fix(sp9): targets as Invariant-1 anchors + divergence sentinel handling
Two targeted fixes to the SP9 Kelly cold-start warmup floor (Fix 37,
commit `48a8b9ee7`) for quirks observed in smoke-test-wrwkz on a
5-epoch L40S smoke. Atomic per `feedback_no_partial_refactor`.
**Quirk 1 — EMA target saturation:** ep1 HEALTH_DIAG showed
`stat_count_tgt=419 div_tgt=105002 temp_tgt=1.0 conf [stat=1.000
bhv=0.333 tmp=1.000] floor=0.0`. Pearl A sentinel-bootstrap on the
3 EMA target updaters (`kelly_*_target_ema_kernel.cu`) caused the
first observation to replace the sentinel directly, so
`target = first_obs`, `current/target = 1.0`, `confidence = 1.0`,
`floor = base × (1 − 1.0) = 0` — defeating the cold-start mechanism.
The principled fix per `feedback_isv_for_adaptive_bounds.md`:
"sufficient" is defined in absolute terms via Invariant-1 numerical
anchors. The 3 target ISV slots become constructor-written constants
(written ONCE in trainer constructor, identical pattern to existing
`config.cql_alpha → ISV[CQL_ALPHA_INDEX=48]` from Plan 1 Task 12).
- ISV[KELLY_SAMPLE_COUNT_TARGET_INDEX=333] = 100.0 (trades)
- ISV[KELLY_DIVERGENCE_TARGET_INDEX=334] = 2.0 (ratio)
- ISV[KELLY_TEMPORAL_TARGET_INDEX=335] = 5.0 (epochs)
**Quirk 2 — divergence ratio explosion:** ep1 showed
`divergence=70005`. Algebraically `intent_f / max(eval_f, 1e-6) =
0.07 / 1e-6 = 70 000` because `ISV[EVAL_DIST_F_INDEX=338]` was still
at Pearl A sentinel 0 before the first val window populated it.
Algebraically the warmup-floor kernel still derived
`behavioral_conf = 0` (correct semantic), but the synthetic 70 000
in HEALTH_DIAG masked the real signal flow.
Fix: explicit sentinel detection in
`intent_eval_divergence_compute_kernel.cu`:
if (eval_f < SENTINEL_THRESHOLD=1e-5)
divergence = SENTINEL_DIVERGENCE=1e6 // forces behavioral_conf=0
else
divergence = intent_f / eval_f
Same `behavioral_conf = 0` outcome but with explicit provenance —
HEALTH_DIAG now reads `divergence=1e6` until eval_f matures.
**Atomic structure:**
- DELETE 3 EMA target updater kernels (`.cu` files)
- DELETE 3 entries from `crates/ml/build.rs::kernels_with_common`
- DELETE 3 cubin static byte arrays + 3 struct fields + 3 cubin
loaders + 3 fields in trainer construction tuple in
`gpu_dqn_trainer.rs`
- DELETE 3 launches in `launch_sp9_kelly_warmup_floor` (chain
shrinks 5 → 2: q_var_mag_ema + main warmup-floor)
- DELETE 3 scratch slots; SP5_SCRATCH_TOTAL 268 → 265;
SCRATCH_SP9_EVAL_DIST_BASE slides 265 → 262
- ADD constructor-write of 3 Invariant-1 anchors (100.0, 2.0, 5.0)
immediately before layout-fingerprint write
- UPDATE 3 dispatch arms in `reset_named_state` to rewrite the
Invariant-1 anchors at fold boundary (NOT sentinel 0) — these
are constants, not stateful EMAs
- UPDATE 3 registry descriptions to reflect Invariant-1 anchor
semantic + smoke-test-wrwkz evidence
- UPDATE `intent_eval_divergence_compute_kernel.cu` with
sentinel-detect branch
- ISV slot layout UNCHANGED (3 target slots @ ISV[333..336)
remain); Wiener-buffer linear span (`SP5_PRODUCER_COUNT=165`)
UNCHANGED — 3 wiener triples for deleted producers become
reserved-unused like Pearl 6's [525..543) carve-out
- audit doc Fix 37.1 entry with full provenance + smoke evidence
**Verification:**
- `SQLX_OFFLINE=true cargo check -p ml` — clean (only pre-existing
18 warnings; no new errors or warnings).
- `SQLX_OFFLINE=true cargo test -p ml --lib state_reset` — 4/4 pass
including `every_fold_and_soft_reset_entry_has_dispatch_arm`.
- `SQLX_OFFLINE=true cargo test -p ml --lib sp5_isv_slots` — 9/9
pass including `sp9_slots_contiguous_above_sp8_block`.
**Expected smoke signature post-fix:**
- floor: starts ≈ base_floor × 1.0, decays as confidence accumulates
- divergence: 1e6 until first val window, small (≤ 5) afterward
- conf: stat=count/100, bhv=0 until eval_dist matures, tmp=ep/5
- combined_conf reaches 1.0 only when at least one axis genuinely
matures (typically temporal at epoch 5 in fold 0)
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This commit is contained in:
jgrusewski
@@ -525,25 +525,28 @@ fn main() {
|
||||
// `pearl_controller_anchors_isv_driven.md`.
