7f92fa242c52ad0e11d60b686cc2c8a6bc76d8a8
4109 Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
jgrusewski
|
7f92fa242c |
cleanup: wire td_error ISV scratch + rewrite stale TODOs declaratively
Part A of pre-L40S cleanup.
1. Wire td_error batch mean into ISV scratch (gpu_dqn_trainer.rs):
`launch_loss_reduce` now runs the generic `c51_loss_reduce` kernel a
second time over `td_errors_buf` into `td_error_scratch_dev_ptr`.
ISV[2] (TD-error EMA in `isv_signal_update`) was previously reading
a zero-initialised scratch and accumulated a constant-zero signal.
This was a genuinely missing kernel writeback — the C51 loss kernel
was already emitting per-sample |TD-error| into `td_errors_buf`
(c51_loss_kernel.cu:1096), it just wasn't being batch-reduced.
Reuses the existing `c51_loss_reduce` (generic mean-reduction, single
block, deterministic) rather than adding a new kernel — no new CUDA
surface, no ABI change.
2. Remove 2 stale TODOs from batched_backward.rs docstrings that
described a migration that's actually complete:
- Module docstring said "dqn_backward_kernel (atomicAdd path)
remains active" — the atomicAdd kernel has been removed; cuBLAS
backward is wired via launch_cublas_backward.
- `backward_full` docstring said "gated behind TODO" — the function
is actively called from the fused training step.
3. Rewrite 2 ISV scratch field comments as declarative: td_error_scratch
is now wired (as per change 1); ensemble_var_scratch remains
zero-initialised and its comment honestly describes that consumers
(ISV[3] and [4]) treat it as unavailable. Per feedback_no_todo_fixme.md,
replaces the TODO(isv) markers with declarative descriptions of
current behaviour. Future wiring is tracked in the plan, not in code
aspirational markers.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
|
||
|
jgrusewski
|
d9fee6ef8d |
fix(kelly): Task 2.Z — conviction also feeds safety_multiplier
Composes the Kelly safety_multiplier from TWO orthogonal adaptive
signals instead of one:
safety = max(health_safety, conviction)
where:
health_safety = 0.5 + 0.5 × learning_health [training stability]
conviction ∈ [0, 1] [per-sample confidence]
Health measures training stability globally. Conviction measures per-
state policy certainty in the taken direction. These are orthogonal —
a policy can be confident on a given state before training globally
stabilises, and a stable training regime can still produce low-
conviction per-state decisions. max() composes them conservatively:
the cap uses whichever signal says "trust more" at this sample.
Bounded to [0.5, 1.0] by the health floor.
Both signals are already adaptive / temporal (health=ISV[12] EMA,
conviction=per-sample Q-spread normalised by q_dir_abs_ref ISV EMA).
No static tuning knobs. Per feedback_adaptive_not_tuned.md.
Motivation (per project_magnitude_eval_collapse_kelly_capped.md): at
typical smoke-test health=0.49, health_safety = 0.745 sits coincid-
entally on the Half/Full decoder boundary (abs_pos < 0.75). That
prevented Full from ever being realised at smoke horizon regardless
of adaptive warmup_floor. Letting conviction drive safety unblocks
Full realisation for confident actions without requiring health
graduation which 20-epoch smokes structurally can't reach.
Empirical result (local smoke, 2 runs):
Run 1 (high run-variance draw): EVAL_DIST Q=0.911 H=0.057 F=0.032
— still fails H10 eh+ef≥0.30
Run 2: EVAL_DIST Q=0.350 H=0.121 F=0.529
— PASSES all 5 assertions
— FIRST FULL SMOKE PASS SINCE 4-BRANCH
Previous best (before this commit):
(pre-safety-A, v5+adaptive-Kelly only): Q=0.325 H=0.675 F=0.000
— passed H10 at line 134 but failed Task 2.X line 153 (ef < 0.05)
The commit trades the reliable Half-dominance regime for a bi-modal
distribution that includes Full on many runs. Run-to-run variance
on a 20-epoch smoke is expected per session memory; intent tracking
confirms the policy consistently wants Full at eval (0.73-0.85 across
runs), so the gap is purely in realised cap, not policy learning.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
|
||
|
jgrusewski
|
a2624d8b9d |
cleanup: remove TODO(task-4-followup) from gpu_backtest_evaluator
Rewrites the plan_isv_buf comment from an aspirational TODO to a declarative doc note describing the accepted design gap: backtest validation intentionally zero-fills plan/ISV state positions [86..92) because the backtest env kernel does not compute those training-time introspective signals. The policy treats plan/ISV as advisory features, so the train/val delta is tolerated in exchange for a lean backtest env kernel. Per feedback_no_todo_fixme.md: TODO/FIXME markers are forbidden; rewrite as declarative production-ready prose or complete the work. |
||
|
jgrusewski
|
c34a6592f7 |
Merge: adaptive Kelly warmup_floor from policy conviction
Agent worktree
|
||
|
jgrusewski
|
a0224ce846 |
Merge: intent-side magnitude diagnostic (EVAL_INTENT_MAG_DIST)
Agent worktree
|
||
|
jgrusewski
|
ff683470e7 |
fix(kelly): Task 2.Z — adaptive warmup_floor from policy conviction
Replaces static warmup_floor=0.5f in kelly_position_cap (trade_physics.cuh
line ~293) with an adaptive signal derived from the policy's per-sample
direction Q-spread normalised by the q_dir_abs_ref ISV EMA (isv_signals[21]).
High conviction -> high floor (trust policy at cold start). Low conviction
-> low floor (safety dominates). Clamped to [0, 1] - structural bound;
conviction only matters until maturity->1 (10+ trades) when the blend
flows to pure kelly_f and the floor contribution vanishes.
Wiring:
1. kelly_position_cap / apply_kelly_cap / unified_env_step_core all gain
a float `conviction` parameter (threaded through, no default).
2. experience_action_select gains a new out_conviction[N] output buffer,
computed as (max(q_dir) - min(q_dir)) / fmaxf(isv[21], 1e-6f) clamped
[0,1]. Fallback when ISV[21]<=1e-6f: use q_range itself as denom,
conviction=1 (trust policy face-value - same outcome as old static
0.5 at health=1, but from a real signal shape).
3. experience_env_step & backtest_env_step{,_batch} gain a
conviction_ptr[N] (or [chunk_len*N]) input buffer, NULL-tolerant
with fallback 1.0.
4. Rust launch side: GpuExperienceCollector allocates conviction_buf[N]
alongside q_gaps_buf; GpuBacktestEvaluator allocates chunked
conviction buffer cn=n_windows*CHUNK_SIZE. Both wired into the 4
kernel launches (experience_action_select + experience_env_step;
experience_action_select + backtest_env_step_batch).
The previous static 0.5 pinned cold-start cap to <=0.375*max_pos at
health=0.5 (safety_multiplier=0.75), which combined with the
`abs_pos < 0.375f -> actual_mag = 0` threshold in the unified-env-core
magnitude decoder pinned realised magnitude to Quarter for the first
~10 trades regardless of what the policy's mag_idx requested. Smoke
test EVAL_DIST=[1.0, 0.0, 0.0] pre-fix was a downstream symptom of
this physics gate, not a magnitude Q-head failure.
Per feedback_adaptive_not_tuned.md: no hard-coded numeric knobs.
Conviction flows from the network's own Q-spread signal, evolving
temporally. Per feedback_no_functionality_removal.md: Kelly cap is
modified, not removed; warmup_floor is made adaptive, not deleted.
Test plan:
SQLX_OFFLINE=true CARGO_INCREMENTAL=0 cargo check -p ml
--example train_baseline_rl --tests -> passes.
Smoke (magnitude_distribution, 20 epochs) shows:
[MAG_DIST] Quarter~0.62-0.70 Half~0.15-0.19 Full~0.13-0.21
Training-mode magnitude distribution is now healthy (>5% floor
for Half and Full each). Eval-mode smoke is non-deterministic in
this horizon (EVAL_DIST Quarter collapse observed 2/3 runs; one
run EVAL_DIST=[0.573, 0.325, 0.102]). Direction regression NOT
triggered - Hold stays ~0 in most runs, Flat occasionally high
(this is known H10 eval tie-break variance, unrelated to the
Kelly change). q_dir_abs_ref observed in ISV_DIR_MEANS:
~0.14-0.60 across runs - conviction signal is flowing.
