feat(sp11): A0 — allocate 20 ISV slots [340..360) + 20 reset entries
Pure infrastructure. No producer kernels, no consumer reads. Existing training paths trace identically because no consumer reads slots [340..360) yet. Layout-fingerprint bumped to ISV_TOTAL_DIM=360. Spec: docs/superpowers/specs/2026-05-04-sp11-reward-as-controlled-subsystem.md
This commit is contained in:
jgrusewski
@@ -903,7 +903,7 @@ const ISV_NETWORK_DIM: usize = 23;
|
||||
/// (shifted 112→116 in Plan 4 Task 6 Commit A).
|
||||
/// Written by the constructor; checked at checkpoint load. Fail-fast only — no migration
|
||||
/// path exists. See spec §4.A.2 and `LAYOUT_FINGERPRINT_CURRENT` for structural-hash rationale.
|
||||
pub(crate) const ISV_TOTAL_DIM: usize = 340; // SP5 + Layer D D1+D2+D3 + SP7 + SP8 + SP9 + SP10: 173 + 167 (164 SP5 slots @ 174..278/280..340, with 2-slot gap before cross-fold-persistent Kelly block; Layer D D1 PnL outputs at [286..290); Layer D D2 health composition outputs at [290..294); Layer D D3 training metrics EMA at [294..297); SP7 T1 loss-balance Wiener stats at [297..313); SP7 activation-flag fix per-(head,branch) flags at [313..321); SP8 Fix 36 train_active_frac canary @ [321..322) + LB_MAX_BUDGET per-(head,branch) at [322..330); SP9 Fix 37 Kelly warmup floor at [330..331) + Q_VAR_MAG_EMA at [331..332) + INTENT_EVAL_DIVERGENCE at [332..333) + 3 EMA targets at [333..336) + 3 eval_dist mag bins at [336..339); SP10 Fix 38 EVAL_THOMPSON_TEMP at [339..340))
|
||||
pub(crate) const ISV_TOTAL_DIM: usize = 360; // SP5 + Layer D D1+D2+D3 + SP7 + SP8 + SP9 + SP10 + SP11: 173 + 187 (164 SP5 slots @ 174..278/280..340, with 2-slot gap before cross-fold-persistent Kelly block; Layer D D1 PnL outputs at [286..290); Layer D D2 health composition outputs at [290..294); Layer D D3 training metrics EMA at [294..297); SP7 T1 loss-balance Wiener stats at [297..313); SP7 activation-flag fix per-(head,branch) flags at [313..321); SP8 Fix 36 train_active_frac canary @ [321..322) + LB_MAX_BUDGET per-(head,branch) at [322..330); SP9 Fix 37 Kelly warmup floor at [330..331) + Q_VAR_MAG_EMA at [331..332) + INTENT_EVAL_DIVERGENCE at [332..333) + 3 EMA targets at [333..336) + 3 eval_dist mag bins at [336..339); SP10 Fix 38 EVAL_THOMPSON_TEMP at [339..340); SP11 Fix 39 reward-subsystem controller at [340..360) — 6 component weights [340..346) + 4 controller scalar outputs [346..350) + 2 val-sharpe canaries [350..352) + 6 mag-ratio canaries [352..358) + saboteur engagement [358] + PnL magnitude EMA [359])
|
||||
/// Legacy alias preserved for call sites that haven't been audited for the
|
||||
/// network-vs-total split. New code should pick `ISV_NETWORK_DIM` (for weight
|
||||
/// tensor sizing) or `ISV_TOTAL_DIM` (for the broadcast bus buffer).
|
||||
@@ -1838,7 +1838,14 @@ const fn layout_fingerprint_seed() -> &'static [u8] {
|
||||
KELLY_DIVERGENCE_TARGET=334;KELLY_TEMPORAL_TARGET=335;\
|
||||
EVAL_DIST_Q=336;EVAL_DIST_H=337;EVAL_DIST_F=338;\
|
||||
EVAL_THOMPSON_TEMP=339;\
|
||||
ISV_TOTAL_DIM=340;\
|
||||
REWARD_POPART_WEIGHT=340;REWARD_CF_WEIGHT=341;REWARD_TRAIL_WEIGHT=342;\
|
||||
REWARD_MICRO_WEIGHT=343;REWARD_OPP_COST_WEIGHT=344;REWARD_BONUS_WEIGHT=345;\
|
||||
CURIOSITY_PRESSURE=346;SABOTEUR_INTENSITY_MULT=347;\
|
||||
REWARD_WEIGHT_FLOOR=348;CURIOSITY_BOUND=349;\
|
||||
VAL_SHARPE_DELTA_EMA=350;VAL_SHARPE_VAR_EMA=351;\
|
||||
REWARD_COMPONENT_MAG_RATIO_BASE=352;\
|
||||
SABOTEUR_ENGAGEMENT_RATE=358;PNL_REWARD_MAGNITUDE_EMA=359;\
|
||||
ISV_TOTAL_DIM=360;\
|
||||
PARAM_W_A_H_S1=0;PARAM_B_A_H_S1=1;PARAM_W_B_H_S1=2;PARAM_B_B_H_S1=3;\
|
||||
PARAM_W_RESIDUAL_H_S1=4;PARAM_GAMMA_H_S1=5;PARAM_BETA_H_S1=6;\
|
||||
PARAM_W_A_H_S2=7;PARAM_B_A_H_S2=8;PARAM_W_B_H_S2=9;PARAM_B_B_H_S2=10;\
|
||||
|
||||
@@ -60,6 +60,7 @@ pub mod q_snapshot;
|
||||
pub mod meta_q_network;
|
||||
pub mod sp4_isv_slots;
|
||||
pub mod sp5_isv_slots;
|
||||
pub mod sp11_isv_slots;
|
||||
pub use sp4_isv_slots::{
|
||||
TARGET_Q_BOUND_INDEX, ATOM_POS_BOUND_BASE, WEIGHT_BOUND_BASE,
|
||||
ADAM_M_BOUND_BASE, ADAM_V_BOUND_BASE, WD_RATE_BASE,
|
||||
|
||||
73
crates/ml/src/cuda_pipeline/sp11_isv_slots.rs
Normal file
73
crates/ml/src/cuda_pipeline/sp11_isv_slots.rs
Normal file
@@ -0,0 +1,73 @@
|
||||
//! SP11 — Reward-subsystem controller ISV slot constants.
|
||||
//!
|
||||
//! Slots [340..360) populated by the SP11 controller chain:
|
||||
//! - 6 component weights [340..346)
|
||||
//! - 4 controller scalar outputs [346..350)
|
||||
//! - 2 val-sharpe canaries [350..352)
|
||||
//! - 6 mag-ratio canaries [352..358)
|
||||
//! - 1 saboteur engagement (358)
|
||||
//! - 1 PnL magnitude EMA (359)
|
||||
//!