|
||||
"loss_balance_max_budget_compute_kernel.cu",
|
||||
// SP9 Fix 37 (2026-05-03): Kelly cold-start warmup floor — fully
|
||||
// ISV-driven. Six new producer kernels per
|
||||
// ISV-driven. Four producer kernels per
|
||||
// `pearl_cold_start_exit_signal_or.md` and
|
||||
// `pearl_controller_anchors_isv_driven.md`:
|
||||
// 1. eval_intent_dist_compute — replaces DtoH read_eval_intent_magnitude_distribution
|
||||
// 2. intent_eval_divergence_compute — behavioral-axis numerator
|
||||
// 3. q_var_mag_ema_compute — base_floor self-relative baseline
|
||||
// 4. kelly_sample_count_target_ema_compute — statistical-axis denominator
|
||||
// 5. kelly_divergence_target_ema_compute — behavioral-axis denominator
|
||||
// 6. kelly_temporal_target_ema_compute — temporal-axis denominator
|
||||
// 7. kelly_warmup_floor_compute — main producer (combines all axes)
|
||||
// All 7 are single-block kernels writing to producer_step_scratch_buf,
|
||||
// 4. kelly_warmup_floor_compute — main producer (combines all axes)
|
||||
// All 4 are single-block kernels writing to producer_step_scratch_buf,
|
||||
// chained through apply_pearls_ad_kernel for Pearl A sentinel-bootstrap
|
||||
// + Pearl D Wiener-α steady-state smoothing.
|
||||
// + Pearl D Wiener-α steady-state smoothing. Per Fix 37.1
|
||||
// (2026-05-03): the 3 EMA target updaters that previously sat between
|
||||
// q_var_mag_ema and kelly_warmup_floor are deleted — the targets at
|
||||
// ISV[333..336) are Invariant-1 numerical anchors (sample-sufficiency
|
||||
// count, healthy divergence ratio, fold-temporal threshold) written
|
||||
// ONCE at constructor time per `feedback_isv_for_adaptive_bounds.md`,
|
||||
// not EMA-tracked. Pearl A sentinel-bootstrap immediately saturated
|
||||
// the EMAs to the first observation, which made conf=1.0 in epoch 1
|
||||
// (`floor=0`) — defeating the cold-start mechanism. Constructor-write
|
||||
// gives a fixed denominator so the ratio yields meaningful confidence.
|
||||
"eval_intent_dist_compute_kernel.cu",
|
||||
"intent_eval_divergence_compute_kernel.cu",
|
||||
"q_var_mag_ema_compute_kernel.cu",
|
||||
"kelly_sample_count_target_ema_kernel.cu",
|
||||
"kelly_divergence_target_ema_kernel.cu",
|
||||
"kelly_temporal_target_ema_kernel.cu",
|
||||
"kelly_warmup_floor_compute_kernel.cu",
|
||||
// SP7 (2026-05-03): GPU dispatch for cql/c51 budget consumer.
|
||||
// Eliminates capture-time host-branch freeze in
|
||||
|
||||
@@ -510,27 +510,6 @@ static SP9_INTENT_EVAL_DIVERGENCE_COMPUTE_CUBIN: &[u8] =
|
||||
static SP9_Q_VAR_MAG_EMA_COMPUTE_CUBIN: &[u8] =
|
||||
include_bytes!(concat!(env!("OUT_DIR"), "/q_var_mag_ema_compute_kernel.cubin"));
|
||||
|
||||
/// SP9 (Fix 37, 2026-05-03): EMA target for the statistical-confidence
|
||||
/// axis. Single block, 1 thread; reads `ISV[KELLY_SAMPLE_COUNT_INDEX=283]`
|
||||
/// and writes to scratch. Downstream apply_pearls_ad_kernel smooths into
|
||||
/// ISV[KELLY_SAMPLE_COUNT_TARGET_INDEX=333].
|
||||
static SP9_KELLY_SAMPLE_COUNT_TARGET_EMA_CUBIN: &[u8] =
|
||||
include_bytes!(concat!(env!("OUT_DIR"), "/kelly_sample_count_target_ema_kernel.cubin"));
|
||||
|
||||
/// SP9 (Fix 37, 2026-05-03): EMA target for the behavioral-confidence
|
||||
/// axis. Single block, 1 thread; reads `ISV[INTENT_EVAL_DIVERGENCE_INDEX=332]`
|
||||
/// and writes to scratch. Downstream apply_pearls_ad_kernel smooths into
|
||||
/// ISV[KELLY_DIVERGENCE_TARGET_INDEX=334].
|
||||
static SP9_KELLY_DIVERGENCE_TARGET_EMA_CUBIN: &[u8] =
|
||||
include_bytes!(concat!(env!("OUT_DIR"), "/kelly_divergence_target_ema_kernel.cubin"));
|
||||
|
||||
/// SP9 (Fix 37, 2026-05-03): EMA target for the temporal-confidence axis.
|
||||
/// Single block, 1 thread; reads `ISV[EPOCH_IDX_INDEX=39]` and writes to
|
||||
/// scratch. Downstream apply_pearls_ad_kernel smooths into
|
||||
/// ISV[KELLY_TEMPORAL_TARGET_INDEX=335].
|
||||
static SP9_KELLY_TEMPORAL_TARGET_EMA_CUBIN: &[u8] =
|
||||
include_bytes!(concat!(env!("OUT_DIR"), "/kelly_temporal_target_ema_kernel.cubin"));
|
||||
|
||||
/// SP9 (Fix 37, 2026-05-03): main Kelly cold-start warmup-floor producer.
|
||||
/// Single block, 1 thread; reads 8 ISV inputs (q_var_mag, q_var_mag_ema,
|
||||
/// kelly_sample_count, sample_count_target, intent_eval_divergence,
|
||||
@@ -1210,12 +1189,14 @@ pub const SP5_WIENER_TOTAL_FLOATS: usize =
|
||||
/// 1 float for kelly_warmup_floor (SCRATCH_SP9_KELLY_WARMUP_FLOOR=259, at [259..260))
|
||||
/// 1 float for q_var_mag_ema (SCRATCH_SP9_Q_VAR_MAG_EMA=260, at [260..261))
|
||||
/// 1 float for intent_eval_divergence (SCRATCH_SP9_INTENT_EVAL_DIVERGENCE=261, at [261..262))
|
||||
/// 1 float for sample_count_target (SCRATCH_SP9_SAMPLE_COUNT_TARGET=262, at [262..263))
|
||||
/// 1 float for divergence_target (SCRATCH_SP9_DIVERGENCE_TARGET=263, at [263..264))
|
||||
/// 1 float for temporal_target (SCRATCH_SP9_TEMPORAL_TARGET=264, at [264..265))
|
||||
/// 3 floats for eval_dist[Q,H,F] (SCRATCH_SP9_EVAL_DIST_BASE=265, at [265..268))
|
||||
/// Combined: 259 + 9 = 268 scratch slots [0..268).