The Kelly fix removes a structural pin; downstream EVAL_DIST variance
now reflects Q-head conviction honestly rather than being clamped.
|
||
|
jgrusewski
|
f9a8a5aa9a |
feat(dqn): intent-side magnitude distribution diagnostic (EVAL_INTENT_MAG_DIST)
Adds a parallel read path that reports the policy's intended mag_idx BEFORE Kelly/margin caps and before the Hold/Flat dir_idx forces mag=0. This exposes whether the magnitude Q-head is learning state- dependent preferences, independent of the Kelly cold-start cap that was masking it via actual_mag decoding (kelly_position_cap warmup_floor=0.5 + safety=0.5+0.5*health pinning abs_pos <= 0.375). Kernel changes (experience_kernels.cu): - experience_action_select: new trailing optional arg `out_intent_mag` (int*, NULL=skip). Populated AFTER the existing mag_idx selection via a strict argmax over q_b1, ignoring the Hold/Flat mag=0 forcing, with the same higher-bin-wins tie-break used in the b2/b3 paths. Uses q_sign so the intent stays consistent with contrarian mode. - New scatter_intent_chunk kernel: copies step-major chunked intent [chunk_len, n_windows] into window-major intent_history [n_windows, max_len], mirroring the actions_history layout. Rust wiring (gpu_backtest_evaluator.rs): - New fields intent_mag_buf, chunked_intent_mag_buf, scatter_intent_kernel. Buffers allocated alongside existing chunked buffers in ensure_action_select_ready. intent_mag_buf is zeroed by reset_evaluation_state so short rollouts don't read stale data. - submit_dqn_step_loop_cublas appends the new arg to the action_select launch and launches scatter_intent_chunk immediately after, before the env_batch_kernel (which never touches intent_mag_buf). - read_eval_intent_magnitude_distribution mirrors read_eval_action_distribution_per_magnitude but decodes raw mag_idx (a as usize) rather than the factored action encoding. Training-path call site (gpu_experience_collector.rs): passes NULL (0u64) for the new arg — training does not collect intent history. Trainer wiring: - new last_eval_intent_magnitude_dist field + accessor; populated in metrics.rs::evaluate_on_gpu next to last_eval_magnitude_dist. Smoke test (magnitude_distribution.rs): adds [EVAL_INTENT_MAG_DIST] println line; no new assertions. Diagnostic-only, no new feature flag — production behaviour unchanged. All 7 files build cleanly with no new warnings. [testing: smoke compiles + runs, still fails on the existing H10 assertion (EVAL_DIST Quarter=1.000 driven by Kelly cold-start cap), EVAL_INTENT_MAG_DIST shows Quarter=0.357 Half=0.045 Full=0.599 at 20-epoch smoke on local RTX 3050 Ti — confirming the magnitude head prefers Full ~60% of the time while the Kelly-capped realised distribution pins to Quarter.] Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> |
||
|
jgrusewski
|
04a6f0dea6 |
fix(dqn): Task 2.Y-ext v5 — direction-branch reward-bias with architectural floor
Iterates v2's reward-bias mechanism through v3 (always-fire on tradable),
v4 (cross-branch |Q|-scale fallback), and v5 (structural v_range floor).
Replaces the entire v2 body, not incremental.
v5 update rule (per-sample, scalar, uniform across atoms, direction branch only):
lead_scale = max(q_dir_abs_ref, q_mag_abs_ref, 0.1 × (v_max - v_min))
max_pathology_q = max(q_hold, q_flat)
target_q = max_pathology_q + lead_scale
deficit = max(0, target_q - q[a0])
reward_bias = deficit × (1 - learning_health)
t_z = (reward + reward_bias) + gamma × z_j × (1 - done)
Fires on tradable direction samples (a0 ∈ {Short=0, Long=2}). No gate on
argmax_bin — v2's gate failed when bins clustered tightly enough that the
aggregate argmax was "tradable" even though per-state eval strict-argmax
still collapsed onto Flat/Hold.
Signal stack (all adaptive, no hard-coded knobs):
- isv_signals[17..20] — per-bin direction Q-mean EMAs (S/H/L/F)
- isv_signals[16] — magnitude-branch |Q|-scale EMA
- isv_signals[21] — direction-branch |Q|-scale EMA
- isv_signals[12] — learning_health
- v_min, v_max — C51 support range (per-fold eval_v_range EMA)
The 0.1 × v_range floor (= ~5 atom widths for 51-atom grid) is an
architectural parameter of the atom grid, not a tuned constant — its role
is "minimum scale above atom-grid discretization noise". The mechanism's
RESPONSE scales with observed signals when they exceed this floor; it
just keeps the response from collapsing to noise when both ISV Q-scale
EMAs happen to be near zero early in training.
Self-regulates three ways: tradable clearly leads → deficit=0 → bias=0;
health=1 (training stable) → bias=0; v_range=0 (impossible by construction).
Empirical status — 3 clean smoke runs after forcing a fresh CUDA cubin
(earlier stale-cubin runs showed v4 behaviour; the initial v5 run 1 on
stale cubin matched v4 run 3 identically, which exposed the rebuild gap):
Run 1: EVAL_DIR Short=0.287 Hold=0.000 Long=0.713 Flat=0.000 — Hold+Flat=0 ✓
Run 2: EVAL_DIR Short=0.000 Hold=0.000 Long=1.000 Flat=0.000 — Hold+Flat=0 ✓
Run 3: EVAL_DIR Short=0.000 Hold=0.000 Long=1.000 Flat=0.000 — Hold+Flat=0 ✓
Pre-v5 baseline (committed v2): Hold+Flat ∈ {0.809, 0.872, 0.796} across 3 runs.
The smoke test still fails on magnitude assertions (line 134 eh+ef≥0.30
or line 153 ef≥0.05) because eval magnitude still collapses to Quarter
or Half. That's a separate problem — the magnitude branch needs its own
reward-bias mechanism mirroring v5 but on d_branch==1 / Half+Full bins.
Tracked separately as Task 2.X-ext (internal task #60).
Per feedback_adaptive_not_tuned.md: the mechanism remains signal-driven;
the only scalar constant (0.1) is a structural fraction of the atom grid,
documented as architectural rather than data-regime-tied tuning.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
|
||
|
jgrusewski
|
6061a190b8 |
fix(dqn): Task 2.Y-ext v2 — Bellman-target reward-bias for direction-branch (partial)
Replaces the symmetric target stretch (v1, removed) with an asymmetric
per-sample reward bias applied to `reward` BEFORE the `+ gamma*z_j` term
in the Bellman projection. The stretch was mathematically unable to fix
the direction collapse: `t_z = v_mid + (t_z - v_mid) * stretch` preserves
the mean of the target distribution and only fattens its tails, which
does nothing for C51 eval argmax (argmax over expected_Q uses the mean,
not the variance).
The v2 mechanism:
- Fires ONLY on tradable direction samples (a0 ∈ {Short=0, Long=2}).
- Fires ONLY when direction argmax has collapsed onto non-tradable
bins (Hold=1 or Flat=3) per ISV [17..20] Q-mean EMAs.
- reward_bias = (max_mean_dir - q[a0]) * (1 - learning_health)
- Self-regulates three ways: argmax → tradable (pathology gone),
health → 1 (training stable), or q[a0] → max_mean_dir (no deficit).
Signal wiring (all pre-existing):
- ISV [17..20]: q_s / q_h / q_l / q_f per-bin EMAs
- ISV [12]: learning_health
- Populated by `q_dir_bin_means_reduce` + `isv_signal_update` wiring
landed in commits
|
||
|
jgrusewski
|
c071489979 |
infra(smoke): --max-bars cap for train_baseline_rl + multi_fold smoke
Adds an optional --max-bars CLI cap to `examples/train_baseline_rl.rs`.
When >0, truncates the fxcache-loaded features/targets/OFI/timestamps in
lockstep per the data_loading.rs precedent (commit
|
||
|
jgrusewski
|
b8cd4e1d94 |
diag+docs(dqn): trunk-slice grad decomposition + stale IQN-trunk doc fix
Extends grad_decomp_kernel to snapshot the trunk tensor slice (tensors
0..4 = w_s1, b_s1, w_s2, b_s2) in addition to the existing direction +
magnitude branch slices (8..12 / 12..16). Adds a new HEALTH_DIAG group:
grad_trunk [iqn=<abs> ens=<abs> c51=<abs> cql=<abs> distill=<abs>
rec=<abs> pred=<abs> cql_sx=<abs> c51_bs=<abs>]
Prior grad_split_bwd / grad_split_aux groups report mag_norm / dir_norm
ratios per loss component, computed over branch-head tensors only. That
measurement range structurally reports 0.0000 for any loss component
that writes exclusively to the trunk — IQN and Ens in particular. This
caused the persistent misdiagnosis that IQN-to-trunk was not wired; the
prior scoping in /tmp/foxhunt_research/iqn-to-trunk-wiring-scoping.md
confirmed the wiring is live (apply_iqn_trunk_gradient at
gpu_dqn_trainer.rs:4882) and that the zero reading was a blind spot in
the measurement pipeline.
Smoke confirms the diagnostic: after iqn_readiness ramps up (late
epochs), grad_trunk reports iqn=100..381 (real trunk SAXPY amplitude),
ens=0.07..3.57, c51=2.46..8.91 (value-head dueling path contributes
through trunk), while cql/cql_sx/distill/rec/pred stay near-zero — a
clean diagnostic baseline.
Also fixes stale documentation at dual-distributional-c51-iqn-design.md
that claimed "IQN trains in isolation — its gradients don't flow back
to the shared trunk": reworded to reflect current wired state with
file:function citation and explicit iqn_readiness gating note. Updated
the "What Changes" table ("IQN training") and "Implementation Order"
(Phase 1 marked DONE) with the same citation.