|
||||
//! Spec: docs/superpowers/specs/2026-05-04-sp11-reward-as-controlled-subsystem.md
|
||||
//! Pearl: pearl_reward_as_controlled_subsystem.md
|
||||
|
||||
// Component weights (controller outputs, consumed by reward shaping kernels)
|
||||
pub const REWARD_POPART_WEIGHT_INDEX: usize = 340;
|
||||
pub const REWARD_CF_WEIGHT_INDEX: usize = 341;
|
||||
pub const REWARD_TRAIL_WEIGHT_INDEX: usize = 342;
|
||||
pub const REWARD_MICRO_WEIGHT_INDEX: usize = 343;
|
||||
pub const REWARD_OPP_COST_WEIGHT_INDEX: usize = 344;
|
||||
pub const REWARD_BONUS_WEIGHT_INDEX: usize = 345;
|
||||
|
||||
// Controller scalar outputs
|
||||
pub const CURIOSITY_PRESSURE_INDEX: usize = 346;
|
||||
pub const SABOTEUR_INTENSITY_MULT_INDEX: usize = 347;
|
||||
pub const REWARD_WEIGHT_FLOOR_INDEX: usize = 348;
|
||||
pub const CURIOSITY_BOUND_INDEX: usize = 349;
|
||||
|
||||
// Canaries — val-sharpe trend + variance
|
||||
pub const VAL_SHARPE_DELTA_EMA_INDEX: usize = 350;
|
||||
pub const VAL_SHARPE_VAR_EMA_INDEX: usize = 351;
|
||||
|
||||
// Canaries — per-component reward-magnitude ratios (6 contiguous slots).
|
||||
// Source: REWARD_POPART_EMA_INDEX..+6 (existing per-component reward
|
||||
// magnitude EMAs from reward_component_ema_kernel). Spec §3.3.1 said
|
||||
// "grad EMA" but the SP4 grad-balancer is per-branch (4 slots), not
|
||||
// per-reward-component. Per `feedback_trust_code_not_docs`: code wins,
|
||||
// terminology corrected. Semantics for the controller are equivalent —
|
||||
// "which component contributes most to total reward" maps the same way
|
||||
// in the winner_weight / diversifier_weight blend.
|
||||
pub const REWARD_COMPONENT_MAG_RATIO_BASE: usize = 352;
|
||||
pub const REWARD_COMPONENT_MAG_RATIO_COUNT: usize = 6;
|
||||
|
||||
// Canaries — saboteur + PnL signal scale
|
||||
pub const SABOTEUR_ENGAGEMENT_RATE_INDEX: usize = 358;
|
||||
pub const PNL_REWARD_MAGNITUDE_EMA_INDEX: usize = 359;
|
||||
|
||||
// Invariant-1 anchors (rate-limiters, NOT regime thresholds).
|
||||
//
|
||||
// Per `feedback_isv_for_adaptive_bounds`: every adaptive bound lives on
|
||||
// ISV. The constants below are the few exceptions allowed by Invariant-1
|
||||
// — they are numerical-stability anchors (epsilon floors), structural
|
||||
// fractions of signal-driven quantities (curiosity-bound = 0.3 × PnL EMA),
|
||||
// and rate-limiter clamps that bound the saboteur multiplier output. Each
|
||||
// is documented in the SP11 spec §3.4.2 (Invariant-1 carve-out table).
|
||||
pub const SP11_EPS_DIV: f32 = 1e-6;
|
||||
pub const SP11_WEIGHT_HARD_FLOOR: f32 = 0.01;
|
||||
pub const SP11_SABOTEUR_MIN: f32 = 0.5;
|
||||
pub const SP11_SABOTEUR_MAX: f32 = 2.0;
|
||||
pub const SP11_CURIOSITY_PERMANENT_FRACTION: f32 = 0.2;
|
||||
pub const SP11_CURIOSITY_BOUND_FRACTION: f32 = 0.3;
|
||||
pub const SP11_WEIGHT_FLOOR_FRACTION: f32 = 0.5;
|
||||
pub const SP11_ENGAGEMENT_FLOOR: f32 = 0.1;
|
||||
|
||||
pub const SP11_SLOT_BASE: usize = 340;
|
||||
pub const SP11_SLOT_END: usize = 360;
|
||||
pub const SP11_PRODUCER_COUNT: usize = SP11_SLOT_END - SP11_SLOT_BASE;
|
||||
|
||||
/// Salt added to `config.seed` when seeding the novelty SimHash projection
|
||||
/// matrix init kernel. Co-located here so the seed-derivation path is
|
||||
/// reproducible from a single source of truth. Consumed by the launcher
|
||||
/// `launch_novelty_simhash_proj_init` (added in Task A2 alongside the
|
||||
/// projection buffer).
|
||||
pub const SP11_PROJ_SEED_SALT: u64 = 0x_5511_0001;
|
||||
@@ -18,8 +18,15 @@
|
||||
//! 321..330 SP8 (Fix 36) train_active_frac canary + LB_MAX_BUDGET (9 slots, fold-reset)
|
||||
//! 330..339 SP9 (Fix 37) Kelly warmup floor + EMA targets + eval_dist (9 slots, fold-reset)
|
||||
//! 339..340 SP10 (Fix 38) eval Thompson temperature (1 slot, fold-reset)
|
||||
//! 340..360 SP11 (Fix 39) reward-subsystem controller (20 slots, fold-reset) — see sp11_isv_slots
|
||||
//!
|
||||
//! Total: 164 new SP5 ISV slots (52 + 24 + 20 + 4 + 4 + 6 + 4 + 4 + 3 + 16 + 8 + 9 + 9 + 1).
|
||||
//! Total: 184 new SP5 ISV slots (52 + 24 + 20 + 4 + 4 + 6 + 4 + 4 + 3 + 16 + 8 + 9 + 9 + 1 + 20).
|
||||
|
||||
// Re-export SP11 slot constants so existing consumers can import every
|
||||
// slot constant from a single location (`crate::cuda_pipeline::sp5_isv_slots::*`).
|
||||
// SP11 owns its file but participates in the SP5 linear-span window
|
||||
// (SP5_SLOT_BASE..SP5_SLOT_END covers all SP5/SP7/SP8/SP9/SP10/SP11 slots).
|
||||
pub use crate::cuda_pipeline::sp11_isv_slots::*;
|
||||
|
||||
pub const SP5_SLOT_BASE: usize = 174;
|
||||
|
||||
@@ -281,7 +288,21 @@ pub const EVAL_DIST_F_INDEX: usize = 338; // [1] eval-side Full m
|
||||
// from the first observation.
|
||||
pub const EVAL_THOMPSON_TEMP_INDEX: usize = 339; // [1] eval Thompson selector temperature
|
||||
|
||||
pub const SP5_SLOT_END: usize = 340;
|
||||
// ── SP11 (Fix 39, 2026-05-04): reward-subsystem controller ────────────
|
||||
//
|
||||
// Per `pearl_reward_as_controlled_subsystem.md` + `pearl_controller_anchors_isv_driven.md`:
|
||||
// the reward subsystem is migrated from a fixed-weight composition (host-side
|
||||
// constants in `unified_reward_kernel.cu` and friends) to a fully ISV-driven
|
||||
// controller producing 6 component weights + 4 controller scalars + 10 canaries.