|
||||
pub const SP5_SCRATCH_TOTAL: usize = 268;
|
||||
/// 3 floats for eval_dist[Q,H,F] (SCRATCH_SP9_EVAL_DIST_BASE=262, at [262..265))
|
||||
/// Fix 37.1 (2026-05-03): the 3 EMA target updaters that previously sat
|
||||
/// at [262..265) (sample_count_target / divergence_target / temporal_target)
|
||||
/// are deleted — their target ISV slots are constructor-written
|
||||
/// Invariant-1 anchors per `feedback_isv_for_adaptive_bounds.md` and
|
||||
/// require no scratch storage. The eval_dist base slides 265 → 262.
|
||||
/// Combined: 259 + 6 = 265 scratch slots [0..265).
|
||||
pub const SP5_SCRATCH_TOTAL: usize = 265;
|
||||
|
||||
/// SP5 Layer D Task D1 (rewrite, 2026-05-02): scratch index base for
|
||||
/// `pnl_aggregation_update`.
|
||||
@@ -1307,17 +1288,20 @@ pub const SCRATCH_LB_MAX_BUDGET_CQL: usize = 251; // 251..255
|
||||
pub const SCRATCH_LB_MAX_BUDGET_C51: usize = 255; // 255..259
|
||||
|
||||
/// SP9 (Fix 37, 2026-05-03): scratch slots for Kelly cold-start warmup
|
||||
/// floor + EMA targets + intent/eval divergence + eval_dist GPU
|
||||
/// migration. Each producer kernel writes to its scratch slot; downstream
|
||||
/// floor + intent/eval divergence + eval_dist GPU migration. Each
|
||||
/// producer kernel writes to its scratch slot; downstream
|
||||
/// `apply_pearls_ad_kernel` smooths into the corresponding ISV slot via
|
||||
/// Pearl A sentinel-bootstrap + Pearl D Wiener-α steady-state smoothing.
|
||||
///
|
||||
/// Fix 37.1 (2026-05-03): the 3 EMA target updater scratch slots
|
||||
/// (sample_count_target / divergence_target / temporal_target) are
|
||||
/// deleted — their target ISV slots are constructor-written
|
||||
/// Invariant-1 anchors per `feedback_isv_for_adaptive_bounds.md` and
|
||||
/// require no scratch storage. The eval_dist base slides 265 → 262.
|
||||
pub const SCRATCH_SP9_KELLY_WARMUP_FLOOR: usize = 259; // → ISV[KELLY_WARMUP_FLOOR_INDEX=330]
|
||||
pub const SCRATCH_SP9_Q_VAR_MAG_EMA: usize = 260; // → ISV[Q_VAR_MAG_EMA_INDEX=331]
|
||||
pub const SCRATCH_SP9_INTENT_EVAL_DIVERGENCE: usize = 261; // → ISV[INTENT_EVAL_DIVERGENCE_INDEX=332]
|
||||
pub const SCRATCH_SP9_SAMPLE_COUNT_TARGET: usize = 262; // → ISV[KELLY_SAMPLE_COUNT_TARGET_INDEX=333]
|
||||
pub const SCRATCH_SP9_DIVERGENCE_TARGET: usize = 263; // → ISV[KELLY_DIVERGENCE_TARGET_INDEX=334]
|
||||
pub const SCRATCH_SP9_TEMPORAL_TARGET: usize = 264; // → ISV[KELLY_TEMPORAL_TARGET_INDEX=335]
|
||||
pub const SCRATCH_SP9_EVAL_DIST_BASE: usize = 265; // → ISV[EVAL_DIST_Q/H/F_INDEX=336..339)
|
||||
pub const SCRATCH_SP9_EVAL_DIST_BASE: usize = 262; // → ISV[EVAL_DIST_Q/H/F_INDEX=336..339)
|
||||
|
||||
/// SP5 Task A7: scratch index base for pearl_8_trail_update trail_dist[4] output block.
|
||||
/// Slots [199..203): per-direction trail-stop distance (Short=0, Hold=1, Long=2, Flat=3).
|
||||
@@ -4605,7 +4589,7 @@ pub struct GpuDqnTrainer {
|
||||
/// SP9 (Fix 37, 2026-05-03): GPU producer for eval-side magnitude
|
||||
/// intent distribution. Replaces DtoH `read_eval_intent_magnitude_distribution()`.
|
||||
/// Single-block, 3-thread kernel (one per Q/H/F bin); writes 3 floats
|
||||
/// to `scratch[SCRATCH_SP9_EVAL_DIST_BASE=265..268)`. Followed by 3
|
||||
/// to `scratch[SCRATCH_SP9_EVAL_DIST_BASE=262..265)`. Followed by 3
|
||||
/// apply_pearls_ad_kernel launches → ISV[EVAL_DIST_Q/H/F_INDEX=336..339).
|
||||
/// Loaded from `eval_intent_dist_compute_kernel.cubin`.
|
||||
sp9_eval_intent_dist_compute_kernel: CudaFunction,
|
||||
@@ -4623,26 +4607,6 @@ pub struct GpuDqnTrainer {
|
||||
/// apply_pearls_ad_kernel launch → ISV[Q_VAR_MAG_EMA_INDEX=331].
|
||||
/// Loaded from `q_var_mag_ema_compute_kernel.cubin`.