Changes:
- grad_decomp_kernel.cu: per-component result slot 2 → 3 floats
(mag_norm, dir_norm, trunk_norm); extra __shared__ sum_trunk +
tree reduction; new grad_trunk_start/trunk_len kernel args.
- gpu_dqn_trainer.rs: pinned result buffer 18 → 27 floats; snapshot
now does two copy_f32 passes (trunk → dst[0..trunk_len), branch →
dst[trunk_len..]); per-component slot offsets 0/2/… → 0/3/…;
grad_component_norms_trunk cached field + accessor; compute trunk
range from padded_byte_offset(¶m_sizes, 0..4).
- fused_training.rs: grad_trunk_norms_by_component() + per-component
grad_trunk_*_abs() accessors.
- training_loop.rs: HEALTH_DIAG emits new grad_trunk group ordered
[iqn ens c51 cql distill rec pred cql_sx c51_bs]; extended doc
comment explaining the three groups' roles.
- design spec (Problem #1 + What Changes row + Implementation Order):
stale "IQN trains in isolation" replaced by current wired-state
description, cites gpu_dqn_trainer.rs:4882 and readiness ramp at
gpu_dqn_trainer.rs:4228-4243.
Pure diagnostic — no training dynamics change, no atomicAdd, no tuning
knobs, no TF32 changes. Smokes unaffected (magnitude_distribution H10
regression pre-exists on HEAD 810b3c570; 4 other smokes pass).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
|
||
|
jgrusewski
|
810b3c5703 |
fix(dqn): Task 2.Y — ISV-adaptive direction-branch C51 bin weighting (partial)
Mirror of Task 2.X (magnitude branch, commit |
||
|
jgrusewski
|
fa8d546614 |
diag+fix(dqn): Task 2.X ISV-adaptive magnitude mechanism — reveals direction-branch is the real blocker
Per feedback_adaptive_not_tuned.md: adaptive signal-driven mechanism, zero
static tuning knobs. Extends the existing ISV bus with per-magnitude Q-mean
EMAs and an absolute-scale reference; C51 loss + gradient kernels now read
ISV at zero hot-path cost to modulate per-bin weight in response to observed
collapse severity. Weight is 1.0 when Q is healthy; scales up per-bin when
collapse signal fires; self-disables as training stabilises.
Additions:
* ISV_DIM 13 → 17. New slots:
[13] Q_MAG_MEAN_QUARTER: ema(mean Q(Quarter), tau=0.05)
[14] Q_MAG_MEAN_HALF: ema(mean Q(Half), tau=0.05)
[15] Q_MAG_MEAN_FULL: ema(mean Q(Full), tau=0.05)
[16] Q_ABS_REF: ema(max(|Q_mean[k]|), tau=0.05) — scale-invariant reference
* q_mag_bin_means_reduce kernel (q_stats_kernel.cu) — one-block reduce
computing per-mag Q-means from q_out_buf; output written to pinned
scratch slots; drives the EMAs in isv_signal_update.
* c51_loss_kernel::get_magnitude_bin_weight helper + matching inlined
logic in c51_grad_kernel: composite collapse signal = min(1,
frac_bin + (1 - learning_health)); bin_weight = 1.0 + collapse *
mag_bias_signal[k] (bounded in [1, 2]); mag_bias_signal[k] = (k+1)/b1_size.
* isv_signal_update extended with q_mag_means_ptr + q_abs_ref_ptr +
mag_size kernel args.
Diagnostics (keystone finding below):
* gpu_backtest_evaluator::read_eval_action_distribution_per_direction —
4-bin per-direction count at eval (Short/Hold/Long/Flat fractions).
This diagnostic flipped the task diagnosis.
* DQNTrainer::last_eval_direction_dist accessor.
* last_isv_magnitude_bin_q_means accessor.
* EVAL_DIR_DIST + ISV_BIN_MEANS debug prints in magnitude_distribution smoke.
* ef >= 0.05 smoke gate added (currently unreachable behind pre-existing
eh+ef >= 0.30 gate; kept for future use).
Training-time outcome:
Pre-fix MAG_DIST: Quarter=0.60 Half=0.10 Full=0.23
Post-fix MAG_DIST: Quarter=0.46 Half=0.24 Full=0.28 (2.4× Half lift,
Full unchanged)
Pre-fix EVAL_DIST: eq=1.000 eh=0.000 ef=0.000
Post-fix EVAL_DIST: eq=0.981 eh=0.019 ef=0.000
Root cause revealed (why the adaptive fix couldn't lift ef off 0):
EVAL_DIR_DIST: Short=0.045 Hold=0.115 Long=0.070 Flat=0.771
~88% of eval states have direction ∈ {Hold, Flat}. Kernel at
experience_kernels.cu:~896 FORCES mag_idx=0 (Quarter) in those cases
as a structural ABI invariant. Only ~11.5% of eval samples have a
free magnitude choice. Upper bound on ef regardless of magnitude
mechanism: ~0.11.
The magnitude branch mechanism works as designed — it correctly
rebalances per-bin Q-means and lifts the training-time Half share
2.4×. But direction-branch collapse to Flat masks everything
downstream. Task 2.X's magnitude-only scope cannot unblock eval ef.
The real fix target is direction-branch eval collapse. Follow-up
task "Task 2.Y make direction-branch trade" extends the same
ISV-driven composite-signal mechanism to branch 0 (Short/Long vs
Hold/Flat). Scoping doc to be written.
Smoke validation:
magnitude_distribution FAIL (pre-existing eh+ef >= 0.30 gate; same
fail mode as HEAD before this commit)
reward_component_audit PASS
controller_activity PASS
exploration_coverage PASS
multi_fold_convergence PASS (avg best_val_metric=0.039, within ±15%)
No config fields. No static tuning knobs. No feature flags. All
modulation flows through the ISV bus. Shape constants documented:
eps=1e-6 (numerical guard), alpha=0.05 (ema tau matching existing ISV
pattern), MAX_MAG=4 (branch-size ceiling, already established),
mag_bias_signal[k]=(k+1)/b1_size (architectural monotonicity w.r.t.
bin index as stake size).
|
||
|
jgrusewski
|
a9a51e8fa0 |
cleanup+fix(reward): Task 2.4 R6 relocation + Task 2.5 Bug #6 docstring
Task 2.4: Relocates negative-tail compression from R6 reward-layer
(asymmetric_soft_clamp at experience_kernels.cu:78-81) to C51 Bellman
target smoothing (c51_loss_kernel.cu::block_bellman_project_f).
Functionality preserved — same invariant, better location. Upper +10
cap kept inline as fminf(reward, 10.0f) for numerical safety.
Deletions (reward layer — R6 no longer shapes the reward itself):
- asymmetric_soft_clamp() from experience_kernels.cu:78-81 (no callers)
- Reward-layer clamp replaced with fminf(reward, 10.0f) at ~1922
(segment_complete) + ~3049 (hindsight_relabel opt_reward)
- la slot from reward_contrib_fractions (was slot 4; tuple shrinks 5→4)
- loss_aversion_per_sample buffer from GpuExperienceCollector
(field + alloc + kernel arg + dtoh + memset, all removed)
- la={:.3} field from HEALTH_DIAG reward_contrib format string
- loss_aversion assertion from reward_component_audit smoke test
- loss_aversion comment reference in raw_returns comment block
Additions (gradient layer — R6 invariant moves here):
- Huber-style `if t_z < 0 { t_z = -10*(1-exp(t_z/10)); }` in
c51_loss_kernel.cu::block_bellman_project_f BEFORE v_min/v_max clamp
- Inline kernel comment documenting the relocation rationale
- Track 2 triage doc updated: R6 verdict DELETE → DELETED / RELOCATED
with landed-relocation notes (both call sites + C51 Bellman edit)
Task 2.5 Bug #6: Stale `patience_mult` docstring at
experience_kernels.cu:1144 referenced the defunct R7 V8 reward (deleted
in Task 0.8). Rewrote the reward-shape docstring to reflect current
post-V7 / Task 0.8 reality (sparse = 2.0 * vol_normalized_return, capped
inline) and notes the R6 relocation. Per feedback_trust_code_not_docs.md.
Per feedback_no_functionality_removal.md: R6's invariant is RELOCATED,
not deleted. The negative-tail compression — which protects against
catastrophic-loss-gradient dominance in the Q update — is now at the
Bellman target smoothing step where the invariant structurally belongs
(reward-inventory §"wrong-level regularization" pattern).