|
||||
//
|
||||
// All 20 slot constants live in `crate::cuda_pipeline::sp11_isv_slots`
|
||||
// (re-exported above). They occupy ISV[340..360) and are FoldReset (sentinel 0;
|
||||
// Pearl A bootstrap on first observation). SP5_SLOT_END bumps 340 → 360 so the
|
||||
// linear-span window covers the new block; the wiener_state_buf lockstep is
|
||||
// handled by the existing bulk memset of `wiener_state_buf` covered by the
|
||||
// `sp4_wiener_state` registry entry's dispatch arm.
|
||||
|
||||
pub const SP5_SLOT_END: usize = 360;
|
||||
|
||||
/// Wiener-buffer producer-count constant. Sizes `wiener_state_buf` via
|
||||
/// `(SP4_PRODUCER_COUNT + SP5_PRODUCER_COUNT) * SP4_WIENER_FLOATS_PER_SLOT`.
|
||||
@@ -359,9 +380,20 @@ pub const SP5_SLOT_END: usize = 340;
|
||||
/// the existing `intent_eval_divergence_compute_kernel` (single producer,
|
||||
/// two outputs); FoldReset (sentinel 0; Pearl A bootstrap on first
|
||||
/// observation).
|
||||
pub const SP5_PRODUCER_COUNT: usize = 166;
|
||||
// linear span = SP5_SLOT_END - SP5_SLOT_BASE = 340 - 174 = 166 wiener triples
|
||||
// unique-slot count = 164 (52 per-branch + 24 Adam + 20 IQN τ + 4 trail
|
||||
///
|
||||
/// SP11 (2026-05-04, Fix 39): allocates 20 new SP11 ISV slots at
|
||||
/// ISV[340..360) — 6 reward component weights + 4 controller scalar
|
||||
/// outputs + 10 canaries (val-sharpe Δ/var, 6 mag-ratios, saboteur
|
||||
/// engagement, PnL EMA). Unique-slot count grows 164 → 184; the linear
|
||||
/// span (and `SP5_PRODUCER_COUNT`) grows 166 → 186. Drives the reward
|
||||
/// subsystem controller per `pearl_reward_as_controlled_subsystem.md` +
|
||||
/// `pearl_controller_anchors_isv_driven.md`. All 20 slots are FoldReset
|
||||
/// (sentinel 0; Pearl A bootstrap on first observation). Constants for
|
||||
/// the SP11 block live in `crate::cuda_pipeline::sp11_isv_slots` and are
|
||||
/// re-exported above.
|
||||
pub const SP5_PRODUCER_COUNT: usize = 186;
|
||||
// linear span = SP5_SLOT_END - SP5_SLOT_BASE = 360 - 174 = 186 wiener triples
|
||||
// unique-slot count = 184 (52 per-branch + 24 Adam + 20 IQN τ + 4 trail
|
||||
// + 4 num_atoms + 6 Kelly + 4 Layer D D1 PnL aggregation
|
||||
// + 4 Layer D D2 health composition
|
||||
// + 3 Layer D D3 training metrics EMA
|
||||
@@ -370,7 +402,8 @@ pub const SP5_PRODUCER_COUNT: usize = 166;
|
||||
// + 1 SP8 train_active_frac canary
|
||||
// + 8 SP8 LB_MAX_BUDGET per-(head,branch)
|
||||
// + 9 SP9 Kelly cold-start warmup floor + targets + eval_dist
|
||||
// + 1 SP10 eval Thompson temperature)
|
||||
// + 1 SP10 eval Thompson temperature
|
||||
// + 20 SP11 reward-subsystem controller)
|
||||
|
||||
// ── Convenience accessors ────────────────────────────────────────────
|
||||
#[inline] pub const fn atom_v_center(b: usize) -> usize { ATOM_V_CENTER_BASE + b }
|
||||
@@ -426,7 +459,14 @@ pub const SP5_LAYOUT_FINGERPRINT_FRAGMENT: &str =
|
||||
KELLY_DIVERGENCE_TARGET=334;KELLY_TEMPORAL_TARGET=335;\
|
||||
EVAL_DIST_Q=336;EVAL_DIST_H=337;EVAL_DIST_F=338;\
|
||||
EVAL_THOMPSON_TEMP=339;\
|
||||
ISV_TOTAL_DIM=340";
|
||||
REWARD_POPART_WEIGHT=340;REWARD_CF_WEIGHT=341;REWARD_TRAIL_WEIGHT=342;\
|
||||
REWARD_MICRO_WEIGHT=343;REWARD_OPP_COST_WEIGHT=344;REWARD_BONUS_WEIGHT=345;\
|
||||
CURIOSITY_PRESSURE=346;SABOTEUR_INTENSITY_MULT=347;\
|
||||
REWARD_WEIGHT_FLOOR=348;CURIOSITY_BOUND=349;\
|
||||
VAL_SHARPE_DELTA_EMA=350;VAL_SHARPE_VAR_EMA=351;\
|
||||
REWARD_COMPONENT_MAG_RATIO_BASE=352;\
|
||||
SABOTEUR_ENGAGEMENT_RATE=358;PNL_REWARD_MAGNITUDE_EMA=359;\
|
||||
ISV_TOTAL_DIM=360";
|
||||
|
||||
#[cfg(test)]
|
||||
mod tests {
|
||||
@@ -534,29 +574,34 @@ mod tests {
|
||||
// SP10 (Fix 38): eval Thompson selector temperature
|
||||
slots.insert(EVAL_THOMPSON_TEMP_INDEX);
|
||||
|
||||
// 1. Exactly 164 unique slots.
|
||||
assert_eq!(slots.len(), 164, "expected 164 unique slots, got {}", slots.len());
|
||||
// SP11 (Fix 39): include the 20 new reward-subsystem controller slots.
|
||||
for s in SP11_SLOT_BASE..SP11_SLOT_END {
|
||||
slots.insert(s);
|
||||
}
|
||||
|
||||
// 1. Exactly 184 unique slots (164 pre-SP11 + 20 SP11).
|
||||
assert_eq!(slots.len(), 184, "expected 184 unique slots, got {}", slots.len());
|
||||
|
||||
// 2. Min slot is SP5_SLOT_BASE = 174.
|
||||
assert_eq!(*slots.iter().min().unwrap(), 174);
|
||||
|
||||
// 3. Max slot is SP5_SLOT_END - 1 = 339.
|
||||
assert_eq!(*slots.iter().max().unwrap(), 339);
|
||||
// 3. Max slot is SP5_SLOT_END - 1 = 359.
|
||||
assert_eq!(*slots.iter().max().unwrap(), 359);
|
||||
|
||||
// 4. Intentional carve-out gap (278, 279) is absent.
|
||||
assert!(!slots.contains(&278), "slot 278 must be absent (carve-out gap)");
|
||||
assert!(!slots.contains(&279), "slot 279 must be absent (carve-out gap)");
|
||||
|
||||
// 5. Set equals {174..278} ∪ {280..340} — no holes (other than the
|
||||
// 5. Set equals {174..278} ∪ {280..360} — no holes (other than the
|
||||
// carve-out gap 278..280), no overlaps. Layer D D1 extended the
|
||||
// upper end 286 → 290; D2 extended it 290 → 294; D3 extends it
|
||||
// 294 → 297; SP7 T1 extended it 297 → 313; SP7 activation-flag
|
||||
// fix extends it 313 → 321; SP8 (Fix 36) extends it 321 → 330;
|
||||
// SP9 (Fix 37) extends it 330 → 339; SP10 (Fix 38) extends it
|
||||
// 339 → 340.