|
||||
sp9_q_var_mag_ema_compute_kernel: CudaFunction,
|
||||
/// SP9 (Fix 37, 2026-05-03): EMA target for the statistical-confidence
|
||||
/// axis. Single-block, single-thread; reads `ISV[KELLY_SAMPLE_COUNT_INDEX=283]`;
|
||||
/// writes to `scratch[SCRATCH_SP9_SAMPLE_COUNT_TARGET=262]`. Followed
|
||||
/// by 1 apply_pearls_ad_kernel launch →
|
||||
/// ISV[KELLY_SAMPLE_COUNT_TARGET_INDEX=333]. Loaded from
|
||||
/// `kelly_sample_count_target_ema_kernel.cubin`.
|
||||
sp9_kelly_sample_count_target_ema_kernel: CudaFunction,
|
||||
/// SP9 (Fix 37, 2026-05-03): EMA target for the behavioral-confidence
|
||||
/// axis. Single-block, single-thread; reads
|
||||
/// `ISV[INTENT_EVAL_DIVERGENCE_INDEX=332]`; writes to
|
||||
/// `scratch[SCRATCH_SP9_DIVERGENCE_TARGET=263]`. Followed by 1
|
||||
/// apply_pearls_ad_kernel launch → ISV[KELLY_DIVERGENCE_TARGET_INDEX=334].
|
||||
/// Loaded from `kelly_divergence_target_ema_kernel.cubin`.
|
||||
sp9_kelly_divergence_target_ema_kernel: CudaFunction,
|
||||
/// SP9 (Fix 37, 2026-05-03): EMA target for the temporal-confidence
|
||||
/// axis. Single-block, single-thread; reads `ISV[EPOCH_IDX_INDEX=39]`;
|
||||
/// writes to `scratch[SCRATCH_SP9_TEMPORAL_TARGET=264]`. Followed by 1
|
||||
/// apply_pearls_ad_kernel launch → ISV[KELLY_TEMPORAL_TARGET_INDEX=335].
|
||||
/// Loaded from `kelly_temporal_target_ema_kernel.cubin`.
|
||||
sp9_kelly_temporal_target_ema_kernel: CudaFunction,
|
||||
/// SP9 (Fix 37, 2026-05-03): main Kelly cold-start warmup-floor producer.
|
||||
/// Single-block, single-thread; reads 8 ISV inputs and computes
|
||||
/// base_floor × (1 − max(statistical_conf, behavioral_conf, temporal_conf))
|
||||
@@ -12096,20 +12060,40 @@ impl GpuDqnTrainer {
|
||||
}
|
||||
|
||||
/// SP9 (Fix 37, 2026-05-03): launch the four cold-path EMA producers
|
||||
/// (q_var_mag_ema + 3 EMA targets) and the main warmup-floor producer.
|
||||
/// (q_var_mag_ema) and the main warmup-floor producer.
|
||||
///
|
||||
/// All five kernels are single-block, single-thread cold-path producers
|
||||
/// Both kernels are single-block, single-thread cold-path producers
|
||||
/// chained through `apply_pearls_ad_kernel`. The launch order matches
|
||||
/// the SP9 dependency DAG:
|
||||
/// 1. q_var_mag_ema_compute — reads ISV[FLATNESS_BASE+1]
|
||||
/// 2. kelly_sample_count_target_ema_compute — reads ISV[KELLY_SAMPLE_COUNT_INDEX]
|
||||
/// 3. kelly_divergence_target_ema_compute — reads ISV[INTENT_EVAL_DIVERGENCE_INDEX]
|
||||
/// 4. kelly_temporal_target_ema_compute — reads ISV[EPOCH_IDX_INDEX]
|
||||
/// 5. kelly_warmup_floor_compute — reads all 8 inputs above + ISV
|
||||
/// targets just produced
|
||||
/// 2. kelly_warmup_floor_compute — reads 8 inputs (q_var_mag,
|
||||
/// q_var_mag_ema, kelly_sample_count,
|
||||
/// sample_count_target, intent_eval_divergence,
|
||||
/// divergence_target, epoch_idx,
|
||||
/// temporal_target). The 3 target
|
||||
/// slots are constructor-written
|
||||
/// Invariant-1 anchors per Fix 37.1
|
||||
/// (`feedback_isv_for_adaptive_bounds.md`).
|
||||
///
|
||||
/// Fix 37.1 (2026-05-03): the 3 EMA target updaters that previously sat
|
||||
/// between #1 and #2 (kelly_sample_count_target_ema /
|
||||
/// kelly_divergence_target_ema / kelly_temporal_target_ema) are deleted.
|
||||
/// Their target slots ISV[333..336) are now numerical-stability anchors
|
||||
/// — `KELLY_SAMPLE_COUNT_TARGET_INDEX = 100.0`,
|
||||
/// `KELLY_DIVERGENCE_TARGET_INDEX = 2.0`,
|
||||
/// `KELLY_TEMPORAL_TARGET_INDEX = 5.0` — written ONCE at trainer
|
||||
/// constructor time. Pearl A sentinel-bootstrap on the original EMA
|
||||
/// path saturated all three to the very first observation in epoch 1
|
||||
/// (`stat_count_tgt=419, div_tgt=70005, temp_tgt=1`), which made
|
||||
/// `current/target = 1.0`, `confidence = 1.0` and
|
||||
/// `floor = base × (1 − 1.0) = 0` — defeating the cold-start floor.
|
||||
/// A fixed denominator yields a meaningful target-relative ratio that
|
||||
/// rises monotonically as samples / divergence-stability / fold-time
|
||||
/// accumulate.