Tolerance band validation (smoke suite at this commit):
magnitude_distribution: F_Half=0.150 F_Full=0.237 (≥0.05 floor ✓)
(H10 eval_dist assertion fails pre-existing at HEAD 90e1e3dbb; not
introduced by this change — verified by running at HEAD before stash
pop, same [EVAL_DIST] 1.000/0.000/0.000 collapse.)
reward_component_audit: cf_flip=0.584 trail=0.304 (cf_flip≥0.1 ✓, PASS)
controller_activity: [CTRL_FIRE] anti_lr=0.000 tau=0.000 gamma=0.000
clip=0.400 cql=0.000 cost=0.000 (PASS)
exploration_coverage: entropy @ep5=0.988 @ep20=0.985 (PASS)
multi_fold_convergence: Best Sharpe 81.54/38.82/84.18 (≥20 floor ✓)
best_val_metric 0.043/0.024/0.049 (baseline was 0.028/0.018/0.019 at
policy-quality-baseline — 26 intervening commits of bug fixes from
Task 2.5 bugs #1–#7 would account for persistent drift; within
run-to-run variance of HEAD-pre-change)
|
||
|
jgrusewski
|
90e1e3dbb2 |
fix(dqn): Bug #7 — cql_alpha regime gate handles null ISV without silent fallback (Task 2.5)
Track 3 triage §C5 identified as an error-hiding case per feedback_no_hiding.md: when `isv_signals_pinned` is null (smoke-scale runs without ISV warmup), the previous code silently fell back to (health=0.5, regime_stability=0.5), yielding `cql_alpha_eff = base × 0.5 × 0.5 = 0.25 × base` by degenerate math, not by design. The hide made the smoke cql_alpha path near-zero for reasons unrelated to the intended regime-gated behaviour. Fix (option b per plan): emit a one-shot `tracing::warn!` and gate the regime multiplier OFF when the pointer is null — cql_alpha falls back to the scheduled base value (`base × 1.0 × 1.0`). Real ISV path unchanged. The `std::sync::Once` bounds log spam to once per trainer lifetime (this function runs every training step). Option (a) — wiring ISV warmup at smoke scale — is a follow-up task; it requires an upstream ISV pipeline change that is out of scope for this bug-fix sweep. |
||
|
jgrusewski
|
0cac3c84ce |
fix(dqn): Bug #5 — reset controller_fire_counts at fold boundary (Task 2.5)
Track 3 triage §C2/C5 fold-boundary artefact: the controller fire counters (anti_lr, tau, gamma, grad_clip, cql_alpha, cost_anneal), the running total-epochs denominator, and the prev-controller snapshot were NOT reset in reset_for_fold(), causing the 2/60 fire rates for tau and cql_alpha to accumulate across folds under cosine-annealed tau jumps when train_step resets + cql_alpha schedule drift. Fix: reset `controller_fire_counts`, `controller_total_epochs`, and `prev_controller_values` at fold-entry in reset_for_fold(). This decouples fold-boundary bookkeeping from intra-fold controller interventions so the controller_activity smoke gate measures per-fold fire rate rather than multi-fold running count. `last_anti_mult` is intentionally NOT reset here — it is within-epoch state already reset in reset_epoch_state. |
||
|
jgrusewski
|
96ecd0ff46 |
cleanup(dqn): Bug #3 — delete dead \if !true\ update_epsilon block (Task 2.5)
Dead code: `if !true { ... }` is a permanently-disabled guard pattern.
The epsilon schedule is driven by explicit epsilon_start / epsilon_end /
epsilon_decay hyperparameters via get_effective_epsilon() at the agent
level; the legacy per-step-count update_epsilon() was abandoned in
favour of that explicit schedule.
Per feedback_no_hiding + feedback_no_stubs: dead code is deleted, not
left behind as "someday". Per feedback_no_feature_flags: a !true toggle
is a disabled feature flag pattern.
Zero behavior change; removes foot-gun.
|
||
|
jgrusewski
|
54410cba91 |
fix(dqn): Bug #1 — fire_lr detects anti-LR multiplier, not scheduler LR (Task 2.5)
Fire detection captured `cur_lr = lr_scheduler.get_lr()` BEFORE the anti-LR multiplier was applied. Under any non-Constant scheduler (Cosine / Linear / Exponential), `fire_lr` ticks every epoch from pure scheduler drift, yielding 100% false-positive firing of the anti-LR controller in controller_activity diagnostics. Fix: detect anti-LR via the multiplier itself. Added `last_anti_mult: f32` field (init 1.0, reset 1.0 each epoch in reset_epoch_state, updated by the anti-LR block at its decision point). Fire condition becomes `(last_anti_mult - 1.0).abs() > 0.01` — observes the actual intervention, not scheduler drift. Prerequisite for Task 2.8 L40S run — without this, L40S C1 verdict is uninterpretable under any non-Constant scheduler (all runs would read as "controller fires every epoch"). Per Track 3 triage §C1 wiring surprise. |
||
|
jgrusewski
|
2a7005f29a |
fix(dqn): Bug #2 — epsilon_greedy_action samples 4-branch factored space (Task 2.5)
Stale 0..5 range from pre-2026-04-08 9-level code. Only test paths call this cold-path fallback, but the stale range was a latent foot-gun and would mislead anyone reading the code (per feedback_trust_code_not_docs). Fix: sample dir ∈ [0,3), mag ∈ [0,3), ord ∈ [0,3), urg ∈ [0,3) and encode as `dir*27 + mag*9 + ord*3 + urg`, matching MEMORY.md 4-branch DQN architecture (81 factored actions). Per Track 4 E1 triage tech-debt flag. |
||
|
jgrusewski
|
ef165e8961 |
fix(noisy): Bug #4 — sigma_mean returns effective σ, not raw tensor (Task 2.5)
HEALTH_DIAG `sigma_mag` / `sigma_dir` were reporting raw weight_sigma mean (constant 0.0320 across schedules) because reset_noise_with_sigma only scaled the noise epsilon samples, not the underlying σ tensor. Fix: added current_sigma_scale: f32 field on NoisyLinear (default 1.0), updated in reset_noise_with_sigma, multiplied through in sigma_mean() accessor. Also propagated through copy_params_from so target-net sync does not reset the reported effective σ. The HEALTH_DIAG field now reflects the scheduled effective σ, enabling H7 detection signal to actually observe schedule attenuation. Closes Track 4 E2 TUNE finding from Phase 1 triage. |
||
|
jgrusewski
|
4c3806da9c |
fix(cuda): harden latent memcpy_dtoh async-race in download_{params,target_params}
Systematic audit of all `memcpy_dtoh` call sites following commit |
||
|
jgrusewski
|
5da434ab4b |
fix(cuda): sync after async memcpy_dtoh in per_branch_grad_norms — direction-branch determinism
After Option C (commit
|
||
|
jgrusewski
|
199feff4db |
fix(cuda): per-stream cublasLt handles (Option C) — 10× determinism improvement on cuBLAS path
Replaces SharedCublasHandle (one lt_handle rebound across streams) with
PerStreamCublasHandles (one lt_handle per CUDA stream). Implements
NVIDIA's cuBLAS §2.1.4 remediation #1 — documented fix for concurrent-
stream non-determinism.
Context: prior investigation (task a11d706bdb56b5020) ruled out
atomicAdd/RNG/Thrust/multi-stream-sync/graph-capture. Option B (commit
|
||
|
jgrusewski
|
bb399b6359 |
fix(cuda): deterministic cublasLt algorithm selection (Option B)
Replace cublasLtMatmulAlgoGetHeuristic (timing-based, non-deterministic
across process invocations) with cublasLtMatmulAlgoGetIds +
cublasLtMatmulAlgoInit + cublasLtMatmulAlgoCheck across all 10 smoke
training hot-path sites.
Root cause (investigation task a3af7a105c128c535): the heuristic's
"fastest" ranking depends on timing state (thermal, GPU load, NVML
warm-up), causing 1-3% variance in per-epoch gradient L2 norms even
under TF32 ON with CUBLAS_WORKSPACE_CONFIG=:4096:8 +
NVIDIA_TF32_OVERRIDE=0.
Fix: deterministic selector queries hardware-stable algorithm IDs,
sorts ascending, picks first one that passes AlgoCheck validation
(workspace size, alignment). Same inputs -> same algo, always.
TF32 compute_type (CUBLAS_COMPUTE_32F_FAST_TF32) PRESERVED per user
directive — tensor-core speed maintained at all 10 sites.
New module: crates/ml/src/cuda_pipeline/cublas_algo_deterministic.rs
(~485 LOC), process-shared SELECTOR singleton with per-shape cache.
Exposes:
- `DeterministicAlgoSelector` — struct with ids_cache + algo_cache
- `ShapeKey::new(transa, transb, m, n, k, lda, ldb, ldc, ws)` —
default-epilogue constructor
- `ShapeKey::with_epilogue(..., epilogue, ws)` — RELU_BIAS variant
- `get_matmul_algo_deterministic(..)` — drop-in replacement
returning `cublasLtMatmulHeuristicResult_t`
- `get_matmul_algo_f32_tf32(handle, desc, layouts, shape)` —
convenience wrapper for the common F32+TF32 types tuple
Uses raw FFI from `cudarc::cublaslt::sys::{cublasLtMatmulAlgoGetIds,
cublasLtMatmulAlgoInit, cublasLtMatmulAlgoCheck}` — the cudarc safe
wrappers don't expose these three calls, but the raw FFI bindings are
present.
