf16,
and d_adv_logits_mse to match the cast kernel's access pattern.
Also reverts c51_alpha_max default to 1.0 — the gradient collapse
was from this buffer bug, not C51 premature convergence. The
c51_alpha_max parameter is kept for hyperopt exploration.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
H100 baseline (20 epochs) showed gradient collapse at epoch 10: C51
cross-entropy converges its distributional fit before the policy converges,
leaving zero gradient signal. The collapse happened 5 epochs after C51
reached alpha=1.0 (pure C51, zero MSE).
Fix: cap the C51 alpha ramp at c51_alpha_max (default 0.5) so MSE always
contributes (1 - alpha_max) of the primary gradient. MSE loss measures
Q-error directly and only goes to zero when Q-values are correct, not
just when the distribution shape is right.
- c51_alpha_max added to DQNHyperparameters (default 0.5)
- Added to PSO search space as 15th dimension (range [0.3, 0.9])
- Added to TOML profiles: smoketest, production, hyperopt
- Training loop caps alpha ramp at alpha_max instead of 1.0
- All 907 ml tests + 300 ml-core tests + 6 smoke tests pass
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Three root causes found and fixed:
1. SUM-reduced gradients without 1/N: all CUDA loss gradient kernels
(C51, MSE, IQN backward, CQL) accumulated per-sample gradients as
raw SUM. At batch=16384 (H100) the raw norm was 282x larger than
batch=58, causing gradient clipping to destroy signal-to-noise ratio
and collapse training at epoch 2-3. Now all kernels multiply by
1/batch_size, making gradient scale batch-invariant.
2. ExposureLevel::target_exposure() used a flat 9-level scale that did
not match the 4-branch dir*mag encoding. The Rust backtest evaluator
computed wrong position sizes (e.g. 4x oversize for Short+Small).
Now uses dir x mag formula. Also fixed is_buy/is_sell/is_hold and
from_trading_action for 4-branch semantics.
3. Rust epsilon-greedy only explored 5/9 exposure combos (0..5 instead
of dir*3+mag), ignored the magnitude branch on greedy, and used wrong
indices for order/urgency (get(1)/get(2) instead of get(2)/get(3)).
LR recalibrated: old gradient_clip_norm was accidentally a batch-size-
dependent LR reducer (~60x at batch=58, ~16000x at batch=16384). With
mean-reduced gradients the clip rarely fires, so LR is now the sole
training speed control. Smoketest 1e-4 -> 2e-6, production 1e-4 -> 1e-5,
hyperopt range [1e-5,3e-4] -> [1e-7,1e-4].
Diagnostics: FOXHUNT_GRAD_DIAG=1 enables per-stage gradient norm logging.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Three fixes:
1. spread_cost removed contract_multiplier ($50) — was computing in dollars
but PnL is in points, making spread 12.5× actual. The model saw every
trade as guaranteed loss → preferred Flat.
2. Mirror universe now negates ALL 17 directional features (returns, MACD,
Bollinger, SMA ratios, regression slope, CUSUM) + flips RSI. Was only
negating 4, teaching wrong direction on 50% of epochs.
3. tx_cost_multiplier updated to 0.18 bps (IBKR ES RT=$4.50/contract).
min_hold_bars=10 for volume bars (~26s at 23 bars/min).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
With separate optimizer, no gradient competition — safe to increase.
Smoke test shows within-group differentiation starting (S100≠S25≠L50).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The exposure aux backward GEMM amplifies gradients through the
hidden activation matrix (d_logits × h_b0), producing grad_norm=5000+.
With max_grad_norm=10, this throttles ALL gradients by 500x (lr/500).
Fix: reduce aux_weight 0.5→0.01 (50x) + clip_grad_buf_inplace after
aux GEMM to cap combined gradient at max_grad_norm.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
