main
23 Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
jgrusewski
|
e877e8aefe |
chore(sp18): archaeology cleanup — strip post-deletion markers + ISV_TOTAL_DIM giant docstring
Pure-comment follow-up to SP18 Phase 1 atomic deletion ( |
||
|
jgrusewski
|
3c318953a7 |
feat(sp18 v2 P1.T3-T6): atomic delete SP13/SP16 hold_cost_scale chain
Deletes the entire reactive Hold-cost controller chain from SP13 P0a / SP16
P2 / SP16 T3 atomically — replaced by the SP18 D-leg structural opportunity-
cost reward landing in Phase 2.
Per DD7(c): observation chain preserved (slots 380, 382). Slot 380 becomes
constructor-init-only diagnostic at HOLD_COST_BASE=0.005, never updated.
18 consumer sites deleted (8 from plan checklist + 10 from A1-A10 audit
expansion):
Production code:
- 3 reward subtraction sites (replaced with Phase 2 placeholder comments)
- hold_cost_scale_update_kernel.cu (file deleted)
- HoldCostScaleUpdateOps struct in gpu_aux_trunk.rs (~120 lines)
- launch_hold_cost_scale_update in gpu_dqn_trainer.rs (forwarder)
- Host controller block in training_loop.rs (~60 lines)
- Slot constants [461..468) in sp14_isv_slots.rs
- HOLD_COST_CONTROLLER_GAIN/FLOOR/CEIL constants in sp13_isv_slots.rs
- state_layout.cuh mirror constants for [461..468)
- build.rs cubin manifest entry
- 7 fold-reset registry entries + 7 dispatch arms
Tests:
- sp14_oracle_tests.rs lines 2173-2926 (11 tests + helpers, ~750 lines)
— these include_bytes! the deleted cubin and cannot survive
- lock_sp18_v2_pp4_retired_chain test (deleted, no inverse-contract
replacement — locking a deletion is pointless ceremony per user call)
Layout fingerprint: range [461..468) is RESERVED gap (SP14-C.1 pattern;
ISV_TOTAL_DIM stays at 507 for checkpoint compatibility). Layout fingerprint
seed updated atomically: HOLD_COST_SCALE + HCS_* slots replaced with
RESERVED_GAP_461_TO_468=SP18_P1_RETIRED marker.
INTERIM STATE: 3 reward sites are now missing the per-bar Hold cost. This
is forbidden to L40S-dispatch until Phase 2 lands the
compute_sp18_hold_opportunity_cost device fn that replaces the deleted
subtraction. The audit doc records this hold.
Validation: cargo check --workspace clean; cargo test -p ml --lib passes
(state_reset_registry::every_fold_and_soft_reset_entry_has_dispatch_arm,
sp16_t3_wiener_welford_slot_layout_locked, sp18_combined_slot_layout_locked,
sp18_shrink_perturb_slot_layout_locked, sp18_fold_reset_entries_present —
all OK). Audit fingerprint: per-bar Hold-cost subtraction sites count
13 → 0 (smoking gun for complete deletion).
Plan: docs/superpowers/plans/2026-05-08-sp18-reward-shape-hold-attractor.md
Audit: docs/sp18-wireup-audit.md (full A1-A10 expansion + post-deletion
fingerprint)
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
|
||
|
jgrusewski
|
25ff1d4195 |
fix(sp16): floor Wiener-α at 0.4 for non-stationary control loops
Wiener-optimal α minimizes MSE for stationary stochastic signals.
The min_hold_temp / hold_cost_scale loops track a target driven by
the adapting policy itself — non-stationary by construction.
Empirical evidence: train-multi-seed-b5gmp sha
|
||
|
jgrusewski
|
641aa0dfde |
fix(sp16-t3): Wiener-optimal adaptive α per pearl — hold_cost_scale + min_hold_temperature
Per train-multi-seed-hjzss validation: SP16 T1+T2 chain was structurally
landed but BEHAVIORALLY INERT in 5-epoch smoke (bit-identical to pfh9n
baseline through epoch 3). Root cause: hardcoded `alpha = 0.05f` in both
producer kernels violates feedback_isv_for_adaptive_bounds AND prevents
convergence in short runs (~60 epochs needed from cold start).
Fix per pearl_wiener_optimal_adaptive_alpha:
α = diff_var / (diff_var + sample_var + ε)
Where sample_var = running variance of target signal (Welford accumulator)
and diff_var = running variance of consecutive one-step differences.
Cold-start: target jumps 1.0 → 6.4 → 7.0 → high diff_var → α ≈ 0.6+
→ near-bootstrap responsiveness in epochs 1-3
Steady-state: signal stabilizes → diff_var drops → α decays naturally
→ smoothing emerges without hardcoded constant
Adds 12 new ISV slots (6 per producer):
- HCS_TARGET_MEAN/M2, HCS_DIFF_MEAN/M2, HCS_PREV_TARGET, HCS_SAMPLE_COUNT
- MHT_TARGET_MEAN/M2, MHT_DIFF_MEAN/M2, MHT_PREV_TARGET, MHT_SAMPLE_COUNT
ISV_TOTAL_DIM 462 → 474.
Both kernels migrated atomically. Pearl-A bootstrap preserved (sentinel
on prev_blended triggers REPLACE; cold-start α=1.0 when N<3 samples).
Defensive bounds [WELFORD_ALPHA_MIN=0.01, WELFORD_ALPHA_MAX=0.95] on the
Wiener-derived α to guard against denormal/underflow corner cases.
HEALTH_DIAG[N] emit extended with `alpha=...` and `sample_count=...` for
direct trajectory observation in validation smoke.