|
||||
let expected: HashSet<usize> = (174..278).chain(280..340).collect();
|
||||
// 339 → 340; SP11 (Fix 39) extends it 340 → 360.
|
||||
let expected: HashSet<usize> = (174..278).chain(280..360).collect();
|
||||
assert_eq!(slots, expected,
|
||||
"slot set does not match expected {{174..278}} ∪ {{280..340}}");
|
||||
"slot set does not match expected {{174..278}} ∪ {{280..360}}");
|
||||
}
|
||||
|
||||
#[test]
|
||||
@@ -583,11 +628,11 @@ mod tests {
|
||||
// is cross-fold-persistent — the contracts must remain disjoint).
|
||||
assert!(PNL_TOTAL_INDEX > LOSS_RATE_SMOOTH_INDEX);
|
||||
// SP5_SLOT_END must reflect the post-SP8 end-of-block.
|
||||
assert_eq!(SP5_SLOT_END, 340);
|
||||
assert_eq!(SP5_SLOT_END, 360);
|
||||
// SP5_PRODUCER_COUNT is the wiener-buffer linear span (slot-range
|
||||
// width including the 2-slot carve-out gap), NOT the unique-slot
|
||||
// count. See SP5_PRODUCER_COUNT docstring for the rationale.
|
||||
assert_eq!(SP5_PRODUCER_COUNT, 166);
|
||||
assert_eq!(SP5_PRODUCER_COUNT, 186);
|
||||
assert_eq!(SP5_PRODUCER_COUNT, SP5_SLOT_END - SP5_SLOT_BASE);
|
||||
}
|
||||
|
||||
@@ -606,7 +651,7 @@ mod tests {
|
||||
// 4-slot block is internally contiguous.
|
||||
assert_eq!(GRAD_NORM_NORM_INDEX - HEALTH_SCORE_INDEX, 3);
|
||||
// SP5_SLOT_END must reflect the post-SP8 end-of-block.
|
||||
assert_eq!(SP5_SLOT_END, 340);
|
||||
assert_eq!(SP5_SLOT_END, 360);
|
||||
// SP5_PRODUCER_COUNT linear-span check matches the new end.
|
||||
assert_eq!(SP5_PRODUCER_COUNT, SP5_SLOT_END - SP5_SLOT_BASE);
|
||||
}
|
||||
@@ -627,12 +672,12 @@ mod tests {
|
||||
assert_eq!(MAX_DD_EMA_INDEX - TRAINING_SHARPE_EMA_INDEX, 1);
|
||||
assert_eq!(LOW_DD_RATIO_INDEX - MAX_DD_EMA_INDEX, 1);
|
||||
// SP5_SLOT_END must reflect the post-SP8 end-of-block.
|
||||
assert_eq!(SP5_SLOT_END, 340);
|
||||
assert_eq!(SP5_SLOT_END, 360);
|
||||
// SP5_PRODUCER_COUNT linear-span check matches the new end. The
|
||||
// wiener buffer must cover the entire linear span — including the
|
||||
// 2-slot carve-out gap (278..280) and the Pearl 6 reserved-but-
|
||||
// unused 6-float block (slots 280..286 don't call apply_pearls).
|
||||
assert_eq!(SP5_PRODUCER_COUNT, 166);
|
||||
assert_eq!(SP5_PRODUCER_COUNT, 186);
|
||||
assert_eq!(SP5_PRODUCER_COUNT, SP5_SLOT_END - SP5_SLOT_BASE);
|
||||
}
|
||||
|
||||
@@ -652,7 +697,7 @@ mod tests {
|
||||
assert_eq!(lb_c51_active(b), LB_C51_ACTIVE_BASE + b);
|
||||
}
|
||||
// SP5_SLOT_END / SP5_PRODUCER_COUNT cover the new range.
|
||||
assert_eq!(SP5_SLOT_END, 340);
|
||||
assert_eq!(SP5_SLOT_END, 360);
|
||||
assert_eq!(SP5_PRODUCER_COUNT, SP5_SLOT_END - SP5_SLOT_BASE);
|
||||
}
|
||||
|
||||
@@ -672,7 +717,7 @@ mod tests {
|
||||
assert_eq!(lb_max_budget_c51(b), LB_MAX_BUDGET_C51_BASE + b);
|
||||
}
|
||||
// Layout: 1 + 4 + 4 = 9 contiguous slots (321..330).
|
||||
assert_eq!(SP5_SLOT_END, 340);
|
||||
assert_eq!(SP5_SLOT_END, 360);
|
||||
assert_eq!(SP5_PRODUCER_COUNT, SP5_SLOT_END - SP5_SLOT_BASE);
|
||||
}
|
||||
|
||||
@@ -705,7 +750,7 @@ mod tests {
|
||||
// Strictly above the SP8 MAX_BUDGET block.
|
||||
assert!(KELLY_WARMUP_FLOOR_INDEX > LB_MAX_BUDGET_C51_BASE + 3);
|
||||
// Block end at ISV[340) post-SP10 (Fix 38).
|
||||
assert_eq!(SP5_SLOT_END, 340);
|
||||
assert_eq!(SP5_SLOT_END, 360);
|
||||
// Linear span matches producer count (no gaps in SP9 block).
|
||||
assert_eq!(SP5_PRODUCER_COUNT, SP5_SLOT_END - SP5_SLOT_BASE);
|
||||
}
|
||||
@@ -719,9 +764,9 @@ mod tests {
|
||||
// Strictly above the SP9 eval_dist block (last entry at 338).
|
||||
assert!(EVAL_THOMPSON_TEMP_INDEX > EVAL_DIST_F_INDEX);
|
||||
// SP5_SLOT_END reflects the post-SP10 end-of-block.
|
||||
assert_eq!(SP5_SLOT_END, 340);
|
||||
assert_eq!(SP5_SLOT_END, 360);
|
||||
// Linear span matches producer count (single new slot at top).