Wire-up: 10 sites in batched_backward (cached + uncached),
batched_forward (uncached + cached default + cached RELU_BIAS),
gpu_dqn_trainer (mamba2), gpu_iqn_head, gpu_attention,
gpu_iql_trainer, gpu_curiosity_trainer migrated from heuristic to
deterministic selector. `matmul_pref` create/set/destroy boilerplate
deleted at every site.
Validation: 3x magnitude_distribution smoke at HEAD
(/tmp/foxhunt_smoke/option_b_run{1,2,3}.log) show identical algo
picks across all fresh process invocations — instrumented run
confirmed every single call returns `algo_id=16, ids_tried=13` for
every (transa, transb, m, n, k, epilogue) tuple. Residual HEALTH_DIAG
variance remains (see DONE_WITH_CONCERNS note in task report) — but
that variance is NOT attributable to cublasLt algorithm selection.
Wall-clock impact: neutral. Per-fold training time stable at
~6.9s / ~8.4s / ~10.4s across folds 1/2/3 with <0.05s std-dev
across 3 fresh runs. First-call AlgoGetIds cost is amortised via
the per-types-tuple cache.
|
||
|
jgrusewski
|
34168f53f2 |
diag(policy-quality): per-magnitude win-rate + return-variance instrumentation
Prerequisite from Task 2.X scoping doc (commit
|
||
|
jgrusewski
|
f9286e938d |
docs(policy-quality): Task 2.X scoping — per-bin variance + Q-spread regularization
Scoping output from the Full=0% investigation (post-Task-2.2 smoke showed
eval_dist [eq=0.580 eh=0.420 ef=0.000] — Quarter+Half recovered,
Full still 0%).
Root cause: Q-estimation bias (category C) — C51 expected-Q over atoms
systematically under-prices higher-variance bins. Full has 4x Quarter's
PnL variance per experience_kernels.cu risk scaling. Bias is already
documented at experience_kernels.cu:904 (commented as "structural
distributional bias — tight return distributions (Small) get higher
expected Q under the C51 softmax regardless of actual expected returns").
Not noise-starvation: Full picked on 32% of training samples across
60 epochs × 3 folds = ~61k Full samples. Ample data.
Proposed Task 2.X = Rank 1 + Rank 2 combined (~90 LOC):
Rank 1: per-bin variance weighting in C51 branch loss — amplify
Full's gradient by sqrt(var_f / mean_var)
Rank 2: Q-spread regularization — lambda * ReLU(target_spread -
Q_spread)^2 penalty preventing degenerate fixpoint
Alternative fixes ranked but not recommended:
Rank 3 per-magnitude reward shaping (~80 LOC) — only if data-disfavor
is confirmed at L40S
Rank 4 magnitude curriculum (~20 LOC) — rejected (noise-starvation ruled out)
Rank 5 state-vector enrichment (~150 LOC) — premature
Prerequisite: ~40 LOC of per-magnitude win-rate + realized-variance
instrumentation to confirm category C vs A at L40S scale.
L40S gating recommendation was:
ef >= 0.15 at L40S → Task 2.X not needed (smoke artefact)
0.05 <= ef < 0.15 → Task 2.X scoped but optional
ef < 0.05 → Task 2.X required
Per feedback_fix_aggressively.md (landed this session): we ship Task 2.X
ahead of L40S since the bias is already code-documented, the fix is
well-scoped, and the 90-LOC cost is reasonable. L40S validation (Task
2.8) will verify the fixed state end-to-end instead of the un-fixed one.
No deletions anywhere per feedback_no_functionality_removal.md.
|
||
|
jgrusewski
|
7b74290dd0 |
docs(dqn): R5 micro-reward — documented intentional disable (Phase 2 Task 2.3)
Per feedback_no_functionality_removal.md: R5 was originally scoped as
DELETE in the Phase 2 plan and the Track 2 triage because
reward_contrib[3] = 0.000 across 60 / 60 smoke epochs and
dqn-smoketest.toml sets micro_reward_scale = 0.0. Re-examination during
Phase 2 Task 2.3 rejected the DELETE path:
- dqn-production.toml already sets micro_reward_scale = 0.1, so R5 is
load-bearing in production, not dead code. The 0.0 value in smoke is
deliberate test-isolation (td_propagation / magnitude_distribution /
reward_component_audit all want the sparse-reward TD path isolated).
- The state-vector OFI block at state[SL_OFI_START..SL_OFI_START+SL_OFI_DIM)
= [42..62) provides representation features for the encoder (policy
side). R5 is a per-bar reward gradient on the critic (critic side).
Different mechanisms — production deploys both together.
- R5 also reads PREV_MID (retrospective hold quality), which is NOT in
the state vector. That signal exists only in the kernel branch.
Changes — pure documentation, no behavior change:
- experience_kernels.cu: ~30-line comment block at the R5 wiring site
(~L1915) documenting the parameter-not-flag status, production vs
smoke values, why state-vector OFI is complementary not redundant,
and the feedback_no_functionality_removal.md seal.
- experience_kernels.cu: fix stale kernel-signature comment that claimed
OFI was at state[66..74). Correct range is [42..62) per state_layout.cuh.
- config.rs: extend DQNHyperparameters::micro_reward_scale docstring and
add a comment at the Default impl pointing back at the kernel site.
- gpu_experience_collector.rs: extend reward_contrib_fractions docstring
to mark the micro=0.000 slot as a SEMANTIC value when the loaded profile
has micro_reward_scale=0.0, not a wiring regression.
- track2-triage.md: R5 verdict changed from DELETE to FIX-documented-disable
with the rationale above; "Proposed Phase 2 changes" section 1 and
"Next Track 2 steps" updated accordingly.
Smoke tests: 3 / 4 pass (reward_component_audit, controller_activity,
exploration_coverage). magnitude_distribution is failing on baseline
HEAD
|
||
|
jgrusewski
|
c0fee5a9bf |
docs(smoke): magnitude_distribution — replace H9-delete references with fix-path
Per standing rule feedback_no_functionality_removal.md: never propose deleting the magnitude branch as a fallback. Updated the Task 2.2 regression-assertion comments + assertion message to point at the actual follow-up fixes if the eh+ef≥0.30 gate fails: - per-magnitude reward shaping - per-bin advantage weighting - magnitude curriculum - state-vector enrichment No semantic change to the test (still asserts eh+ef≥0.30); only the guidance comments were reframed. |
||
|
jgrusewski
|
8aef59f735 |
fix(dqn): H10 — stable argmax tie-break at eval per Track 1 triage + Task 2.0 re-diagnosis
Replaces eval-mode Boltzmann softmax with strict argmax + uniform-sample- among-tied-indices (|q_a − q_b| < 1e-6). Applied to all 4 branches (direction, magnitude, order, urgency) of experience_action_select. Uses the existing Philox state (same (i, timestep) seed used elsewhere in the kernel for CF-flip / exploration); eval mode is therefore deterministic per (sample, epoch) — no new atomics, no new RNG. Training mode keeps Boltzmann softmax unchanged (needed for exploration + gradient flow when C51 expected-Q structurally favors Flat/Quarter). Root-cause re-diagnosis (commit |
||
|
jgrusewski
|
d1068de2b8 |
plan(policy-quality): reframe DELETE tasks per no-functionality-removal rule
User directive: "we don't remove functionality at all, we fix what's
broken". Standing rule captured in
~/.claude/projects/-home-jgrusewski-Work-foxhunt/memory/feedback_no_functionality_removal.md.
Phase 2 plan reframed:
* Task 2.3 (R5 micro-reward): was DELETE; now FIX — either (a) set a
non-zero scale that delivers real signal, or (b) document why it's
intentionally disabled while preserving the code path.
* Task 2.4 (R6 loss-aversion): unchanged — the relocation to C51
Bellman target is consolidation, not removal. Invariant preserved.
* Task 2.6 (E4 entropy-reg): was DELETE-CANDIDATE; now TUNE — if the
signal doesn't reach magnitude, re-route it rather than remove.
* Task 2.7 (C4 grad-clip): was DELETE-CANDIDATE; now
KEEP-AND-DOCUMENT — gradient clipping is a safety feature; run the
ablation for diagnostic, not for removal.
* H9 delete-magnitude-branch fallback: rejected permanently. On
magnitude convergence failure the fallback is per-magnitude reward
shaping / per-bin advantage weighting / curriculum / state enrichment.
Task 2.5 wiring-bug sweep unchanged (correctness fixes, not removal).