Behavioral tests verify:
- α high during signal jumps (>0.3 at epoch 3 post-cold-start)
- α low in steady state (mean tail α<0.4 under converging signal)
- Pearl-A bootstrap fires on first observation (Welford state advances
regardless of REPLACE branch)
- α stays within [WELFORD_ALPHA_MIN, WELFORD_ALPHA_MAX] over 50 epochs
(post-cold-start; cold-start α=1.0 by design)
- No 0.05f hardcoded literal remains in blend math (regression-locked
via host-only string scan)
5 GPU + host tests pass: sp16_phase3_alpha_high_during_signal_jump,
alpha_low_in_steady_state, pearl_a_bootstrap_first_obs,
alpha_naturally_bounded, no_hardcoded_alpha. sp14 + sp15 oracle suites
unchanged (34 GPU tests + 4 host tests).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
|
||
|
jgrusewski
|
1a3bcf97b8 |
feat(sp16-p2): adaptive Hold cost scale via ISV[461]
Per train-multi-seed-pfh9n post-mortem: observed_hold_rate climbed 0.25 → 0.52 across training while cost penalty (~0.006) was 100× smaller than per-bar reward magnitudes (popart=0.97, cf=0.65). Hold action was effectively free, allowing Q(Hold) to dominate via structural low-variance bias. Fix: scale Hold cost adaptively. ISV[HOLD_COST_SCALE_INDEX=461] tracks: scale = clamp(1.0 + 24.0 × max(0, observed - target) / max(target, 0.01), 1.0, 25.0) Effective cost at 100% overrun (observed=2× target): 0.006 × 25 = 0.15, competitive with per-bar reward magnitudes (~0.01-0.1). At/below target: scale = 1.0 (no extra penalty). Pearl-A bootstrap + Welford slow EMA (α=0.05). Mirrors T1's MIN_HOLD_TEMPERATURE pattern (same input signals: ISV[382] observed, ISV[381] target). Producer: hold_cost_scale_update_kernel.cu — single-thread cold-path, per-epoch boundary, AFTER MIN_HOLD_TEMPERATURE in training_loop.rs. Consumer migration (atomic per feedback_no_partial_refactor): 3 sites in experience_kernels.cu — segment_complete branch (line ~3089), per-bar positioned-Hold branch (line ~3553), per-bar flat-Hold branch (line ~3617). Cold-start fallback: scale=1.0 when slot ≤ 0 or out-of-bounds (bit-identical pre-Phase-2 cost magnitude). ISV_TOTAL_DIM: 461 → 462. Behavioral tests (5/5 PASS on RTX 3050): - sp16_phase2_hold_cost_scale_climbs_with_overrun - sp16_phase2_hold_cost_scale_at_target_is_one - sp16_phase2_hold_cost_scale_under_target_is_one - sp16_phase2_hold_cost_scale_bounds_clamp - sp16_phase2_hold_cost_scale_pearl_a_bootstrap Regression: SP14 oracle suite 30/30 PASS, SP15 phase 1 oracle suite 36/36 PASS. Instrumentation: HEALTH_DIAG[N]: hold_cost_scale_diag obs/tgt/norm/scale. Per feedback_isv_for_adaptive_bounds + feedback_no_partial_refactor. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> |
||
|
jgrusewski
|
2c469af2f7 |
fixup(sp16-p1): correct stale build.rs comment + remove tautological test assertion
Code-quality review of
|
||
|
jgrusewski
|
0426ce8887 |
fix(sp16-p1): MIN_HOLD_TEMPERATURE signal swap to hold-rate overrun + slot 330 non-bug closure
Per train-multi-seed-pfh9n post-mortem follow-up: slot 460 stuck at 50 in Fold 1 was NOT a launch-lifecycle bug. Producer fires per-epoch but kernel had early-return guard on AUX_DIR_ACC_SHORT_EMA (slot 373) at sentinel 0.5. Slot 373 reset on fold boundary; aux dir-acc EMA either didn't fire or settled within ε of 0.5 → kernel kept early-returning. Fix: drop slot 373 dependency entirely. Drive temperature from observed hold-rate vs target overrun: overrun = max(0, observed_hold_rate - target_hold_rate) overrun_norm = clamp(overrun / max(target, 0.01), 0, 1) new_temp = TEMP_MIN + (TEMP_MAX - TEMP_MIN) × overrun_norm blended_temp = Welford EMA α=0.05 with Pearl-A bootstrap When over-holding: temp HIGH → exit ramp permissive (matches design intent). When at/under target: temp LOW → exit penalty strict. Survives fold reset: hold-rate measurement starts fresh with real data immediately, no chained-input-sentinel masking. Slot 330 (KELLY_WARMUP_FLOOR) investigated and confirmed NON-BUG: producer behaves correctly per pearl_kelly_cap_signal_driven_floors cross-fold- persistence. floor=0 post-warmup is correct steady state. Behavioral tests: - sp16_phase1_min_hold_temp_climbs_with_hold_overrun - sp16_phase1_min_hold_temp_strict_when_at_target - sp16_phase1_min_hold_temp_strict_when_under_target - sp16_phase1_min_hold_temp_no_longer_reads_slot_373 Instrumentation: HEALTH_DIAG[N]: min_hold_temp_diag obs/target/overrun/temp. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> |
||
|
jgrusewski
|
fb24614f07 |
feat(sp16-p0): Q-by-action HEALTH_DIAG diagnostic for Hold-bias observation
Per train-multi-seed-pfh9n post-mortem: structural-Q-bias hypothesis (Adam's m/sqrt(v) prefers low-variance Hold over noisy direction Q-targets) needs direct per-action Q observations to verify or refute. WR plateaued at ~0.46 across both folds while Hold% climbed 15% → 52% and Q-mean climbed 0.19 → 0.41 — but per-action Q was unobservable. Adds HEALTH_DIAG emit each epoch: HEALTH_DIAG[N]: q_by_action [hold=X long=Y short=Z flat=W] Reads via host-side averaging of `q_out_buf [B, total_actions]` direction-branch columns [0..b0=4]. No new kernel needed — q_out_buf is already populated row-major by compute_expected_q. Modeled directly on the existing Task 0.3 magnitude-bucket diagnostic (same dtoh-and- average cold path). Action-index ordering canonical, see state_layout.cuh:123-126: DIR_SHORT=0, DIR_HOLD=1, DIR_LONG=2, DIR_FLAT=3. Emit slot order is [hold, long, short, flat] (Hold first because the hypothesis is about Hold's ascent dominating direction Q-magnitudes). New surface: - gpu_dqn_trainer.rs: q_dir_means_cached field + update_q_dir_means_cached method + q_direction_action_means accessor + free static helper compute_q_dir_means_from_host_buf (factored for testability). - fused_training.rs: FusedTrainingCtx wrappers parallel to q_mag pair. - training_loop.rs: emit block adjacent to q_var_per_branch. Behavioral test: sp16_phase0_q_by_action_diagnostic_reads_four_action_means seeds known per-column constants, asserts canonical-dir-idx → emit-slot mapping at 1e-3 tolerance, and includes sentinel-leak guard (non-direction columns at 99.0 fail any wrong index→slot mapping with margin >12). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> |
||
|
jgrusewski
|
9fb980da2b |
fixup(class-a-audit-batch-4b): invert MIN_HOLD_TEMPERATURE dir_acc → temp mapping
Post-commit kernel-formula audit of `compute_min_hold_penalty` at
trade_physics.cuh:567-577 showed the original `temp = TEMP_MIN +
(TEMP_MAX - TEMP_MIN) × skill` mapping was BACKWARDS for the
WR-plateau scenario this 8-commit chain targets.