|
||||
assert_eq!(SP5_PRODUCER_COUNT, SP5_SLOT_END - SP5_SLOT_BASE);
|
||||
assert_eq!(SP5_PRODUCER_COUNT, 166);
|
||||
assert_eq!(SP5_PRODUCER_COUNT, 186);
|
||||
}
|
||||
}
|
||||
|
||||
@@ -765,6 +765,121 @@ impl StateResetRegistry {
|
||||
category: ResetCategory::FoldReset,
|
||||
description: "ISV[EVAL_THOMPSON_TEMP_INDEX=339] — SP10 Fix 38 eval Thompson selector temperature. Produced as a side-output of `intent_eval_divergence_compute_kernel` (`temp = clamp(divergence/divergence_target, 0.5, 2.0)`). Consumed in `experience_action_select` direction-branch unconditionally (training AND eval) for the temperature-blended Thompson sample `q_eff[d] = E[Q][d] + temp · (q_sample[d] − E[Q][d])`. MIN_TEMP=0.5 enforces permanent stochasticity per `pearl_blend_formulas_must_have_permanent_floor.md`. Eliminates the val-Flat-collapse pathology (T10 train-multi-seed-khr7c on `8a25b330f`: dir_entropy=0, trade_count=1 in 214k bars) caused by the prior `eval_mode ⇒ argmax(E[Q])` selector branch. FoldReset sentinel 0; Pearl A bootstraps from the first observation; the consumer's defensive clamp keeps the selector at MIN_TEMP between fold-reset and first producer launch.",
|
||||
},
|
||||
// ── SP11 (Fix 39, 2026-05-04): reward-subsystem controller ─────
|
||||
//
|
||||
// Per `pearl_reward_as_controlled_subsystem.md` +
|
||||
// `pearl_controller_anchors_isv_driven.md`: 20 new ISV slots at
|
||||
// [340..360) populated by the SP11 controller chain — 6 component
|
||||
// weights + 4 controller scalar outputs + 10 canaries (val-sharpe
|
||||
// delta/var, 6 mag-ratios, saboteur engagement, PnL EMA). All
|
||||
// slots are FoldReset sentinel 0; Pearl A bootstraps from the
|
||||
// first producer launch on the new fold (per
|
||||
// `pearl_first_observation_bootstrap.md`). Producers and
|
||||
// consumers are introduced in subsequent A1/A2/B tasks; A0 is
|
||||
// additive — no consumer reads any of these slots yet, so the
|
||||
// trainer continues to behave identically.
|
||||
//
|
||||
// The 20 entries below mirror the registry layout established by
|
||||
// the SP9/SP10 entries above (one entry per uniquely-zeroed slot
|
||||
// / per-related-block). The 6 mag-ratio canaries occupy a
|
||||
// contiguous 6-slot block at [352..358) and share a single
|
||||
// entry — the dispatch arm zeroes the whole range with one loop,
|
||||
// matching the `sp5_pnl_aggregation`/`sp5_health_composition`
|
||||
// pattern.
|
||||
//
|
||||
// Wiener-state companion: the bulk memset of `wiener_state_buf`
|
||||
// covered by the `sp4_wiener_state` registry entry's dispatch
|
||||
// arm includes the new SP11 triples (the buffer linear span
|
||||
// grew from `(SP4_PRODUCER_COUNT + 166) × 3` to
|
||||
// `(SP4_PRODUCER_COUNT + 186) × 3` = +60 floats). Both halves
|
||||
// of the Pearl A sentinel contract reset together per
|
||||
// `feedback_no_partial_refactor.md`.
|
||||
//
|
||||
// The novelty-hash device buffer (`novelty_hash_buf`) is NOT an
|
||||
// ISV slot but a separate device-side hash table sized for ~1M
|
||||
// 16-bit codes. Its registry entry + dispatch arm
|
||||
// (`fill_zero_async` at fold boundary) lands in Task A2
|
||||
// alongside the buffer field's introduction on `GpuDqnTrainer`.
|
||||
// Adding the entry here would be an orphan against a
|
||||
// non-existent buffer field, violating
|
||||
// `feedback_wire_everything_up.md`. A2 owns the contract end-
|
||||
// to-end: buffer field, registry entry, dispatch arm, and
|
||||
// launcher invocation in the same commit.
|
||||
RegistryEntry {
|
||||
name: "sp11_reward_popart_weight",
|
||||
category: ResetCategory::FoldReset,
|
||||
description: "ISV[REWARD_POPART_WEIGHT_INDEX=340] — SP11 Fix 39 PopArt reward component weight. Produced by `reward_subsystem_controller_kernel` from improvement_z + per-component mag-ratio canaries (winner/diversifier blend). Consumed in unified-reward composition (Layer B). FoldReset sentinel 0; Pearl A bootstraps from the first controller emit (per `pearl_first_observation_bootstrap.md`). Spec: docs/superpowers/specs/2026-05-04-sp11-reward-as-controlled-subsystem.md §3.3.",
|
||||
},
|
||||
RegistryEntry {
|
||||
name: "sp11_reward_cf_weight",
|
||||
category: ResetCategory::FoldReset,
|
||||
description: "ISV[REWARD_CF_WEIGHT_INDEX=341] — SP11 Fix 39 counterfactual (CF) reward component weight. Produced by `reward_subsystem_controller_kernel`. Consumed in `mse_loss_kernel` and `c51_loss_kernel` (Layer B replaces the hardcoded `cf_weight=0.3f`). FoldReset sentinel 0; Pearl A bootstraps from the first controller emit (per `pearl_first_observation_bootstrap.md`). Spec §3.3.",
|
||||
},
|
||||
RegistryEntry {
|
||||
name: "sp11_reward_trail_weight",
|
||||
category: ResetCategory::FoldReset,
|
||||
description: "ISV[REWARD_TRAIL_WEIGHT_INDEX=342] — SP11 Fix 39 trail-stop reward component weight. Produced by `reward_subsystem_controller_kernel`. Consumed in trail-shaping reward path (Layer B). FoldReset sentinel 0; Pearl A bootstraps from the first controller emit. Spec §3.3.",
|
||||
},
|
||||
RegistryEntry {
|
||||
name: "sp11_reward_micro_weight",
|
||||
category: ResetCategory::FoldReset,
|
||||
description: "ISV[REWARD_MICRO_WEIGHT_INDEX=343] — SP11 Fix 39 microstructure reward component weight. Produced by `reward_subsystem_controller_kernel`. Consumed in microstructure-shaping reward path (Layer B). FoldReset sentinel 0; Pearl A bootstraps from the first controller emit. Spec §3.3.",
|
||||
},
|
||||
RegistryEntry {
|
||||
name: "sp11_reward_opp_cost_weight",
|
||||
category: ResetCategory::FoldReset,
|
||||
description: "ISV[REWARD_OPP_COST_WEIGHT_INDEX=344] — SP11 Fix 39 opportunity-cost reward component weight. Produced by `reward_subsystem_controller_kernel`. Consumed in opportunity-cost shaping reward path (Layer B). FoldReset sentinel 0; Pearl A bootstraps from the first controller emit. Spec §3.3.",
|
||||
},
|
||||
RegistryEntry {
|
||||
name: "sp11_reward_bonus_weight",
|
||||
category: ResetCategory::FoldReset,
|
||||
description: "ISV[REWARD_BONUS_WEIGHT_INDEX=345] — SP11 Fix 39 exploration-bonus reward component weight. Produced by `reward_subsystem_controller_kernel`. Consumed in bonus-shaping reward path (Layer B). FoldReset sentinel 0; Pearl A bootstraps from the first controller emit. Spec §3.3.",