|
||||
///
|
||||
/// Must run AFTER:
|
||||
/// * `launch_intent_eval_divergence_compute` (for input #3)
|
||||
/// * `launch_intent_eval_divergence_compute` (so ISV[332] is fresh
|
||||
/// before the warmup-floor kernel reads it)
|
||||
/// Must run BEFORE:
|
||||
/// * any `unified_env_step_core` (which reads ISV[KELLY_WARMUP_FLOOR_INDEX])
|
||||
pub(crate) fn launch_sp9_kelly_warmup_floor(&self) -> Result<(), MLError> {
|
||||
@@ -12185,34 +12169,9 @@ impl GpuDqnTrainer {
|
||||
"q_var_mag_ema_compute",
|
||||
)?;
|
||||
|
||||
// 2. kelly_sample_count_target_ema (reads ISV[283])
|
||||
launch_simple_ema(
|
||||
&self.sp9_kelly_sample_count_target_ema_kernel,
|
||||
KELLY_SAMPLE_COUNT_INDEX,
|
||||
SCRATCH_SP9_SAMPLE_COUNT_TARGET,
|
||||
KELLY_SAMPLE_COUNT_TARGET_INDEX,
|
||||
"kelly_sample_count_target_ema_compute",
|
||||
)?;
|
||||
|
||||
// 3. kelly_divergence_target_ema (reads ISV[332])
|
||||
launch_simple_ema(
|
||||
&self.sp9_kelly_divergence_target_ema_kernel,
|
||||
INTENT_EVAL_DIVERGENCE_INDEX,
|
||||
SCRATCH_SP9_DIVERGENCE_TARGET,
|
||||
KELLY_DIVERGENCE_TARGET_INDEX,
|
||||
"kelly_divergence_target_ema_compute",
|
||||
)?;
|
||||
|
||||
// 4. kelly_temporal_target_ema (reads ISV[39] EPOCH_IDX_INDEX)
|
||||
launch_simple_ema(
|
||||
&self.sp9_kelly_temporal_target_ema_kernel,
|
||||
EPOCH_IDX_INDEX,
|
||||
SCRATCH_SP9_TEMPORAL_TARGET,
|
||||
KELLY_TEMPORAL_TARGET_INDEX,
|
||||
"kelly_temporal_target_ema_compute",
|
||||
)?;
|
||||
|
||||
// 5. kelly_warmup_floor_compute — combines all 8 inputs.
|
||||
// 2. kelly_warmup_floor_compute — combines all 8 inputs (target
|
||||
// slots ISV[333..336) read directly from constructor-written
|
||||
// Invariant-1 anchors per Fix 37.1; no producer chain).
|
||||
let q_var_mag_idx_i32: i32 = (FLATNESS_BASE + 1) as i32;
|
||||
let q_var_mag_ema_idx_i32: i32 = Q_VAR_MAG_EMA_INDEX as i32;
|
||||
let kelly_sample_count_idx_i32: i32 = KELLY_SAMPLE_COUNT_INDEX as i32;
|
||||
@@ -15552,27 +15511,6 @@ impl GpuDqnTrainer {
|
||||
module.load_function("q_var_mag_ema_compute")
|
||||
.map_err(|e| MLError::ModelError(format!("sp9 q_var_mag_ema_compute load: {e}")))?
|
||||
};
|
||||
let sp9_kelly_sample_count_target_ema_kernel = {
|
||||
let module = stream.context()
|
||||
.load_cubin(SP9_KELLY_SAMPLE_COUNT_TARGET_EMA_CUBIN.to_vec())
|
||||
.map_err(|e| MLError::ModelError(format!("sp9 kelly_sample_count_target_ema cubin load: {e}")))?;
|
||||
module.load_function("kelly_sample_count_target_ema_compute")
|
||||
.map_err(|e| MLError::ModelError(format!("sp9 kelly_sample_count_target_ema_compute load: {e}")))?
|
||||
};
|
||||
let sp9_kelly_divergence_target_ema_kernel = {
|
||||
let module = stream.context()
|
||||
.load_cubin(SP9_KELLY_DIVERGENCE_TARGET_EMA_CUBIN.to_vec())
|
||||
.map_err(|e| MLError::ModelError(format!("sp9 kelly_divergence_target_ema cubin load: {e}")))?;
|
||||
module.load_function("kelly_divergence_target_ema_compute")
|
||||
.map_err(|e| MLError::ModelError(format!("sp9 kelly_divergence_target_ema_compute load: {e}")))?
|
||||
};
|
||||
let sp9_kelly_temporal_target_ema_kernel = {
|
||||
let module = stream.context()
|
||||
.load_cubin(SP9_KELLY_TEMPORAL_TARGET_EMA_CUBIN.to_vec())
|
||||
.map_err(|e| MLError::ModelError(format!("sp9 kelly_temporal_target_ema cubin load: {e}")))?;
|
||||
module.load_function("kelly_temporal_target_ema_compute")
|
||||
.map_err(|e| MLError::ModelError(format!("sp9 kelly_temporal_target_ema_compute load: {e}")))?
|
||||
};
|
||||
let sp9_kelly_warmup_floor_compute_kernel = {
|
||||
let module = stream.context()
|
||||
.load_cubin(SP9_KELLY_WARMUP_FLOOR_COMPUTE_CUBIN.to_vec())
|
||||
@@ -17674,6 +17612,48 @@ impl GpuDqnTrainer {
|
||||
* `training_loop.rs` prevents the producer launch until
|
||||
* a meaningful prev value exists). */
|
||||
*sig_ptr.add(Q_DRIFT_RATE_INDEX) = 0.0_f32;
|
||||
// SP9 Fix 37.1 (2026-05-03): the 3 Kelly cold-start exit-axis
|
||||
// targets are Invariant-1 numerical anchors per
|
||||
// `feedback_isv_for_adaptive_bounds.md`, not EMA-tracked.
|
||||
// The original Fix 37 routed each through a single-input
|
||||
// EMA producer + Pearl A sentinel-bootstrap + Pearl D
|
||||
// Wiener-α; smoke-test-wrwkz revealed this caused the
|
||||
// first observation to replace the sentinel directly,
|
||||
// making target == current_obs in epoch 1 (`stat_count_tgt
|
||||
// = 419`, `temp_tgt = 1`). Consequently `current/target =
|
||||
// 1.0`, `confidence = 1.0`, `floor = base × (1 − 1.0) = 0`
|
||||
// — defeating the cold-start mechanism entirely.