Formula recap:
soft_factor = deficit / (deficit + T)
HIGH T → soft_factor → 0 → forgiving (no exit penalty)
LOW T → soft_factor → 1 → sharp (max exit penalty)
The deleted schedule's `START=50 → END=5` therefore meant
"permissive early, strict late" (classic explore-then-exploit), not
"permissive when committed, strict when uncertain" as the
substituted mapping assumed.
WR-plateau scenario: dir_acc ≈ 0.46-0.48 (model committed but
wrong) — the model needs a permissive (HIGH T) exit ramp to bail
out of bad committed directions, not maximum exit penalty (LOW T)
locking it into them.
Inverted the mapping using `confusion = 1 - skill`:
- dir_acc ≈ 0.5 (random / plateau) → confusion=1 → temp=50
(permissive, plateau exit ramp)
- dir_acc → 1.0 (saturated skill) → confusion=0 → temp=5
(strict, force informed commitment)
This preserves the deleted schedule's epoch-0 anchor (T=50
cold-start) while reacting to actual realized skill instead of an
epoch-time anchor.
Files touched (atomic per `feedback_no_partial_refactor`):
- min_hold_temperature_update_kernel.cu — formula + header doctring.
- sp14_isv_slots.rs — slot-doc comment expanded with new mapping
+ fixup-rationale paragraph.
- gpu_aux_trunk.rs — `MinHoldTemperatureUpdateOps` docstring.
- gpu_dqn_trainer.rs — cubin docstring + ISV_TOTAL_DIM doc-comment.
- tests/sp14_oracle_tests.rs — Tests 3 + 4 flipped (high dir_acc
now expects temp=TEMP_MIN; low dir_acc now expects TEMP_MAX);
Tests 1, 2, 5 numerically unchanged (sentinel guard fires before
the mapping; midpoint dir_acc=0.75 has skill=confusion=0.5 so
pre/post-fixup numerics match).
- docs/dqn-wire-up-audit.md — Item 4 entry rewritten with the
fixup rationale + mapping table updated.
Verification:
cargo check -p ml --tests --all-targets PASS
cargo test sp14_audit_4b (lib) 4/4 PASS (slot layout
locks unchanged)
GPU oracle re-run deferred to next L40S smoke (kernel was rebuilt
in-place so cubin contents change; the 5 oracles encode the new
mapping and will exercise the new cubin).
Per `pearl_first_observation_bootstrap.md` (sentinel→target
replace; sentinel anchors unchanged) +
`pearl_symmetric_clamp_audit.md` (bilateral clamp on target_temp
preserved) + `feedback_no_partial_refactor.md`.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
|
||
|
jgrusewski
|
0b9ea77dc4 |
fix(class-a-audit-batch-4b): plan_threshold floor adaptive + MIN_HOLD_TEMPERATURE ISV-driven
Per Class A audit-fix Batch 4-B (final 2 of 4 deferred items from
P1-wiring/P1-producer). Completes the 8-commit WR-plateau intervention
chain. Validation deferred to next L40S smoke.
Item 3: plan_threshold adaptive floor (Design Y - inline producer)
- NEW slot PLAN_THRESHOLD_FLOOR_ADAPTIVE_INDEX=459
- Writes slow-EMA shadow of `0.5 * readiness_ema` from inside the
existing update kernel (no new file, no new launch).
- Pearl-A first-observation bootstrap (sentinel 0.1 matches pre-fix
hardcoded value for bit-identical cold-start) + Welford alpha=0.005
slow EMA.
- Bilateral clamp [0.05, 0.50] (probability units) per
pearl_symmetric_clamp_audit.
- Consumer reads isv[459] as the floor in the same launch's final
fmaxf; cold-start sentinel REPLACES with threshold_target so the
pre-fix `fmaxf(0.1, 0.5*ema)` semantic is preserved bit-identical
for any readiness EMA above 0.20.
Item 4: MIN_HOLD_TEMPERATURE -> ISV-driven (driving signal: dir_acc skill)
- NEW slot MIN_HOLD_TEMPERATURE_ADAPTIVE_INDEX=460
- NEW kernel min_hold_temperature_update_kernel.cu (single-thread
cold-path, per-epoch boundary launch).
- Driving signal: dir_acc skill = clamp((short_ema - 0.5)/0.5, 0, 1)
from ISV[AUX_DIR_ACC_SHORT_EMA_INDEX=373]. When committing skillfully
(high dir_acc) -> temp HIGH (permissive). When at random baseline
(~0.5) -> temp LOW (sharp commitment pressure). Substituted for
the audit-spec's `dir_entropy_deficit` because no dir_entropy ISV
slot exists - dir_acc skill is the closest semantically-equivalent
signal that preserves the spec intent.
- Pearl-A bootstrap (sentinel 50.0 matches the deleted
MIN_HOLD_TEMPERATURE_START=50 anchor) + alpha=0.05 mid-cadence EMA.
- Bounds [5, 50] (matches the deleted schedule range).
- Decouples temperature from epoch number - the old schedule pinned
LOW temp (sharp) at end of training, exactly when a WR-plateaued
model needed forgiveness to escape.
- DELETED: state_layout.cuh::MIN_HOLD_TEMPERATURE_{START, END, DECAY}
#defines + training_loop.rs::min_hold_temperature_for_epoch helper
function (kept docstring tombstone explaining the deletion). Both
call sites migrated to the new ISV reader. Per
feedback_no_legacy_aliases + feedback_no_partial_refactor.
ISV_TOTAL_DIM: 459 -> 461.
Cumulative WR-plateau fix series (final commit, #8):
-
|
||
|
jgrusewski
|
7e9a8f6ef1 |
fix(class-a-audit-batch-4a): DD saturation floor adaptive + legacy DD path Case A
Per Class A audit-fix Batch 4-A (deferred from P1-wiring/P1-producer due to
audit-doc errors). Fixes 2 of 4 deferred items; Batch 4-B handles plan_threshold
floor + MIN_HOLD_TEMPERATURE in a separate commit.