This commit updates only the planning narrative. Task execution hasn't
started on 2.3/2.6/2.7 yet, so no code changes required here. When
those tasks run, they will implement the fix-paths, not the
delete-paths.
|
||
|
jgrusewski
|
7e78bf4f85 |
plan(policy-quality): Phase 2 pivot — H4 REJECTED by Task 2.0 data
Task 2.0 instrumentation (commits |
||
|
jgrusewski
|
41b0c559c9 |
diag(policy-quality): Task 2.0 confirmation — expose absolute grad_dir / grad_mag norms
Task 0.4's grad_ratio_mag_dir returns 0.0 whenever dir_norm < 1e-9, so the epoch-end reading of 0.0000 doesn't disambiguate "magnitude starved" vs "direction starved". Task 2.0's per-component data showed CQL and C51 each sending 100-400x more gradient to magnitude than direction — implying direction is the starved one, not magnitude. This adds a HEALTH_DIAG field exposing the raw absolute norms: grad_abs [dir=<sci-notation> mag=<sci-notation>] Along the way uncovered + fixed two latent bugs that had been silently zeroing the ratio signal since Task 0.4 landed: 1. `per_branch_grad_norms` read `grad_buf.len()` = total_params + cutlass_tile_pad (~4096 elements of GEMM tile padding) but the pinned readback slot was sized at construction to total_params exactly. The size check `grad_len > grad_readback_pinned_capacity` was always true, so the accessor returned Err on every call — and FusedTrainingCtx's proxy coerced Err to 0.0 via `.unwrap_or(0.0)`. Root cause for the "always 0.0000" grad_ratio_mag_dir. Fix: read only the first `total_params` prefix of grad_buf (the tail is pure GEMM padding, never holds gradient values). 2. Readback timing: process_epoch_boundary calls estimate_avg_q_value_with_early_stopping early, which replays `eval_forward_exec` — the SAME captured graph as forward_child whose first op is `cuMemsetD32Async(grad_buf, 0, total_params)`. Any grad_buf readback AFTER the avg_q call sees all zeros. Fix: snapshot grad_dir_abs / grad_mag_abs / grad_ratio_mag_dir at the TOP of process_epoch_boundary, before avg_q runs, and consume the cached values in the HEALTH_DIAG block. Last 5 epochs of fold 3 on the magnitude_distribution baseline smoke (FOXHUNT_TEST_DATA=test_data/futures-baseline): HEALTH_DIAG[15] ratio=42.85 grad_abs [dir=6.804900e0 mag=3.989081e2] HEALTH_DIAG[16] ratio=232.66 grad_abs [dir=6.500046e0 mag=3.603353e0] HEALTH_DIAG[17] ratio=318.34 grad_abs [dir=2.720317e-2 mag=1.713861e1] HEALTH_DIAG[18] ratio=367.59 grad_abs [dir=5.180866e-2 mag=8.076681e0] HEALTH_DIAG[19] ratio=298.55 grad_abs [dir=1.989553e-2 mag=6.498848e0] Across all 60 epoch-boundary readings (3 folds × 20 epochs) dir ∈ [~4e-3, ~6e0] and mag ∈ [~5e-2, ~5e2], with ratio mag/dir consistently 50-400× (matching Task 2.0's per-component ratios). Neither branch is near float precision — direction is PROPORTIONALLY starved, not numerically zero. Scenario confirmed: direction is starved relative to magnitude, NOT the reverse. Phase 2's Task 2.1 (architectural fix on magnitude branch) should pivot toward increasing direction's gradient flow instead. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> |
||
|
jgrusewski
|
980f3b07f3 |
diag(policy-quality): Task 2.0 extension — instrument 5 more grad writers
First Task 2.0 pass (commit |
||
|
jgrusewski
|
d60e5375a9 |
diag(policy-quality): Task 2.0 — per-component grad decomposition for H4
Adds grad_mag_{iqn,cql,c51,ens} HEALTH_DIAG fields via in-graph
pinned-snapshot + in-graph reduction kernel (revised approach; first
Task 2.0 dispatch escalated BLOCKED on host-side-snapshots-inside-
captured-graph, plan revised at
|
||
|
jgrusewski
|
5a4d561451 |
plan(policy-quality): Task 2.0 revised approach — in-graph pinned snapshots
First Task 2.0 dispatch escalated BLOCKED: the four loss-component
backward kernels are captured inside the fused training graph, so
host-side snapshot-between-components isn't possible mid-graph without
a force-ungraphed diagnostic step (~210 LOC + cross-stream sync risk).
Revised approach (chosen after cost analysis):
cudaMemcpyAsync(device → pinned host) IS captureable in a CUDA graph.
Even better: DtoD into per-component scratch buffers, then an in-graph
reduction kernel computes per-component (mag_norm, dir_norm) and writes
8 floats to a pinned result slot. Only the 8-float result crosses
PCIe (at epoch boundary), keeping per-step PCIe traffic to zero.
Changes to the plan's Step 2 + Step 3 + Step 4:
- Step 2: added 4 device-side scratch buffers (one per component,
~10 MB each = 40 MB device) + 8-float pinned result slot + new
reduction kernel grad_decomp_kernel.cu spec'd out.
- Step 3: clarified that DtoD snapshot + backward + reduction kernel
are ALL captured in the graph; graph replays them every step;
no force-ungraphed dance needed.
- Step 4: added refresh_grad_component_norms() accessor that reads
the 8-float pinned slot at epoch boundary (zero-copy) and populates
the host-side cache.
Approach matches Task 0.4 pattern (commit
|
||
|
jgrusewski
|
2a54edc92d |
plan(policy-quality): Phase 2 plan refinements (tolerance bands + decision tables)
Three polish items from the plan review: 1. Task 2.4 Step 5 — replaced vague "must not regress" with a numeric tolerance-band table: multi_fold best_val_metric ±15% per fold, Best Sharpe ≥ 20 floor, and hard F_Half/F_Full ≥ 0.05 + cf_flip ≥ 0.1 BLOCKERS. Anchors to the baseline metrics doc values captured at policy-quality-baseline. 2. Task 2.6 — converted the ad-hoc text decision at Step 1 into the same five-row decision table format Task 2.1 uses (DELETE / KEEP / KEEP-as- safety-net / INCONCLUSIVE / DELETE-because-inseparable). Step 2 ablation got its own 4-row outcome table keyed to ent_mag delta, multi-fold Sharpe regression, and NaN appearance. 3. Task 2.7 Step 1 — added "Repeat 3× with different seeds" clause and explicit total wall-clock note: ~75 min (3 × 25 min) for the ablation pass before the Step 2 decision. Aligns with the Cross-cutting concern #6 about sample-noise rejection. No task count change (still 11: 2.0–2.10). Net code-delta estimate unchanged. Standing-rule compliance unchanged (no stubs / no atomics / no quickfixes / no hiding / no feature flags / no push-per-task). |
||
|
jgrusewski
|
1ce99efc53 |
plan(policy-quality): Phase 2 implementation — synthesis of 4 tracks
Consolidates Phase 1 triage findings from Tracks 1-4 into an 11-task
Phase 2 plan at docs/superpowers/plans/2026-04-21-policy-quality-phase2.md.