|
||||
//
|
||||
// The principled fix per the pearl: define what
|
||||
// "sufficient" means in absolute terms via Invariant-1
|
||||
// anchors. The target-relative ratio
|
||||
// `current_observation / fixed_anchor` then yields a
|
||||
// self-normalised confidence in [0, 1] that rises
|
||||
// monotonically as samples / divergence-stability /
|
||||
// fold-time accumulate.
|
||||
//
|
||||
// ISV[KELLY_SAMPLE_COUNT_TARGET_INDEX = 333] = 100.0
|
||||
// — minimum trade samples before statistical confidence
|
||||
// releases the cap.
|
||||
// ISV[KELLY_DIVERGENCE_TARGET_INDEX = 334] = 2.0
|
||||
// — divergence ratio above which intent-vs-eval are
|
||||
// considered diverged enough to keep the cap closed.
|
||||
// ISV[KELLY_TEMPORAL_TARGET_INDEX = 335] = 5.0
|
||||
// — minimum epochs in fold before temporal confidence
|
||||
// releases the cap (safety net per Risk 3 in spec).
|
||||
//
|
||||
// FoldReset rewrites the same anchors (in `reset_named_state`)
|
||||
// since these are constants, not stateful EMAs — they must
|
||||
// never reach 0 sentinel between folds.
|
||||
use crate::cuda_pipeline::sp5_isv_slots::{
|
||||
KELLY_SAMPLE_COUNT_TARGET_INDEX,
|
||||
KELLY_DIVERGENCE_TARGET_INDEX,
|
||||
KELLY_TEMPORAL_TARGET_INDEX,
|
||||
};
|
||||
*sig_ptr.add(KELLY_SAMPLE_COUNT_TARGET_INDEX) = 100.0_f32;
|
||||
*sig_ptr.add(KELLY_DIVERGENCE_TARGET_INDEX) = 2.0_f32;
|
||||
*sig_ptr.add(KELLY_TEMPORAL_TARGET_INDEX) = 5.0_f32;
|
||||
|
||||
// Layout fingerprint (ISV[58..60)). Compile-time structural hash of
|
||||
// the slot layout; checkpoint load fails-fast on mismatch.
|
||||
// Stored as a u64 split across two f32 lanes using raw bit-cast so
|
||||
@@ -18799,9 +18779,6 @@ impl GpuDqnTrainer {
|
||||
sp9_eval_intent_dist_compute_kernel,
|
||||
sp9_intent_eval_divergence_compute_kernel,
|
||||
sp9_q_var_mag_ema_compute_kernel,
|
||||
sp9_kelly_sample_count_target_ema_kernel,
|
||||
sp9_kelly_divergence_target_ema_kernel,
|
||||
sp9_kelly_temporal_target_ema_kernel,
|
||||
sp9_kelly_warmup_floor_compute_kernel,
|
||||
lb_budget_dispatch_kernel,
|
||||
saxpy_f32_dev_ptr_aux,
|
||||
|
||||
@@ -26,8 +26,11 @@
|
||||
// ratio rises sharply, signalling the controller. The behavioral_conf axis
|
||||
// of the warmup-floor formula maps `clamp(1 − divergence/divergence_target, 0, 1)`
|
||||
// — divergence above target → confidence = 0 (cap stays closed); divergence
|
||||
// at or below target → confidence rises (cap relaxes). EMA target itself is
|
||||
// adaptive (Wiener-α via Pearl D) — see `kelly_divergence_target_ema_kernel.cu`.
|
||||
// at or below target → confidence rises (cap relaxes). The divergence
|
||||
// target ISV[KELLY_DIVERGENCE_TARGET_INDEX=334] is an Invariant-1 numerical
|
||||
// anchor (2.0) per Fix 37.1 (`feedback_isv_for_adaptive_bounds.md`),
|
||||
// constructor-written once and rewritten on FoldReset — see audit Fix 37.1
|
||||
// for why the original Fix 37 EMA-tracked target was abandoned.
|
||||
//
|
||||
// Single block, single thread; cheap arithmetic. No atomicAdd
|
||||
// (`feedback_no_atomicadd.md`); `__threadfence_system()` after the write.