|
||||
},
|
||||
RegistryEntry {
|
||||
name: "sp11_curiosity_pressure",
|
||||
category: ResetCategory::FoldReset,
|
||||
description: "ISV[CURIOSITY_PRESSURE_INDEX=346] — SP11 Fix 39 curiosity-pressure scalar. Produced by `reward_subsystem_controller_kernel` from val-sharpe trend + mag-ratio dispersion. Consumed at REPLAY-time (not collect-time) as `r_with_curiosity = r_base + curiosity_pressure · novelty_signal(state, action)`. Curiosity-bound capped at `SP11_CURIOSITY_BOUND_FRACTION=0.3 × pnl_reward_magnitude_ema`, with a permanent-fraction floor of `SP11_CURIOSITY_PERMANENT_FRACTION=0.2 × bound` per `pearl_blend_formulas_must_have_permanent_floor.md`. FoldReset sentinel 0; Pearl A bootstraps. Spec §3.3, §3.4.2.",
|
||||
},
|
||||
RegistryEntry {
|
||||
name: "sp11_saboteur_intensity_mult",
|
||||
category: ResetCategory::FoldReset,
|
||||
description: "ISV[SABOTEUR_INTENSITY_MULT_INDEX=347] — SP11 Fix 39 saboteur curriculum intensity multiplier. Produced by `reward_subsystem_controller_kernel` from saboteur_engagement_rate canary. Consumed in `gpu_experience_collector` (Layer B multiplies `saboteur_perturbation_scale` by this slot). Clamped to `[SP11_SABOTEUR_MIN=0.5, SP11_SABOTEUR_MAX=2.0]` (Invariant-1 rate-limiter anchor per `feedback_isv_for_adaptive_bounds.md` §3.4.2). FoldReset sentinel 0; Pearl A bootstraps. Spec §3.3, §3.4.2.",
|
||||
},
|
||||
RegistryEntry {
|
||||
name: "sp11_reward_weight_floor",
|
||||
category: ResetCategory::FoldReset,
|
||||
description: "ISV[REWARD_WEIGHT_FLOOR_INDEX=348] — SP11 Fix 39 controller-output minimum-weight floor (`SP11_WEIGHT_FLOOR_FRACTION=0.5 × hard_floor`). Produced by `reward_subsystem_controller_kernel`. Consumed by the controller itself (re-read on next step) to enforce that no reward component drops below `SP11_WEIGHT_HARD_FLOOR=0.01` after blending — Invariant-1 rate-limiter floor preventing winner-take-all collapse. FoldReset sentinel 0; Pearl A bootstraps. Spec §3.4.2.",
|
||||
},
|
||||
RegistryEntry {
|
||||
name: "sp11_curiosity_bound",
|
||||
category: ResetCategory::FoldReset,
|
||||
description: "ISV[CURIOSITY_BOUND_INDEX=349] — SP11 Fix 39 curiosity-pressure structural cap (`SP11_CURIOSITY_BOUND_FRACTION=0.3 × pnl_reward_magnitude_ema`). Produced by `reward_subsystem_controller_kernel`. Consumed at the curiosity-pressure clamp — keeps curiosity scaled to the run's PnL magnitude rather than a tuned constant per `feedback_isv_for_adaptive_bounds.md`. FoldReset sentinel 0; Pearl A bootstraps. Spec §3.4.2.",
|
||||
},
|
||||
RegistryEntry {
|
||||
name: "sp11_val_sharpe_delta_ema",
|
||||
category: ResetCategory::FoldReset,
|
||||
description: "ISV[VAL_SHARPE_DELTA_EMA_INDEX=350] — SP11 Fix 39 val-sharpe trend canary. Produced by `val_sharpe_delta_compute_kernel` (two-pass delta_ema + delta_var_ema, A1). Smoothed by Pearls A+D via chained `apply_pearls_ad_kernel` per `feedback_no_cpu_compute_strict.md`. Consumed by the SP11 controller as the improvement_z numerator. FoldReset sentinel 0; Pearl A bootstraps from the first val-sharpe sample of the new fold. Spec §3.3.1.",
|
||||
},
|
||||
RegistryEntry {
|
||||
name: "sp11_val_sharpe_var_ema",
|
||||
category: ResetCategory::FoldReset,
|
||||
description: "ISV[VAL_SHARPE_VAR_EMA_INDEX=351] — SP11 Fix 39 val-sharpe variance canary (squared-deviation EMA against the running delta_ema). Produced alongside `sp11_val_sharpe_delta_ema` by `val_sharpe_delta_compute_kernel` (A1). Smoothed by Pearls A+D. Consumed by the SP11 controller as the improvement_z denominator (`improvement_z = delta_ema / sqrt(var_ema + SP11_EPS_DIV)`). FoldReset sentinel 0; Pearl A bootstraps. Spec §3.3.1.",
|
||||
},
|
||||
RegistryEntry {
|
||||
name: "sp11_reward_component_mag_ratios",
|
||||
category: ResetCategory::FoldReset,
|
||||
description: "ISV[REWARD_COMPONENT_MAG_RATIO_BASE=352..358) — SP11 Fix 39 per-component reward-magnitude ratios (6 contiguous slots: popart, cf, trail, micro, opp_cost, bonus). Produced by `reward_component_mag_ratio_compute_kernel` (A1) reading existing per-component reward-magnitude EMAs (`REWARD_POPART_EMA_INDEX..+6`) and normalising to a unit simplex. Smoothed by Pearls A+D via chained `apply_pearls_ad_kernel`. Consumed by the SP11 controller for the winner_weight / diversifier_weight blend. The single registry entry covers the whole 6-slot block — the dispatch arm zeroes all 6 with one loop, matching the `sp5_pnl_aggregation` block-reset pattern. Per `feedback_trust_code_not_docs.md`: spec §3.3.1 said `grad EMA` but the SP4 grad-balancer is per-branch (4 slots), not per-reward-component — code wins, terminology corrected to `mag-ratio` (semantics for the controller are equivalent). FoldReset sentinel 0; Pearl A bootstraps. Spec §3.3.1.",
|
||||
},
|
||||
RegistryEntry {
|
||||
name: "sp11_saboteur_engagement_rate",
|
||||
category: ResetCategory::FoldReset,
|
||||
description: "ISV[SABOTEUR_ENGAGEMENT_RATE_INDEX=358] — SP11 Fix 39 saboteur engagement-rate canary (per-bar fraction of bars where the saboteur curriculum was actively perturbing experience tuples). Produced by `saboteur_engagement_compute_kernel` (A1) — block tree-reduce over the per-bar engagement flag, then EMA. Smoothed by Pearls A+D. Consumed by the SP11 controller to drive `saboteur_intensity_mult` toward the engagement floor (`SP11_ENGAGEMENT_FLOOR=0.1`). FoldReset sentinel 0; Pearl A bootstraps. Spec §3.3.1.",
|
||||
},
|
||||
RegistryEntry {
|
||||
name: "sp11_pnl_reward_magnitude_ema",
|
||||
category: ResetCategory::FoldReset,
|
||||
description: "ISV[PNL_REWARD_MAGNITUDE_EMA_INDEX=359] — SP11 Fix 39 PnL-reward magnitude EMA (running `|reward_total|` smoothed with Wiener-α). Produced by the unified-reward kernel chain (existing producer; SP11 wires the EMA output to this slot in A1 alongside the val-sharpe and engagement canaries). Consumed by the SP11 controller as the curiosity-bound scale factor (`curiosity_bound = SP11_CURIOSITY_BOUND_FRACTION=0.3 × this slot`) per `feedback_isv_for_adaptive_bounds.md` §3.4.2. FoldReset sentinel 0; Pearl A bootstraps. Spec §3.3.1, §3.4.2.",
|
||||
},
|
||||
// SP5 Task A1: Wiener-state companion reset. The wiener_state_buf
|
||||
// now covers SP4 (71 producers × 3 = 213 floats) + SP5 (110 × 3 = 330
|
||||
// floats) = 543 floats total. The SP5 triples start at offset 213.