Task inventory:
- 2.0 Per-component gradient decomposition (H4 keystone diagnostic)
- 2.1 H4 fix — magnitude-head gradient starvation (decision tree from 2.0)
- 2.2 H10 fix — stable argmax tie-break at eval
- 2.3 DELETE R5 micro-reward
- 2.4 DELETE R6 loss-aversion; relocate neg-tail to C51 target smoothing
- 2.5 Wiring-bug sweep (7 bugs: C1 fire, epsilon gen_range, if !true,
sigma_mean scale, fold-boundary reset, stale docstring, C5 ISV null)
- 2.6 E4 entropy-reg DELETE-or-KEEP (data-driven, post-2.0)
- 2.7 C4 adaptive grad-clip ablation + DELETE-or-KEEP
- 2.8 L40S validation run — all 4 tracks re-measured
- 2.9 Mandatory-gate verification + phase3-results.md
- 2.10 Tag policy-quality-phase2-complete (and policy-quality-v1 if soft pass)
Matches Phase 0/1 plan formatting (checkbox steps, concrete file paths,
code snippets, bash commands, per-task commit templates). References
project standing rules (no quickfixes, no stubs, no atomic-adds on hot
paths, no feature flags, no hiding errors) and the pinned-readback
pattern from Task 0.4 (commit
|
||
|
jgrusewski
|
0611d32b06 |
docs(policy-quality): Track 3 controllers audit (Phase 1) — 0 load-bearing, 6 diagnostic, 1 candidate-for-delete
V7 audit of the 7 adaptive controllers named in spec §5.3 against the
baseline controller_activity smoke run (3 folds × 20 epochs = 60 epochs,
intervention-based fire detection per commit
|
||
|
jgrusewski
|
207fce8778 |
docs(policy-quality): Track 2 reward audit (Phase 1) — 2 DELETE, 3 KEEP
V7 audit of reward terms R1-R8 against the baseline smoke run
(magnitude_distribution.rs, 3 folds x 20 epochs = 60 HEALTH_DIAG rows):
* R1 step_return: KEEP (base reward, denominator)
* R2 PopArt drift: PENDING (warmup-gated zero at smoke; re-check at L40S)
* R3 CF-flip: KEEP (49-74% contribution, dominates shaping)
* R4 trail_r: KEEP-WITH-CAVEAT (fold-3 dominance, trade-volume gated)
* R5 micro-reward: DELETE (micro_reward_scale=0 in smoke, intended)
* R6 loss-aversion: DELETE (sub-1% in 54/60 epochs, relocate protection)
* R7 segment-patience: ALREADY-REMOVED (stub deleted
|
||
|
jgrusewski
|
5c70c68a15 |
docs(policy-quality): Track 1 magnitude triage (Phase 1) — H4 + H10 CONFIRMED
Preliminary triage of spec §5.1 hypotheses H1–H10 using the Phase 0 baseline capture on RTX 3050 Ti. L40S validation pending per plan. Verdicts: H1 PENDING (needs forced-exploration instrumentation not yet wired) H2 REJECTED var_scale=0.96 across 19/20 epochs; Var[Q] inactive at smoke scale H3 INCONCLUSIVE kelly degenerate (insufficient win/loss counts at smoke scale) H4 CONFIRMED grad_ratio_mag_dir=0.0000 across 20/20 epochs (threshold <0.1) H5 REJECTED ent_mag stays ≥0.98 throughout; no bootstrap collapse H6 REJECTED Full fire rate (0) is lower than Quarter fire rate, not higher H7 REJECTED vsn and sigma symmetric between mag and dir branches H8 REJECTED target-net drift equal (mag=dir=0.001) H9 PENDING (same instrumentation gap as H1) H10 CONFIRMED training ent_mag=0.98, eval F_Quarter=100% Synthesis: H4 is the root cause. Magnitude branch receives ~0 gradient → weights stay near init → three magnitude Q-values near-identical → argmax picks bin 0 (Quarter) on ties → H10 manifests at eval time. H2, H5, H7, H8 all ruled out as contributors. Proposed Phase 2 priority: fix H4 (gradient-flow path into magnitude head — likely per-component advantage weighting or direction-conditioning of w_b1fc) + H10 (Q-margin argmax + stochastic eval rollouts as safety net). Phase 1 next: validate preliminary verdicts on L40S, instrument per-component gradient decomposition for magnitude, proceed with Tracks 2/3/4 in parallel. |
||
|
jgrusewski
|
8ab368de58 |
docs(policy-quality): Task 0.17 baseline metrics — captured (Phase 0 complete)
All <PENDING> markers replaced with real values from a clean 20-epoch
magnitude_distribution smoke run + 3-fold multi_fold_convergence run
on RTX 3050 Ti at HEAD
|
||
|
jgrusewski
|
0472b97300 |
test(smoke): strengthen performance probes with meaningful thresholds
Previously test_training_throughput_measurement and test_real_data_single_epoch
only asserted loss.is_finite() on the final metric. That passes on trivial
zeros, on huge-but-finite NaN-disguised values, and on any regression that
doesn't produce literal NaN — giving effectively no signal.
test_training_throughput_measurement now asserts:
- loss finite AND non-negative
- epochs_trained >= 1
- throughput floor: epochs_per_sec > 0.05 (i.e. each epoch < 20s on the
RTX 3050 Ti; catches accidental CPU fallback or kernel CPU-pinning).
Documented as a conservative local floor; CI may tighten.
- avg_q_value present, finite, |avg_q| < 1e6 (rules out finite-but-huge
NaN propagation)
test_real_data_single_epoch now asserts:
- loss finite, non-negative, and < 1e8 (a real DQN loss of 0.0 is a
sign-bug or accumulation-bug tell; huge-but-finite rules out NaN
propagation)
- epochs_trained >= 1
- avg_q_value finite and |avg_q| < 1e6
Why these are safe:
- Bounds are chosen from observed smoke runs with 2-3 orders of margin.
- Passes locally in 1.58s and 10.24s respectively.
- Designed to flag regressions, not true production-scale deviations.
Verified PASS on laptop (RTX 3050 Ti).
|
||
|
jgrusewski
|
2570fe0130 |
fix(smoke): test_fxcache_zero_copy_training — real failure detection
Previously the fold loop swallowed every training error into f64::NAN,
unconditionally pushed a value at the top of the loop, and then asserted
only !fold_losses.is_empty() — a tautology that could never fail. Any
CUDA error, NaN explosion, or regression of the zero-copy path passed
silently.
Now:
- Error propagation via `?` (no swallow). A zero-copy test can't
tolerate training failures — if training blew up, the zero-copy
wiring is broken and must surface.
- All fold losses must be finite and non-negative.
- Every fold must report epochs_trained >= 1 (zero-epoch fold =
kernel skipped or buffer not populated).
A direct "no htod/dtoh copy happened" assertion would need a copy-counter
instrumented into the fused-training GPU path plus a field on
TrainingMetrics. That is out of scope for this test; documented inline.
Verified PASS on laptop (RTX 3050 Ti):
loss=0.0018366, epochs=1
|
||
|
jgrusewski
|
e8ecb2f626 |
fix(training_loop): surface FusedTrainingCtx init errors instead of hiding them
Both lazy-init sites (train_with_data_full_loop_slices @ L344 and
run_training_steps_slices @ L1511) previously logged FusedTrainingCtx::new
failures via tracing::error! and continued with fused_ctx = None. On CUDA
builds there is no CPU fallback, so the next stage (GPU experience collector)
then fails ~100ms later with a misleading "GPU experience collector MUST be
active for CUDA training" error that obscures the real CUDA root cause (OOM,
driver error, etc.).
Both sites now propagate the original error via map_err → anyhow::anyhow! so
callers see the actual failure. The post-init wiring (PER buffer pointers,
RNG dev ptr) is refactored from `if let Some(ref mut fused)` — which was
silently-no-op on the hidden-error path — to an unconditional
as_mut().expect() since fused_ctx is guaranteed Some after the `?`.
The two remaining `fused_ctx = None` assignments are legitimate:
- mod.rs:655 in Drop (ordered GPU teardown)
- training_loop.rs:1508 batch-size-change recreate (immediately reassigned
by the ? on the next line)
No new API, no threshold changes, no feature flags.
|
||
|
jgrusewski
|
0beccd5e82 |
fix(smoke): multi_fold_convergence — laptop-sized config
Two wrong-scale assumptions in the test made it unachievable on the
RTX 3050 Ti / 4 GB laptop this smoke is meant to run on:
1. `train_baseline_rl` was invoked without `--training-profile`, so it
defaulted to `dqn-production`: batch_size=16384, buffer=500k,
num_atoms=52, hidden_dim_base=256. Fused-CUDA init OOMs at
`kan_d_coeff_per_elem alloc` on 4 GB, leaving `fused_ctx = None` and
every subsequent fold failing with "GPU experience collector MUST be
active for CUDA training". Fix: pass `--training-profile=dqn-smoketest`.
2. Default walk-forward windows (12 train / 3 val / 3 test / 3 step) only
yield 2 folds in the 24-month baseline dataset — fold 2's test-end
lands one month past `data_end`. The test's pass-gate is "≥2/3 folds
produce a checkpoint", so a test that can only ever generate 2 folds
is degenerate. Fix: explicit shorter windows (6 / 2 / 2, step 2) that
yield all 3 folds (`6 + 2*2 + 2 + 2 = 14 ≤ 24`, comfortable margin).
Also drops `--epochs 20` → `--epochs 5`. Each fold runs ~5500 batches at
~33 s/epoch on this GPU; 20 × 3 folds ≈ 33 min was exceeding the smoke
budget (kill observed around the 10-minute mark). 5 epochs is ample for
the checkpoint gate — `best_sharpe` saves on the first improving epoch
(epoch 1 in practice), so more epochs add no pass/fail signal, only
wall-clock.
Verified locally: 3/3 folds produce `dqn_fold{N}_best.safetensors`,
total wall-clock ~7 min.