|
||||
@@ -47,10 +50,14 @@ extern "C" __global__ void intent_eval_divergence_compute(
|
||||
) {
|
||||
if (blockIdx.x != 0 || threadIdx.x != 0) return;
|
||||
|
||||
/* Numerical-stability anchor (Invariant 1) — prevents divide-by-zero
|
||||
* when intent_f or eval_f is exactly zero. Same EPS scale used by
|
||||
* `train_active_frac_compute_kernel` and the SP4 Pearl D recipe. */
|
||||
const float EPS_DIV = 1e-6f;
|
||||
/* Numerical-stability anchor (Invariant 1) — distinguishes a freshly-
|
||||
* fold-reset / never-populated `eval_f` (Pearl A sentinel = 0) from a
|
||||
* tiny-but-real `eval_f` value. Anything strictly below this threshold
|
||||
* is treated as "uninitialised" because `eval_dist` writes are
|
||||
* naturally bounded by [0, 1] from `eval_intent_dist_compute_kernel`'s
|
||||
* normalised output, and 1e-5 sits well below any realisable value
|
||||
* after even a single observation enters the Pearl A bootstrap path. */
|
||||
const float SENTINEL_THRESHOLD = 1e-5f;
|
||||
|
||||
/* Sum action_counts over Full magnitude bins: exp_idx = dir*3 + 2 for
|
||||
* dir ∈ {0..4}. Bins 2 (Short Full), 5 (Hold Full), 8 (Long Full),
|
||||
@@ -74,11 +81,30 @@ extern "C" __global__ void intent_eval_divergence_compute(
|
||||
if (eval_f < 0.0f) eval_f = 0.0f;
|
||||
if (eval_f > 1.0f) eval_f = 1.0f;
|
||||
|
||||
/* Divergence = intent_f / eval_f. Both numerator and denominator are
|
||||
* floored at EPS_DIV to prevent NaN/Inf when either is 0. Range under
|
||||
* healthy training: ~1; under val-Flat-collapse: large (intent up,
|
||||
* eval down). */
|
||||
float divergence = fmaxf(intent_f, EPS_DIV) / fmaxf(eval_f, EPS_DIV);
|
||||
/* Fix 37.1 (2026-05-03): handle Pearl A sentinel for `eval_f`. Before
|
||||
* the first val-window has populated `ISV[EVAL_DIST_F_INDEX=338]`,
|
||||
* the slot still carries its sentinel 0 (per
|
||||
* `pearl_first_observation_bootstrap.md`). Reading it as a real
|
||||
* denominator and flooring at any `EPS_DIV` produced a synthetic
|
||||
* `divergence ≈ intent_f / 1e-6 ≈ 70 000`, which the warmup-floor
|
||||
* kernel then mapped to `behavioral_conf = clamp(1 − 70000/2, 0, 1) = 0`
|
||||
* — algebraically the right answer "no behavioral signal yet", but
|
||||
* coupled to a magic-floor `EPS_DIV` and easily missed in
|
||||
* smoke logs (the synthetic 70 000 looked like a runaway).
|
||||
* Instead: explicitly emit `SENTINEL_DIVERGENCE` (large enough that
|
||||
* `divergence / divergence_target ≫ 1` regardless of target) so the
|
||||
* downstream warmup-floor kernel's `behavioral_conf` clamp lands
|
||||
* cleanly at 0 with provenance traceable in HEALTH_DIAG. Once
|
||||
* `eval_dist_compute` fires its first val-window observation,
|
||||
* Pearl A's first-observation-replacement raises `eval_f` above the
|
||||
* sentinel threshold and the real ratio takes over. */
|
||||
const float SENTINEL_DIVERGENCE = 1e6f;
|
||||
float divergence;
|
||||
if (eval_f < SENTINEL_THRESHOLD) {
|
||||
divergence = SENTINEL_DIVERGENCE;
|
||||
} else {
|
||||
divergence = intent_f / eval_f;
|
||||
}
|
||||
|
||||
scratch_buf[scratch_idx] = divergence;
|
||||
__threadfence_system();
|
||||
|
||||
@@ -1,40 +0,0 @@
|
||||
// crates/ml/src/cuda_pipeline/kelly_divergence_target_ema_kernel.cu
|
||||
//
|
||||
// SP9 (Fix 37, 2026-05-03): GPU producer for EMA-tracked intent vs eval
|
||||
// divergence target — the behavioral-confidence axis denominator of the
|
||||
// OR'd cold-start exit per `pearl_cold_start_exit_signal_or.md`.
|
||||
//
|
||||
// Reads ISV[INTENT_EVAL_DIVERGENCE_INDEX=332] (the SP9 ratio
|
||||
// `intent_f / eval_f` produced by `intent_eval_divergence_compute_kernel.cu`).
|
||||
// Writes that observation to `scratch_buf[scratch_idx]`.
|
||||
//
|
||||
// Downstream `apply_pearls_ad_kernel` chains:
|
||||
// - Pearl A (sentinel 0 at fold boundary → first observation replaces slot)
|
||||
// - Pearl D (Wiener-optimal α adapts to divergence-stability)
|
||||
// → ISV[KELLY_DIVERGENCE_TARGET_INDEX=334].
|
||||
//
|
||||
// Per `pearl_controller_anchors_isv_driven.md`: the "healthy divergence"
|
||||
// threshold is itself signal-driven via Pearl D smoothing — when divergence
|
||||
// is stable around any value, that value becomes the target; when divergence
|
||||
// shifts (val-Flat-collapse onset), the EMA lags and the consumer's ratio
|
||||
// `divergence / target` rises sharply, dropping behavioral_conf to 0.
|
||||
//
|
||||
// Single-block, single-thread cold-path kernel. No atomicAdd; no shared
|
||||
// memory; `__threadfence_system()` after the write.
|
||||
|
||||
#include <cuda_runtime.h>
|
||||
|
||||
extern "C" __global__ void kelly_divergence_target_ema_compute(
|
||||
const float* __restrict__ isv_signals,
|
||||
int intent_eval_divergence_isv_index, // INTENT_EVAL_DIVERGENCE_INDEX = 332
|
||||
float* __restrict__ scratch_buf,
|
||||
int scratch_idx
|
||||
) {
|
||||
if (blockIdx.x != 0 || threadIdx.x != 0) return;
|
||||
|
||||
float observed = isv_signals[intent_eval_divergence_isv_index];
|
||||
if (observed < 0.0f) observed = 0.0f;
|
||||
|
||||
scratch_buf[scratch_idx] = observed;
|
||||
__threadfence_system();
|
||||
}
|
||||
@@ -1,42 +0,0 @@
|
||||
// crates/ml/src/cuda_pipeline/kelly_sample_count_target_ema_kernel.cu
|
||||
//
|
||||
// SP9 (Fix 37, 2026-05-03): GPU producer for EMA-tracked Kelly
|
||||
// sample-sufficiency target — the statistical-confidence axis denominator
|
||||
// of the OR'd cold-start exit per `pearl_cold_start_exit_signal_or.md`.
|
||||
//
|
||||
// Reads ISV[KELLY_SAMPLE_COUNT_INDEX=283] (Pearl 6's cumulative-via-max
|
||||
// trade count). Writes that observation to `scratch_buf[scratch_idx]`.