|
||||
|
||||
@@ -6968,6 +6968,122 @@ impl DQNTrainer {
|
||||
fused.trainer().write_isv_signal_at(EVAL_THOMPSON_TEMP_INDEX, 0.0);
|
||||
}
|
||||
}
|
||||
// ── SP11 (Fix 39, 2026-05-04): reward-subsystem controller ─────
|
||||
//
|
||||
// 15 registry entries covering 20 slots [340..360). All are
|
||||
// FoldReset sentinel 0 so Pearl A's first-observation
|
||||
// replacement fires on the new fold's first producer launch
|
||||
// (per `pearl_first_observation_bootstrap.md`). Wiener-state
|
||||
// companion is reset by the existing bulk memset of
|
||||
// `wiener_state_buf` covered by the `sp4_wiener_state`
|
||||
// registry entry's dispatch arm — `SP5_PRODUCER_COUNT` linear
|
||||
// span grew 166 → 186, the buffer expanded in lockstep.
|
||||
//
|
||||
// No consumer reads any of these slots in A0 — the trainer
|
||||
// continues to behave identically. Layer A1/A2 introduce the
|
||||
// producer kernels and Layer B atomically migrates consumers.
|
||||
"sp11_reward_popart_weight" => {
|
||||
if let Some(ref fused) = self.fused_ctx {
|
||||
use crate::cuda_pipeline::sp5_isv_slots::REWARD_POPART_WEIGHT_INDEX;
|
||||
fused.trainer().write_isv_signal_at(REWARD_POPART_WEIGHT_INDEX, 0.0);
|
||||
}
|
||||
}
|
||||
"sp11_reward_cf_weight" => {
|
||||
if let Some(ref fused) = self.fused_ctx {
|
||||
use crate::cuda_pipeline::sp5_isv_slots::REWARD_CF_WEIGHT_INDEX;
|
||||
fused.trainer().write_isv_signal_at(REWARD_CF_WEIGHT_INDEX, 0.0);
|
||||
}
|
||||
}
|
||||
"sp11_reward_trail_weight" => {
|
||||
if let Some(ref fused) = self.fused_ctx {
|
||||
use crate::cuda_pipeline::sp5_isv_slots::REWARD_TRAIL_WEIGHT_INDEX;
|
||||
fused.trainer().write_isv_signal_at(REWARD_TRAIL_WEIGHT_INDEX, 0.0);
|
||||
}
|
||||
}
|
||||
"sp11_reward_micro_weight" => {
|
||||
if let Some(ref fused) = self.fused_ctx {
|
||||
use crate::cuda_pipeline::sp5_isv_slots::REWARD_MICRO_WEIGHT_INDEX;
|
||||
fused.trainer().write_isv_signal_at(REWARD_MICRO_WEIGHT_INDEX, 0.0);
|
||||
}
|
||||
}
|
||||
"sp11_reward_opp_cost_weight" => {
|
||||
if let Some(ref fused) = self.fused_ctx {
|
||||
use crate::cuda_pipeline::sp5_isv_slots::REWARD_OPP_COST_WEIGHT_INDEX;
|
||||
fused.trainer().write_isv_signal_at(REWARD_OPP_COST_WEIGHT_INDEX, 0.0);
|
||||
}
|
||||
}
|
||||
"sp11_reward_bonus_weight" => {
|
||||
if let Some(ref fused) = self.fused_ctx {
|
||||
use crate::cuda_pipeline::sp5_isv_slots::REWARD_BONUS_WEIGHT_INDEX;
|
||||
fused.trainer().write_isv_signal_at(REWARD_BONUS_WEIGHT_INDEX, 0.0);
|
||||
}
|
||||
}
|
||||
"sp11_curiosity_pressure" => {
|
||||
if let Some(ref fused) = self.fused_ctx {
|
||||
use crate::cuda_pipeline::sp5_isv_slots::CURIOSITY_PRESSURE_INDEX;
|
||||
fused.trainer().write_isv_signal_at(CURIOSITY_PRESSURE_INDEX, 0.0);
|
||||
}
|
||||
}
|
||||
"sp11_saboteur_intensity_mult" => {
|
||||
if let Some(ref fused) = self.fused_ctx {
|
||||
use crate::cuda_pipeline::sp5_isv_slots::SABOTEUR_INTENSITY_MULT_INDEX;
|
||||
fused.trainer().write_isv_signal_at(SABOTEUR_INTENSITY_MULT_INDEX, 0.0);
|
||||
}
|
||||
}
|
||||
"sp11_reward_weight_floor" => {
|
||||
if let Some(ref fused) = self.fused_ctx {
|
||||
use crate::cuda_pipeline::sp5_isv_slots::REWARD_WEIGHT_FLOOR_INDEX;
|
||||
fused.trainer().write_isv_signal_at(REWARD_WEIGHT_FLOOR_INDEX, 0.0);
|
||||
}
|
||||
}
|
||||
"sp11_curiosity_bound" => {
|
||||
if let Some(ref fused) = self.fused_ctx {
|
||||
use crate::cuda_pipeline::sp5_isv_slots::CURIOSITY_BOUND_INDEX;
|
||||
fused.trainer().write_isv_signal_at(CURIOSITY_BOUND_INDEX, 0.0);
|
||||
}
|
||||
}
|
||||
"sp11_val_sharpe_delta_ema" => {
|
||||
if let Some(ref fused) = self.fused_ctx {
|
||||
use crate::cuda_pipeline::sp5_isv_slots::VAL_SHARPE_DELTA_EMA_INDEX;
|
||||
fused.trainer().write_isv_signal_at(VAL_SHARPE_DELTA_EMA_INDEX, 0.0);
|
||||
}
|
||||
}
|
||||
"sp11_val_sharpe_var_ema" => {
|
||||
if let Some(ref fused) = self.fused_ctx {
|
||||
use crate::cuda_pipeline::sp5_isv_slots::VAL_SHARPE_VAR_EMA_INDEX;
|
||||
fused.trainer().write_isv_signal_at(VAL_SHARPE_VAR_EMA_INDEX, 0.0);
|
||||
}
|
||||
}
|
||||
"sp11_reward_component_mag_ratios" => {
|
||||
// Block-reset the 6 contiguous mag-ratio canaries at
|
||||
// ISV[REWARD_COMPONENT_MAG_RATIO_BASE=352..358). Matches
|
||||
// the `sp5_pnl_aggregation`/`sp5_health_composition`
|
||||
// block-reset pattern (one registry entry, one dispatch
|
||||
// arm, multi-slot zero-write).