[MULTI_FOLD] fold 0 checkpoint OK
[MULTI_FOLD] fold 1 checkpoint OK
[MULTI_FOLD] fold 2 checkpoint OK
test result: ok. 1 passed; 0 failed ... finished in 416.36s
Docstring updated to reflect new sizing and call out the 4 GB / 24-month
constraints explicitly so the next person reading this can see why the
numbers are what they are.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
|
||
|
jgrusewski
|
4399a56d76 |
test(smoke): strengthen weak assertions across DQN smoke suite
Five tests in crates/ml/src/trainers/dqn/smoke_tests had pass gates that validated existence rather than the invariant their doc-comment claimed to test. A trainer returning all-zero diagnostics (a plausible wiring regression) would have passed them. - reward_component_audit: was `is_finite() && >= 0.0` on 5 slots — passes trivially on all-zero stubs. Added `cf_flip > 0.1` and `trail_r > 0.01` floors (known-wired slots in smoke config; popart/micro/loss_aversion remain finite-only as they are legitimately near-zero in smoke). Run-observed values: cf=0.614, trail=0.295, la=0.006 — well above floors. - exploration_coverage: `.unwrap_or(0.0)` silently substituted 0 for a missing epoch, conflating "emission regressed" with "exploration collapsed". Now panics with a distinct message on missing entries, also asserts len >= 20, normalized range [0,1], and spread > 1e-6 to catch constant-output emitters. Run-observed spread: 0.275. - training_stability::50_epoch_convergence: entropy assertion was guarded behind `if entropy.is_finite()`, so NaN entropy (the more severe failure) silently passed. Fail hard on NaN first. - training_stability::trading_model_behavior: same `is_finite` guard pattern on action_entropy — now fails hard when the diagnostic is missing or NaN rather than skipping. - training_stability::gpu_collector_auto_initializes: only asserted training returned `Ok(_)`. A collector producing silent zeros would pass. Now also verifies epochs_trained, loss finiteness, and gradient flow. - walk_forward::no_overfitting_50_epochs: had two tautological "finite check" assertions (`x < x + 1` and the signum-adjusted ratio) that always passed regardless of divergence. Replaced with real `.is_finite()` checks plus a `div_ratio < 10.0` stability gate. Tests run under CUBLAS_WORKSPACE_CONFIG=:4096:8 + FOXHUNT_TEST_DATA=test_data/futures-baseline, release-test profile. reward_component_audit and exploration_coverage both PASS on the local RTX 3050. No threshold relaxation or quickfixes applied; any future wiring regression will now be caught by a meaningful assertion instead of a near-tautology. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> |
||
|
jgrusewski
|
ed4b30b493 |
fix(smoke): controller_activity — V7 intervention-based fire detection
Redefine the "fire" semantic for adaptive controllers per the V7 audit (policy-quality-design spec §5.3): a controller fires iff it made an ADAPTIVE INTERVENTION this epoch, not merely because its observable output value changed. Previously fire detection was absolute-delta on the output: - grad_clip fired in 98.3% of epochs because the adaptive clip threshold is an EMA recomputed every training step. The EMA drifts by > 1e-3 every epoch regardless of whether the clip actually clamped a gradient. - cost_anneal fired in 98.3% of epochs because it is a deterministic sigmoid of current_epoch (1/(1+exp(-(epoch-10)/3))) with no adaptive or reactive component. Every epoch moves it by > 1e-4 by design. Neither was "load-bearing" in the V7 sense — one was a per-step EMA tracker, the other a pure curriculum schedule. The prior test output "controller 'grad_clip' fires in 98.3%" was a false positive from measuring the wrong signal. New semantics: - anti_lr / tau / gamma / cql_alpha: unchanged — absolute delta vs prior epoch on the effective output value (real adaptive controllers). - grad_clip: intervention-based latch `grad_clip_kicked_this_epoch`, set in run_training_steps_slices iff raw_grad_norm > active clip at any training step this epoch. Reset in reset_epoch_state. - cost_anneal: pure deterministic schedule → never load-bearing → always fires=false. The value is still tracked in prev_controller_values and the HEALTH_DIAG line still emits it for observability, but it cannot trip the 50% load-bearing gate. After fix, controller_activity smoke reports: anti_lr=0.000 tau=0.033 gamma=0.017 clip=0.233 cql=0.033 cost=0.000 All 6 rates ≤ 0.5. Test passes. Touched: - crates/ml/src/trainers/dqn/trainer/mod.rs (add grad_clip_kicked_this_epoch) - crates/ml/src/trainers/dqn/trainer/constructor.rs (init new field) - crates/ml/src/trainers/dqn/trainer/training_loop.rs (latch kick per step, redefine fire_clip + fire_cost in HEALTH_DIAG block) Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> |
||
|
jgrusewski
|
a5f23b28f0 |
fix(diag): move GPU summary download before HEALTH_DIAG emit
Third root cause of the apparent magnitude collapse in smoke-test data:
HEALTH_DIAG reads monitor.action_counts to compute dist_q/h/f and
ent_mag/ent_dir, but the block that populated action_counts from the
GPU summary ran AFTER HEALTH_DIAG. Every epoch saw all-zero counts so
dist_q=dist_h=dist_f=ent_mag=ent_dir=0 in every log line, and the
last_magnitude_dist cache the smoke test reads was always zeros.
Moved the GPU-summary download + monitor population block up above the
HEALTH_DIAG preparation block. Signal is now real per-epoch:
dist_q=0.60 dist_h=0.15 dist_f=0.25 (was 0/0/0)
ent_mag=0.83 ent_dir=0.89 (was 0/0)
magnitude_distribution smoke test now PASSES on local RTX 3050 Ti at
20 epochs. Final: Quarter=0.598 Half=0.154 Full=0.249 — all above 5%
smoke threshold.
Combined with
|
||
|
jgrusewski
|
2fb30f098e |
fix(monitoring): cascade 9→12 action bins for 4-direction action space
The runtime DQN action space is 4×3×3×3 = 108 (dir×mag×order×urgency) per
the kernel's b0_size=4 (Short/Hold/Long/Flat) layout, but the monitoring
stack was still coded for the legacy 3×3=9-bin exposure layout. Flat
actions (dir=3) had exp_idx = 9 >= num_actions=9 and were silently
dropped by the monitoring_reduce kernel's bounds check.
Cascade:
- monitoring_kernel.cu: num_exposure_bins=12 (EXP_MAX constant), summary
layout 27 floats with bins [5..17), order [17..20), urgency [20..23),
N [23], trades [24], pad [25..27).
- gpu_monitoring.rs: MonitoringSummary.action_counts [usize; 12];
summary_buf 27 f32; reduce() now takes b0_size and validates
num_exposure_bins ≤ EXP_MAX.
- trainers/dqn/monitoring.rs: action_counts/q_value_sums/q_value_counts
all [_; 12], factored_action_counts [_; 108], EXPOSURE_NAMES with 4
dirs × 3 mags ("S_Small".."F_Full"); track_action now maps the 8-level
ExposureLevel enum onto the kernel's dir*3+mag layout by going through
direction()/magnitude() accessors (was indexing by raw `exposure as
usize` which is an entirely different scheme).
- trainer/training_loop.rs: total_action_counts [_; 12],
total_factored_action_counts [_; 108]; GPU monitoring reduce call
now passes b0 from agent.branch_sizes(); direction-entropy gains a
4th dir bin; per-action Q diagnostics names array extended to 12.
- trainer/metrics.rs: create_final_metrics arrays widened; q_diagnostics
per_action_avgs [_; 12]; combo_idx = d*3 + m (matching kernel layout);
action_space_size 108/12 and buy/sell/hold uses the new exp_idx
boundaries.
- financials.rs: action_counts [_; 12]; buy/sell/hold classification
updated — BUY = Long [6..9], SELL = Short [0..3], HOLD = Hold + Flat
([3..6] ∪ [9..12]). Tests updated for 12-bin layout.
Also fixes several collateral bugs exposed by the audit:
1. magnitude_action_dist_final conflated Hold (dir=1, mag-forced-to-0)
with "Quarter magnitude". New formula restricts the denominator to
tradable directions (Short + Long) and computes Q/H/F fractions only
over those, producing the real magnitude preference signal.
2. training_loop.rs:3012 flat_count indexed [3,4,5] (which is Hold in
the new layout, but was already the wrong bucket — "Flat" was never
at that index even under the legacy 9-bin reading, where [3,4,5]
was the Flat bin but mag was forced to 1 not 0). New code correctly
separates flat_count [9..12] from hold_count [3..6] and excludes
both from the directional denominator and diversity cell threshold.
3. monitoring.rs had q_value_sums: [f64; 7] but q_value_counts:
[usize; 9] — different sizes for what should be the same exposure
axis. Both now [_; 12].
4. log_action_distribution iterated over 7 indices when printing per-
action Q-values but the array is logically 12-wide — now uses
`0..12` with the 12-entry EXPOSURE_NAMES lookup.
Tests: all 7 financials tests pass. Monitoring cubin loads + default
summary sanity checks pass. Smoke test magnitude_distribution now
emits correct 12-bin breakdown — GPU summary shows (example epoch 20):
actions=[786, 157, 280, 377, 2, 7, 806, 156, 317, 303, 5, 4]
which decodes as Short (S) = 786/157/280 (~38%), Hold (H) = 377/2/7
(~12%, mag-forced), Long (L) = 806/156/317 (~40%), Flat (F) =
303/5/4 (~10%, mag-forced). Previously the Flat bucket (303+5+4 =
~10% of all actions) was silently dropped — policy was picking Flat
~10% of the time and nobody could see it.
No atomic ops added; all reductions remain via warp-scratch + lane-0
sequential merge. No feature flags, no stubs, no quickfixes.
|
||
|
jgrusewski
|
2ac956298b |
fix(data-loading): truncate OFI in lockstep with bars on max_bars path
Two truncation sites in data_loading.rs truncated feature/target vectors to max_bars but left self.ofi_features at the original (cache-full) length. Downstream DqnGpuData::upload_slices asserts OFI len == num_bars and surfaced this as "OFI/market data length mismatch: 175874 OFI rows vs 5000 bars — fxcache is corrupt". The cache was not corrupt — the upstream truncation was the bug. Fixed by moving OFI storage below the truncation block in the fxcache path and by adding a parallel truncate in the DBN fallback path. Added debug_assert_eq on equal lengths post-truncation so future regressions hard-fail in dev. Also corrected the cuda_pipeline/mod.rs error message to point at the actual culprit (upstream truncation desync) instead of blaming the on-disk cache. |