Item 1: DD saturation floor (the upper end of the DD ramp at trade_physics.cuh:154
in apply_margin_cap, NOT line 548 as the audit doc claimed — that line is a
magnitude action constant; the actual saturation floor lives in apply_margin_cap)
- NEW slot DD_SATURATION_FLOOR_ADAPTIVE_INDEX=458
- Producer dd_saturation_floor_update_kernel.cu — p75(per-env DD_MAX) × 1.5
via Welford `mean + Z_75 × sigma` estimator with `max(p75, mean)` robustness
guard, mirrors P0-A REWARD_POS_CAP producer pattern (Pearl-A bootstrap +
Welford α=0.01)
- Cold-start fallback: 0.25f (DD_SATURATION_FLOOR_DEFAULT in state_layout.cuh)
- Bounds: [0.10, 0.50] (Category-1 dimensional safety)
- Distinct from SP15_DD_THRESHOLD_INDEX=421 (the SP15 quadratic DD-penalty
*trigger* threshold, a *lower* bound; this slot is the *upper* end of the
linear position-size scaling ramp dd_scale = max(0.05, 1.0 − dd_frac/floor))
- Threaded `isv_signals_ptr` into `apply_margin_cap` with NULL-tolerant
cold-start fallback to DD_SATURATION_FLOOR_DEFAULT
- 4 oracle tests (Pearl-A bootstrap, no-DD guard, bounds clamp, Welford EMA)
Item 2: Legacy compute_drawdown_penalty path → Case A (DELETED)
- Decision rationale: SP15's quadratic asymmetric DD penalty
(compute_sp15_final_reward_kernel.cu:154 via sp15_dd_penalty helper) runs
unconditionally as a post-modifier on the SP11-composed reward with
ISV-driven λ_dd (slot 420) and DD threshold (slot 421). Layering the legacy
linear-ramp penalty inside the SP11 composer on top of the SP15 quadratic
creates double-counting of DD shaping — exactly the code-smell the Class A
audit was designed to eliminate. Per `feedback_no_legacy_aliases.md` and
`feedback_no_partial_refactor.md`.
- Atomic deletion across:
- `compute_drawdown_penalty` device function (trade_physics.cuh)
- Single call site at experience_kernels.cu:3822
- `dd_threshold` and `w_dd` kernel arguments
- `w_dd` Rust config field (gpu_experience_collector.rs +
trainers/dqn/config.rs DQNHyperparameters)
- `w_dd` profile section + dispatch (training_profile.rs RewardSection,
OptimizableParameterRanges, FixedRewardParameters, ParamLookup
dispatch, profile→hyperparam mapping, test assertion)
- `w_dd *= rki` risk-intensity multiplier (config.rs)
- `w_dd` TOML keys (dqn-hyperopt.toml × 2, dqn-localdev.toml,
dqn-production.toml, dqn-smoketest.toml)
- Stale doc comments on hyperopt/adapters/dqn.rs + config.rs
risk_intensity field
- `config.dd_threshold` SURVIVES (still consumed by `launch_sp15_dd_state`
as the dd_budget for DD_PCT scaling). Documented in field comment.
ISV_TOTAL_DIM: 458 → 459 (Item 1 adds 1 slot; Item 2 is pure deletion)
Cumulative WR-plateau fix series (this is commit 7):
- Class C bug 1 + P0-B (
|
||
|
jgrusewski
|
87d597d5d7 |
fix(class-a-p1-producer): adaptive Bayesian Kelly priors per-fold-end
Replace 4 hardcoded Bayesian priors in Kelly cap calculations
(prior_wins=2.0, prior_losses=2.0, prior_sum_wins=0.01,
prior_sum_losses=0.01) with ISV-driven slow-EMA values fed by a new
producer kernel that aggregates realized PS_KELLY_* fields across envs
from the same portfolio_state buffer kelly_cap_update_kernel reads
from at the same per-epoch boundary.
Slots [454..458):
- KELLY_PRIOR_WINS, KELLY_PRIOR_LOSSES,
KELLY_PRIOR_SUM_WINS, KELLY_PRIOR_SUM_LOSSES
ISV_TOTAL_DIM bumps 454 -> 458.
Producer:
- kelly_bayesian_priors_update_kernel.cu (per-fold-end / per-epoch
boundary, block-tree-reduce in shmem, no atomicAdd). Pearl-A
first-observation bootstrap (sentinels 2.0/2.0/0.01/0.01 match
pre-P1-Producer hardcoded values for bit-identical cold-start) +
alpha=0.005 slow EMA. Bounds counts in [0.5, 100], sums in [0.001,
1.0] per feedback_isv_for_adaptive_bounds. Launched RIGHT BEFORE
launch_kelly_cap_update so that kernel sees the freshly-blended
priors.
Consumers migrated atomically (same commit):
- kelly_cap_update_kernel.cu:39-42 -> ISV[454..458) with cold-start
fallback via kelly_prior_or_default helper (range guard).
- trade_physics.cuh::kelly_position_cap:304-307 -> NULL-tolerant
isv_signals_ptr threaded through apply_kelly_cap (single caller in
unified_env_step_core line 898 already had the bus pointer; mirrors
the existing kelly_f_smooth / kelly_warmup_floor_sp9 patterns).
State reset:
- 4 FoldReset registry entries (sp14_p1_kelly_prior_*) +
4 dispatch arms in training_loop.rs::reset_named_state.
Oracle tests (sp14_oracle_tests.rs, GPU-gated #[ignore]):
- Pearl-A bootstrap (sentinel -> REPLACE with aggregated targets)
- No realized trades -> ISV preserved bit-exactly
- Bounds clamp on extreme aggregates
- Slow EMA blend after bootstrap
CPU tests passing:
- sp14_p1_kelly_prior_slot_layout_locked
- all_sp14_p1_slots_fit_within_isv_total_dim
- every_fold_and_soft_reset_entry_has_dispatch_arm (C.10 lesson)
- layout_fingerprint_bumps_after_sp14_wire
DEFERRED — Item 2 (MIN_HOLD_TEMPERATURE EMA): the audit-spec said
"MIN_HOLD_TEMPERATURE = 0.5f hardcoded somewhere" but the actual code
has MIN_HOLD_TEMPERATURE_{START=50.0f, END=5.0f, DECAY=20.0f} as a
PER-EPOCH ANNEALING SCHEDULE driven by min_hold_temperature_for_epoch
in training_loop.rs:68-73. The kernel takes T as a runtime scalar
specifically to enable Phase 2 ISV-driven lift "without recompiling
cubin" per the SP12 v3 design comment. The audit's claim of "0.5f
hardcoded" does not match reality; the right Phase 2 signal is a
separate spec decision and should not be guessed at per
feedback_no_quickfixes. Reporting back per the prompt's hard rule #8.