|
||||
if let Some(ref fused) = self.fused_ctx {
|
||||
use crate::cuda_pipeline::sp5_isv_slots::{
|
||||
REWARD_COMPONENT_MAG_RATIO_BASE, REWARD_COMPONENT_MAG_RATIO_COUNT,
|
||||
};
|
||||
for i in 0..REWARD_COMPONENT_MAG_RATIO_COUNT {
|
||||
fused.trainer().write_isv_signal_at(
|
||||
REWARD_COMPONENT_MAG_RATIO_BASE + i,
|
||||
0.0,
|
||||
);
|
||||
}
|
||||
}
|
||||
}
|
||||
"sp11_saboteur_engagement_rate" => {
|
||||
if let Some(ref fused) = self.fused_ctx {
|
||||
use crate::cuda_pipeline::sp5_isv_slots::SABOTEUR_ENGAGEMENT_RATE_INDEX;
|
||||
fused.trainer().write_isv_signal_at(SABOTEUR_ENGAGEMENT_RATE_INDEX, 0.0);
|
||||
}
|
||||
}
|
||||
"sp11_pnl_reward_magnitude_ema" => {
|
||||
if let Some(ref fused) = self.fused_ctx {
|
||||
use crate::cuda_pipeline::sp5_isv_slots::PNL_REWARD_MAGNITUDE_EMA_INDEX;
|
||||
fused.trainer().write_isv_signal_at(PNL_REWARD_MAGNITUDE_EMA_INDEX, 0.0);
|
||||
}
|
||||
}
|
||||
"sp5_pnl_aggregation" => {
|
||||
// Layer D Task D1: zero ISV[286..290) — total/mean/var/max_dd —
|
||||
// so the new fold's first `launch_sp5_pnl_aggregation` triggers
|
||||
|
||||
@@ -183,3 +183,50 @@ gradient contribution = `iqn_budget` (same as SP5 Layer B). The averaged
|
||||
|
||||
**Files changed:** `crates/ml/src/cuda_pipeline/gpu_iqn_head.rs`,
|
||||
`crates/ml/src/trainers/dqn/fused_training.rs`.
|
||||
|
||||
## SP11: Reward as controlled subsystem (Task A0 reservation)
|
||||
|
||||
SP11 Task A0 (Fix 39, 2026-05-04) extends `ISV_TOTAL_DIM` from 340 → 360 by
|
||||
allocating 20 contiguous slots at `[340..360)` for the reward-subsystem
|
||||
controller. Constants live in `crates/ml/src/cuda_pipeline/sp11_isv_slots.rs`
|
||||
and are re-exported via `cuda_pipeline::sp5_isv_slots::*`. **A0 is reservation-
|
||||
only — no producer kernels or consumers are wired yet. All 20 slots remain
|
||||
zero-initialized at construction and behave as no-ops until A1/A2/Layer-B
|
||||
land producers and atomic consumer migration.**
|
||||
|
||||
| Range | Name constant | Family | Purpose |
|
||||
|---|---|---|---|
|
||||
| [340] | `REWARD_POPART_WEIGHT_INDEX` | scalar | Component weight: PopArt reward |
|
||||
| [341] | `REWARD_CF_WEIGHT_INDEX` | scalar | Component weight: counterfactual (replaces hardcoded `cf_weight=0.3f`) |
|
||||
| [342] | `REWARD_TRAIL_WEIGHT_INDEX` | scalar | Component weight: trail-stop |
|
||||
| [343] | `REWARD_MICRO_WEIGHT_INDEX` | scalar | Component weight: microstructure |
|
||||
| [344] | `REWARD_OPP_COST_WEIGHT_INDEX` | scalar | Component weight: opportunity cost |
|
||||
| [345] | `REWARD_BONUS_WEIGHT_INDEX` | scalar | Component weight: exploration bonus |
|
||||
| [346] | `CURIOSITY_PRESSURE_INDEX` | scalar | Controller output: curiosity scale (replay-time bonus) |
|
||||
| [347] | `SABOTEUR_INTENSITY_MULT_INDEX` | scalar | Controller output: saboteur intensity multiplier ∈ [0.5, 2.0] |
|
||||
| [348] | `REWARD_WEIGHT_FLOOR_INDEX` | scalar | Controller output: minimum-weight floor |
|
||||
| [349] | `CURIOSITY_BOUND_INDEX` | scalar | Controller output: curiosity hard cap (= 0.3 × pnl_magnitude_ema) |
|
||||
| [350] | `VAL_SHARPE_DELTA_EMA_INDEX` | scalar canary | val-sharpe trend EMA (improvement_z numerator) |
|
||||
| [351] | `VAL_SHARPE_VAR_EMA_INDEX` | scalar canary | val-sharpe variance EMA (improvement_z denominator) |
|
||||
| [352..358) | `REWARD_COMPONENT_MAG_RATIO_BASE` (+6) | per-component canary | Per-component reward-magnitude ratio (winner/diversifier blend input) |
|
||||
| [358] | `SABOTEUR_ENGAGEMENT_RATE_INDEX` | scalar canary | Per-bar saboteur engagement rate EMA |
|
||||
| [359] | `PNL_REWARD_MAGNITUDE_EMA_INDEX` | scalar canary | `|reward_total|` EMA (curiosity-bound scale factor) |
|
||||
|
||||
Invariant-1 fraction constants (rate-limiter anchors per spec §3.4.2):
|
||||
`SP11_EPS_DIV=1e-6`, `SP11_WEIGHT_HARD_FLOOR=0.01`,
|
||||
`SP11_SABOTEUR_MIN=0.5`, `SP11_SABOTEUR_MAX=2.0`,
|
||||
`SP11_CURIOSITY_PERMANENT_FRACTION=0.2`, `SP11_CURIOSITY_BOUND_FRACTION=0.3`,
|
||||
`SP11_WEIGHT_FLOOR_FRACTION=0.5`, `SP11_ENGAGEMENT_FLOOR=0.1`. Plus
|
||||
`SP11_PROJ_SEED_SALT=0x_5511_0001` for the A2 novelty-SimHash projection-init
|
||||
launcher's seed derivation.
|
||||
|
||||
All 20 slots are FoldReset (sentinel 0; Pearl A bootstraps from first
|
||||
producer launch on each fold per `pearl_first_observation_bootstrap.md`).
|
||||
Producers and consumers wire in subsequent A1/A2/Layer-B tasks. The
|
||||
novelty-hash device buffer (separate from the ISV bus, sized for ~1M 16-bit
|
||||
codes) lands in A2 alongside its registry entry and dispatch arm — adding
|
||||
the entry in A0 would orphan it against a non-existent buffer field, so
|
||||
A2 owns the contract end-to-end.
|
||||
|
||||
**Spec:** `docs/superpowers/specs/2026-05-04-sp11-reward-as-controlled-subsystem.md`
|
||||
**Plan:** `docs/superpowers/plans/2026-05-04-sp11-reward-as-controlled-subsystem.md`
|
||||
|
||||
Reference in New Issue
Block a user