Per feedback_isv_for_adaptive_bounds + feedback_no_partial_refactor +
pearl_controller_anchors_isv_driven.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
|
||
|
jgrusewski
|
394de7d434 |
feat(class-a-p0a): REWARD_POS/NEG_CAP → ISV-driven adaptive caps from realized return distribution
Per Class A audit ranking, the highest-suspected-impact fix for the months-long WR-stuck-at-46-48% plateau across 11 superprojects. Hardcoded REWARD_POS_CAP=+5.0f / REWARD_NEG_CAP=-10.0f (state_layout.cuh:266-267) was structurally clipping the upper tail of realized alpha. Controller signals (sharpe EMA, var_q, q_gap) all derive from this CAPPED buffer, so the controller cannot select for trades it cannot see. Selectivity gradient evaporates — small wins clip to +5 alongside large wins also clipping to +5. The state_layout comment lines 261-264 explicitly deferred Phase 2 (ISV-driven) "IF Phase 1 validation reveals adaptive need". Phase 1 has been running 11 SPs without budging WR — adaptive need revealed. Architecture: - 2 new ISV slots [452..454): REWARD_POS_CAP_ADAPTIVE, REWARD_NEG_CAP_ADAPTIVE. - Producer kernel reward_cap_update_kernel.cu: block-tree-reduce Welford `mean + Z_99 × sigma` p99 estimator over winning realized returns + conservative `max(p99, max_win)` takeover, × 1.5 safety factor → POS cap. NEG cap = -2 × POS cap (preserves Kahneman 2:1 asymmetry per pearl_audit_unboundedness_for_implicit_asymmetry — asymmetry stays, but moved from hardcoded scalar to producer-time multiplier; single source of truth, no consumer applies the 2× ratio itself). - Pearl-A first-observation bootstrap from sentinel (5.0 / -10.0, matching pre-P0-A hardcoded values for bit-identical cold-start). Welford EMA α=0.01 thereafter (slow blend — reward distribution is the foundation of training and shouldn't move fast). - Bounds: POS in [1, 50], NEG in [-100, -2] (Category-1 dimensional safety per feedback_isv_for_adaptive_bounds, NOT tuning). - 3 consumer sites migrated atomically per feedback_no_partial_refactor: experience_kernels.cu:3112-3114 (segment_complete cap), compute_sp15_final_reward_kernel.cu:163 (Stage 4 helper invocation), sp15_reward_axis_helpers.cuh:211 (sp15_apply_sp12_cap device fn signature change to take isv ptr). - Cold-start fallback: when ISV slot at sentinel OR outside [REWARD_POS_CAP_MIN_BOUND=1, REWARD_POS_CAP_MAX_BOUND=50], consumers fall back to original macros (still defined in state_layout.cuh). - 2 new device-ptr accessors on the experience collector (step_ret_per_sample_dev_ptr, trade_close_per_sample_dev_ptr) — reuses existing per-sample buffers; no new buffer allocated. - Per-epoch boundary launch (cold path) in training_loop.rs alongside launch_aux_horizon_chain. - Reset registry entries + dispatch arms in reset_named_state per the C.10 lesson (missing dispatch causes runtime crash). - Layout fingerprint seed updated: ISV_TOTAL_DIM 452→454 + AUX_PRED_HORIZON_BARS=450 + AVG_WIN_HOLD_TIME_BARS=451 (previously missing from seed) + REWARD_POS_CAP_ADAPTIVE=452 + REWARD_NEG_CAP_ADAPTIVE=453. Per feedback_isv_for_adaptive_bounds: every adaptive bound in ISV. Verification: - cargo check -p ml --tests --all-targets: clean (19 pre-existing warnings, 0 new). - sp14_isv_slots tests: 8/8 pass (4 layout + 4 fits-within). - sp14_oracle_tests with --features cuda --ignored: 8/8 pass (4 existing q_disagreement/dir_concat + 4 new P0-A tests covering Pearl-A bootstrap, no-winning-trades preservation, bounds clamping to [1,50], Welford α=0.01 EMA blend). - sp15_phase1_oracle_tests with --features cuda --ignored: 36/36 pass (no regression from sp15_apply_sp12_cap signature change). Cumulative WR-plateau fix series: - Class C bug 1 ( |
||
|
jgrusewski
|
4f372a49a8 |
refactor(sp14-c): atomic α machinery deletion + aux trunk ISV slot allocation
Phase C.1 of SP14 Layer C separate-aux-trunk refactor. Single atomic
commit per feedback_no_partial_refactor and feedback_no_legacy_aliases —
no DEPRECATED stage.
Deleted (per C.0 audit
|
||
|
jgrusewski
|
9d0c124cee |
fix(sp14-egf): gate q_disagreement EMA update on total_cnt > 0 — fixes training-time decay-to-zero of rollout signal
Root cause from train-6fcml 5-epoch trajectory (commit
|
||
|
jgrusewski
|
c0fc28e455 |
fix(sp14): delete warmup_gate — let variance-driven k_aux/k_q handle warmup (ISV-driven)
Per `feedback_isv_for_adaptive_bounds`, the hardcoded `warmup_gate = (fold_step_counter / WARMUP_STEPS_FALLBACK).min(1.0)` ramp violated the rule: adaptive bounds in ISV, never hardcoded constants. The variance-driven k_aux/k_q sigmoid steepness already provides warmup behavior intrinsically: - High variance (cold-start, EMAs still moving) → k → K_MIN → flat sigmoid → gate ≈ 0.5 regardless of input. That IS the warmup. - Low variance (settled) → k → K_BASE → sharp sigmoid → gates respond correctly to driver signals. Adding a separate hardcoded step-counter multiplier on top was double-counting + tuning-driven (the 1000-step threshold had no principled basis). Removed entirely. Removed (per `feedback_no_partial_refactor`, all atomically): - `WARMUP_STEPS_FALLBACK` constant in `sp14_isv_slots.rs` - `warmup_gate: f32` parameter in `alpha_grad_compute_kernel.cu` - `gate1 * gate2 * warmup_gate` → `gate1 * gate2` in kernel - `warmup_gate` arg from `launch_sp14_alpha_grad_compute` - `fold_step_counter: usize` field on the trainer struct - `fold_step_counter = 0` reset in `reset_for_fold` - `fold_step_counter` init in trainer constructor - `let warmup_gate: f32 = 1.0;` and `.arg(&warmup_gate)` from B.4 oracle tests (4 launches: 2 in alpha_grad_schmitt_hysteresis, 20 in alpha_grad_adaptive_beta loop) Build: clean, 18 warnings (pre-existing baseline). Tests: cargo test --no-run on sp14_oracle_tests succeeds. Net result: EGF gate's warmup behavior now lives entirely in the variance-driven k_aux/k_q sigmoid steepness controller (ISV slots 388/var_aux, 389/var_q). No hardcoded step counter. Honors `feedback_isv_for_adaptive_bounds` and `pearl_controller_anchors_isv_driven`. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> |
||
|
jgrusewski
|
857722e774 |
feat(sp14): B.11 — orchestrator wire-up for 3 EGF producer kernels + var_aux gap closure
Per-step launches (in graph capture order): 1. Forward (existing) 2. Action select (existing) → q_dir_logits available 3. launch_sp14_q_disagreement_update → ISV[383, 384, 389] 4. launch_sp14_alpha_grad_compute → ISV[385..395] (consumes q_disagreement) 5. Backward (existing) — wire-col scale at B.10 reads ISV[393] Per-epoch launch (end of epoch): 6. launch_sp14_gradient_hack_detect → circuit breaker α_short=0.3, α_long=0.05, α_var=0.05 per spec; warmup_gate derived from steps_in_fold / WARMUP_STEPS_FALLBACK. Var_aux producer gap closed (option C from B.4): alpha_grad_compute_kernel now also writes ISV[VAR_AUX_INDEX=388] via Welford EMA against (aux_dir_acc_short - aux_dir_acc_long). Adaptive k_aux is now functional (was degenerate at K_BASE_AUX=20.0 constant pre-B.11). Closes the "adaptive_k_aux currently degenerate" concern flagged in B.4 commit. After this commit, the EGF pearl is FULLY ACTIVE end-to-end: - Forward: aux signal feeds direction Q-head input (B.8/B.9) - Backward: wire-col gradient gated by α_grad_smoothed (B.10) - Producers: α_grad computed every step from real driver signals (B.11) - Pre-B.11 force-closed gate (sentinel 0.0) → post-B.11 responsive gate Build clean: 18 warnings pre-existing baseline, 0 new. Tests: 4/4 P0b aux_w tests pass (no regression). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> |
||
|
jgrusewski
|
6715ab4ea1 |
feat(sp14): B.8 — direction Q-head input dim SH2 → SH2+1 + fingerprint
Forward wire (weight-side only — dispatch consumer lands in B.9): direction Q-head's first FC weight tensor `w_b0fc` (param-table index 17) input dim grows by 1 to accept aux_softmax_diff. The (rows, cols) shape table at the trainer-init Xavier site is updated in lock-step with `compute_param_sizes()`. Output dim (adv_h) and the bias `b_b0fc` are unchanged — bias is per-output, not per-input. `layout_fingerprint_seed()` entry renamed `PARAM_W_B0FC` → `PARAM_W_B0FC_AUX1`, forcing the FNV-1a hash to bump per Invariant 8 (old checkpoints fail-fast at load via `check_layout_fingerprint` instead of silently aliasing onto the new architecture). Per `feedback_no_legacy_aliases`, no `_DEPRECATED` shim — straight in-place rename. New weight column zero-initialised in trainer construction (mirrors the OFI column zeroing for w_b2fc/w_b3fc): the model starts ignoring the new input and learns to use it through gradient descent. Xavier- init on this column would inject day-0 noise the trunk would have to denoise — let the EGF gate decide when the aux signal is trustworthy. Atomic-migration consumers (`feedback_no_partial_refactor`) updated: - `gpu_dqn_trainer.rs` — Xavier (rows, cols) table at index 17 grows to (adv_h, SH2+1); spectral-norm descriptor entry [4] for W_a1 grows in_dim from sh2 to sh2+1; spec_v_a1 power-iteration vector grows from sh2 to sh2+1 floats. - `dqn/smoke_tests/gradient_budget.rs` — both `alloc_dueling` fixtures' slot 8 grow `cfg.adv_h * cfg.shared_h2` → `cfg.adv_h * (cfg.shared_h2 + 1)`. - `docs/dqn-wire-up-audit.md` — new SP14 Layer B B.8 entry per Invariant 7. Note: forward GEMM dispatch still uses `K = shared_h2` until B.9 lands the concat-then-SGEMM consumer (per plan §2358). Until then the new column reads as ignored padding; this is safe because (a) it's zero- initialised, (b) GPU-only smoke tests are skipped on this CPU CI, (c) the fingerprint bump invalidates any pre-SP14 checkpoint that would attempt to load. Test: `layout_fingerprint_bumps_after_sp14_wire` (CPU-only, in `sp14_oracle_tests.rs`) hashes the pre-B.8 seed verbatim with the single difference `PARAM_W_B0FC` (vs post-B.8 `PARAM_W_B0FC_AUX1`) and asserts `LAYOUT_FINGERPRINT_CURRENT` differs — any silent revert of the rename trips this test. Mirrors the `fingerprint_bumped_from_ pre_b1_1a` pattern from `sp13_layer_b_oracle_tests.rs`. Verified: - `SQLX_OFFLINE=true cargo check -p ml` clean, 18 warnings (baseline) - `cargo test -p ml --test sp14_oracle_tests` 2 passed, 6 ignored (GPU) - `cargo test -p ml --test sp13_layer_b_oracle_tests fingerprint_bumped_from_pre_b1_1a` still passes (sister test) Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> |
||
|
jgrusewski
|
4527d8c852 |
feat(sp14): B.6 — dir_concat_qaux_kernel (pre-SGEMM forward-wire concat)
Mirrors mag_concat_qdir precedent (experience_kernels.cu:4560). Concats [h_s2 ; aux_softmax_diff] into scratch buffer [B, SH2+1] for the direction Q-head's first FC SGEMM. aux_softmax_diff = softmax[b, 1] - softmax[b, 0] in [-1, +1] computed inline; structurally bounded by softmax components per pearl_bounded_modifier_outputs_require_structural_activation. Pure per-thread map; no reduction; no atomicAdd. Does not read or write ISV slots — purely data-movement. Launch order constraint: aux head forward MUST complete before this concat reads aux_nb_softmax_buf (enforced by orchestrator in B.10/B.11). Test: dir_concat_qaux_correct verifies row-wise contiguous concat + correct softmax diff values for both 'down' (-0.8) and 'up' (+0.8) synthetic aux predictions. B.3+B.4+B.5 regression: 5 GPU tests unchanged (6 total pass). Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> |
||
|
jgrusewski
|
82fe6cea66 |
feat(sp14): B.5 — gradient_hack_detect_kernel (anti-mesa-opt circuit breaker)
Detects suspected gradient hacking: when gate1 is open AND aux_dir_acc post-open-minimum drops > LOCKOUT_TRIGGER_DROP (0.05) below the Schmitt open-threshold (target + SCHMITT_BAND = target + 0.03) AND q_disagreement rises > LOCKOUT_TRIGGER_DIS_RISE (0.10) above the analytic random-alignment baseline 0.5, simultaneously. Action: force gate1_open_state = 0 (ISV[391]); set lockout_remaining = 2.0 epochs (LOCKOUT_EPOCHS). During lockout, gate1 stays force-closed regardless of alpha_grad_compute_kernel output. Tracks AUX_DIR_ACC_POST_OPEN_MIN (ISV[394]): running minimum of aux_dir_acc since gate1 last opened; resets to 1.0 sentinel when gate closes naturally or when circuit breaker fires. Slot indices shifted +2 from original plan (SP13 closeout added HOLD_RATE_TARGET=381 + HOLD_RATE_OBSERVED_EMA=382): Q_DIS_SHORT=383, GATE1=391, POST_OPEN_MIN=394, LOCKOUT=395. Matches sp14_isv_slots.rs. Single-thread state-machine kernel (threadIdx.x==0 guard); runs at end of each epoch after alpha_grad_compute_kernel. No atomicAdd per feedback_no_atomicadd.md. 1 oracle test: gradient_hack_circuit_breaker_fires verifies trigger conditions (aux_drop=0.08 > 0.05, q_rise=0.15 > 0.10) cause lockout=2.0 and gate1 force-close=0.0. B.3+B.4 regression: 4 GPU tests unchanged (5 total GPU pass, 1 host pass). Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> |
||
|
jgrusewski
|
49cdf90ecc |
feat(sp14): B.4 — alpha_grad_compute_kernel (EGF heart)
Single-thread state-machine kernel that is the heart of the Earned Gradient Flow pearl. Reads driver signals from the global ISV bus, runs Schmitt-trigger Gate 1, computes adaptive k_aux/k_q/β, evaluates two sigmoids, multiplies with a host-supplied warmup gate, applies a β-rate-limiter, and writes 7 outputs back to ISV. Per-step pipeline: 1. Read aux_dir_acc (slot 373), q_disagreement (slot 383), Welford variance EMAs (388, 389, 390), persistent Schmitt state (391), alpha_smoothed_prev (393). 2. Compute adaptive k_aux = K_BASE_AUX/(1 + var_aux/VARIANCE_REF_AUX) and k_q analogously (B.2.5; floor at K_MIN = 1.0). 3. Run Schmitt-trigger Gate 1 state update (open at target+0.03, close at target-0.03; intentional discontinuity at transition is smoothed by the β rate-limiter downstream). 4. Evaluate Gate 1 sigmoid (aux competence, distance from threshold) and Gate 2 sigmoid (Q-aux disagreement vs analytic 0.5 baseline). 5. alpha_grad_raw = gate1 × gate2 × warmup_gate (structurally bounded to [0, 1] per pearl_bounded_modifier_outputs_require_structural_ activation; no runtime clamp). 6. Update Welford variance of alpha_grad_raw → adaptive β (B.2.8; floor BETA_BASE = 0.5, ceiling BETA_MAX = 0.95). 7. alpha_grad_smoothed = β × prev + (1-β) × raw (rate-limited). 8. Write back 7 outputs: k_aux (385), k_q (386), β (387), var_alpha (390), gate1_state (391), alpha_raw (392), alpha_smoothed (393). Sigmoid arguments clipped to [-30, 30] before __expf for fp32 overflow guard (precision-neutral; sigmoid saturates bit-equal at those bounds). Per pearl_bounded_modifier_outputs_require_structural_activation: sigmoid composition produces structurally-bounded [0, 1] output. KNOWN LIMITATION: as of B.4 landing, NO upstream kernel writes ISV[388] (AUX_DIR_ACC_VARIANCE_EMA). The grep at status-report time finds only the sp14_isv_slots.rs declaration. Effect: var_aux stays at sentinel 0.0 forever, so k_aux is degenerate-but-non-fatal at K_BASE_AUX (constant). Gate 1 still works, the sigmoid just doesn't soften under noisy aux_dir_acc. To be resolved in B.11 producer- chain orchestrator OR a separate fix-up task that adds a Welford- variance update next to the existing AUX_DIR_ACC_SHORT_EMA producer. var_q (389) IS written by q_disagreement_update_kernel (B.3), so adaptive k_q is fully functional from B.4 onward. Slot indices hardcoded inside the kernel via const int locals — must match crates/ml/src/cuda_pipeline/sp14_isv_slots.rs (and 372/373 from sp13_isv_slots.rs). The plan originally documented 381/383/ 384/385/386/387/388/389/390/391 for SP14 slots; the actual values are +2 because SP13 closeout added HOLD_RATE_TARGET=381 + HOLD_RATE_OBSERVED_EMA=382 after the plan was written. Tests (RTX 3050 Ti pass; B.3's 2 tests still pass — no regression): - alpha_grad_schmitt_hysteresis: 4-step trajectory verifies the closed→open→open→closed transition. Closed at aux=0.55 (below open=0.58); opens at aux=0.60; stays open at aux=0.54 (in hysteresis band [close=0.52, open=0.58]); finally closes at aux=0.50 (below close=0.52). - alpha_grad_adaptive_beta: 20-oscillation regime verifies β grows above β_base=0.5 and remains bounded by β_max=0.95. docs/dqn-wire-up-audit.md updated per Invariant 7 with full B.4 behaviour contract, per-step pipeline, single-thread launch convention, sigmoid clip rationale, Schmitt discontinuity note, and the var_aux Known Limitation. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> |
||
|
jgrusewski
|
d3a35cc6e9 |
feat(sp14): B.3 — q_disagreement_update_kernel with K=4↔K=2 mapping
First of four producer kernels in the Earned Gradient Flow pearl chain.
Per-step computes the K=4↔K=2 mapped argmax mismatch between the Q-head's
4-way direction action (DIR_SHORT/DIR_HOLD/DIR_LONG/DIR_FLAT) and the
aux head's 2-way next-bar prediction (down/up), updates fast + slow
EMAs of the disagreement rate, and updates a Welford-style variance
EMA on the same signal in a single launch.
Mapping (per state_layout.cuh):
DIR_SHORT (=0) → aux down (=0) contributes
DIR_HOLD (=1) → masked no contribution
DIR_LONG (=2) → aux up (=1) contributes
DIR_FLAT (=3) → masked no contribution
Hold and Flat are masked because they represent "no NEW directional
commitment" (Hold = keep prior position; Flat = close all positions).
Penalising the aux head for not matching them would conflate
position-management actions with directional predictions.
Block tree-reduce on numerator + count separately, then divide and
update EMAs in a single thread (tid == 0). No atomicAdd per
feedback_no_atomicadd.md. Pure GPU compute per
feedback_no_cpu_compute_strict.md.
Pearl-A first-observation: when ISV[Q_DISAGREEMENT_SHORT_EMA=383] AND
ISV[Q_DISAGREEMENT_LONG_EMA=384] both equal sentinel 0.5f AND batch
has at least one valid contribution (total_cnt > 0), both EMAs are
replaced directly with the first observation per
pearl_first_observation_bootstrap.md. The 0.5f exact-match is safe
because 0.5 is exactly representable in IEEE 754 single precision
(mantissa = 1.0, exponent = -1). The Welford variance EMA at slot
389 drives the adaptive k_q sigmoid steepness consumed by the
alpha_grad_compute_kernel in B.4.
Slot indices are hardcoded inside the kernel via #define — must match
crates/ml/src/cuda_pipeline/sp14_isv_slots.rs (currently 383, 384, 389).
The kernel header documents the coupling explicitly.
Launch contract:
grid_dim = (1, 1, 1)
block_dim = (256, 1, 1)
shared_mem_bytes = 2 * 256 * sizeof(float) = 2048
A shared_mem_bytes = 0 launch reads garbage and corrupts the EMA — the
kernel header documents the launcher requirement; oracle tests pass
2048 explicitly.
Files:
- crates/ml/src/cuda_pipeline/q_disagreement_update_kernel.cu (NEW)
- crates/ml/build.rs — register cubin in kernels_with_common
- crates/ml/tests/sp14_oracle_tests.rs — append #[cfg(feature="cuda")]
mod gpu with cubin handle + 2 GPU oracle tests
- docs/dqn-wire-up-audit.md — SP14 B.3 section (Invariant 7)
Tests (both pass on RTX 3050 Ti):
- q_disagreement_k4_k2_mapping: 8-row batch with 2 agreements,
2 disagreements, 4 masked Hold/Flat → first-obs Pearl-A replaces
both EMAs with batch_mean = 0.5 (asserted within 1e-4)
- q_disagreement_all_hold_no_contribution: all-Hold edge case;
total_cnt = 0 so batch_mean = 0/1 = 0; sentinel-bootstrap guard
(total_cnt > 0) keeps EMA from collapsing to 0; current kernel
blends to 0.35, test bound [0.0, 0.5] tolerant of either current
blend or future skip-update refinement, asserts is_finite()
Wire-up status: producer kernel exists and is exercised only by the
oracle tests. The Rust launcher and graph-capture integration land in
B.7+ alongside the consumer (alpha_grad_compute in B.4 reads slots
383/384/389). This is one producer kernel of four (B.3 q_disagreement,
B.4 alpha_grad_compute, B.5 gradient_hack_detect, B.6 dir_concat_qaux);
known-orphan for the duration of the producer chain per
feedback_wire_everything_up.md (the same-commit wire-up rule is
relaxed for atomic chained-producer-consumer landings, with each
orphan documented in the audit doc; this orphan is acknowledged in
docs/dqn-wire-up-audit.md SP14 B.3 section).
Build + test verification:
- SQLX_OFFLINE=true CUDA_COMPUTE_CAP=86 cargo check -p ml --features cuda
→ clean (only the 18 pre-existing warnings)
- SQLX_OFFLINE=true cargo test -p ml --test sp14_oracle_tests
set_aux_weight → host-only A.2 still passes (no shared dependency)
- SQLX_OFFLINE=true CUDA_COMPUTE_CAP=86 cargo test -p ml
--test sp14_oracle_tests --features cuda q_disagreement
-- --ignored --nocapture → 2 PASS
Slot indices reflect the SP13-closeout +2 shift documented in
sp14_isv_slots.rs (the original plan's 381/382/387 became 383/384/389
because HOLD_RATE_TARGET=381 and HOLD_RATE_OBSERVED_EMA=382 landed
between when the plan was written and B.1 was implemented).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
|
||
|
jgrusewski
|
731cae4c80 |
fix(sp14): A.2 — lift set_aux_weight clamp to SP13 P0b range [0.15, 1.5]
The pre-fix clamp at gpu_dqn_trainer.rs:14722 was [0.05, 0.3] — the SP11-era cap. P0b's controller computes aux_w in [0.15, 1.5] (base 0.5 × [0.3, 3.0]) but the setter silently chopped everything above 0.3, masking the deficit-amplification term entirely. Smoke A trace confirmed: Fold 0/1: raw aux_w = 0.66-0.80 → clamped to 0.30 (deficit invisible) Fold 2: raw aux_w = 0.164 (stagnation; below clamp) → 45% deficit Post-fix: deficit-amp term `(1 + 5 × deficit)` actually expresses through to the trainer. Fold 2 stagnation will get the designed floor 0.15 instead of being capped at 0.3 — but the upper range 1.5 also opens, so deficit-amp can pull aux_w up when accuracy is below target. Constants imported from sp13_isv_slots.rs (AUX_W_BASE=0.5, AUX_W_HARD_FLOOR_RATIO=0.3, AUX_W_HARD_CEIL_RATIO=3.0). No new slots; existing constants exposed as the clamp bounds. Test: set_aux_weight_clamp_range verifies constants resolve correctly. A.1 (C51 atom-probability floor) deferred — Phase 0 verification found the spec's stated `−log(p)/p · ∂p/∂z` divide does NOT exist in c51_grad_kernel.cu at HEAD 037c24116; actual kernel uses the numerically-stable `expf(lp) - proj`. The 1109 GRAD_CLIP_OUTLIER events in Smoke A are real but their mechanism is different. Will be re-spec'd as a separate task post-SP14. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> |