|
||||
//
|
||||
// Downstream `apply_pearls_ad_kernel` chains:
|
||||
// - Pearl A (sentinel 0 at fold boundary → first observation replaces slot)
|
||||
// - Pearl D (Wiener-optimal α — variance ratio adapts the smoothing rate)
|
||||
// → ISV[KELLY_SAMPLE_COUNT_TARGET_INDEX=333].
|
||||
//
|
||||
// Why an EMA-tracked target instead of a hardcoded constant: per
|
||||
// `pearl_controller_anchors_isv_driven.md`, the threshold for "enough
|
||||
// statistical confidence to release the cap" should track the running
|
||||
// value, not be tuned. Pearl D's variance-ratio-derived α slows the EMA
|
||||
// when the signal is stable and speeds it up when it shifts — Wiener-
|
||||
// optimal blending. The target-relative ratio `kelly_sample_count /
|
||||
// sample_count_target` becomes a self-normalised confidence in [0, 1]
|
||||
// (clamped in the consumer).
|
||||
//
|
||||
// Single-block, single-thread cold-path kernel. No atomicAdd; no shared
|
||||
// memory; `__threadfence_system()` after the write.
|
||||
|
||||
#include <cuda_runtime.h>
|
||||
|
||||
extern "C" __global__ void kelly_sample_count_target_ema_compute(
|
||||
const float* __restrict__ isv_signals,
|
||||
int kelly_sample_count_isv_index, // KELLY_SAMPLE_COUNT_INDEX = 283
|
||||
float* __restrict__ scratch_buf,
|
||||
int scratch_idx
|
||||
) {
|
||||
if (blockIdx.x != 0 || threadIdx.x != 0) return;
|
||||
|
||||
float observed = isv_signals[kelly_sample_count_isv_index];
|
||||
if (observed < 0.0f) observed = 0.0f;
|
||||
|
||||
scratch_buf[scratch_idx] = observed;
|
||||
__threadfence_system();
|
||||
}
|
||||
@@ -1,41 +0,0 @@
|
||||
// crates/ml/src/cuda_pipeline/kelly_temporal_target_ema_kernel.cu
|
||||
//
|
||||
// SP9 (Fix 37, 2026-05-03): GPU producer for EMA-tracked temporal target
|
||||
// — the temporal-confidence axis denominator of the OR'd cold-start exit
|
||||
// per `pearl_cold_start_exit_signal_or.md`.
|
||||
//
|
||||
// Reads ISV[EPOCH_IDX_INDEX=39] (the existing per-fold epoch counter).
|
||||
// Writes that observation to `scratch_buf[scratch_idx]`.
|
||||
//
|
||||
// Downstream `apply_pearls_ad_kernel` chains:
|
||||
// - Pearl A (sentinel 0 at fold boundary → first observation replaces slot)
|
||||
// - Pearl D (Wiener-optimal α adapts to the steady-state epoch growth rate)
|
||||
// → ISV[KELLY_TEMPORAL_TARGET_INDEX=335].
|
||||
//
|
||||
// Per `pearl_controller_anchors_isv_driven.md`: even the temporal target
|
||||
// is signal-driven — the Wiener-EMA tracks the running epoch_idx so the
|
||||
// consumer's `epoch_idx / temporal_target` ratio stays in a useful range
|
||||
// regardless of fold length or maximum epoch count. Per Risk 3 in the SP9
|
||||
// spec: this axis is the safety net — if statistical and behavioral both
|
||||
// stay dormant (e.g. fold 0 has no validation samples yet), temporal alone
|
||||
// eventually fires once epoch_idx_in_fold reaches its EMA-target.
|
||||
//
|
||||
// Single-block, single-thread cold-path kernel. No atomicAdd; no shared
|
||||
// memory; `__threadfence_system()` after the write.
|
||||
|
||||
#include <cuda_runtime.h>
|
||||
|
||||
extern "C" __global__ void kelly_temporal_target_ema_compute(
|
||||
const float* __restrict__ isv_signals,
|
||||
int epoch_idx_isv_index, // EPOCH_IDX_INDEX = 39
|
||||
float* __restrict__ scratch_buf,
|
||||
int scratch_idx
|
||||
) {
|
||||
if (blockIdx.x != 0 || threadIdx.x != 0) return;
|
||||
|
||||
float observed = isv_signals[epoch_idx_isv_index];
|
||||
if (observed < 0.0f) observed = 0.0f;
|
||||
|
||||
scratch_buf[scratch_idx] = observed;
|
||||
__threadfence_system();
|
||||
}
|
||||
@@ -24,11 +24,11 @@
|
||||
// - q_var_mag ← ISV[FLATNESS_BASE+1=207] (Pearl 2 magnitude flatness)
|
||||
// - q_var_mag_ema ← ISV[Q_VAR_MAG_EMA_INDEX=331] (SP9 EMA-tracked baseline)
|
||||
// - kelly_sample_count ← ISV[KELLY_SAMPLE_COUNT_INDEX=283] (Pearl 6 cumulative)
|
||||
// - sample_count_target ← ISV[KELLY_SAMPLE_COUNT_TARGET_INDEX=333] (SP9 EMA target)
|
||||
// - sample_count_target ← ISV[KELLY_SAMPLE_COUNT_TARGET_INDEX=333] (Fix 37.1 Invariant-1 anchor = 100.0)
|
||||
// - intent_eval_divergence← ISV[INTENT_EVAL_DIVERGENCE_INDEX=332] (SP9 producer)
|
||||
// - divergence_target ← ISV[KELLY_DIVERGENCE_TARGET_INDEX=334] (SP9 EMA target)
|
||||
// - divergence_target ← ISV[KELLY_DIVERGENCE_TARGET_INDEX=334] (Fix 37.1 Invariant-1 anchor = 2.0)
|
||||
// - epoch_idx ← ISV[EPOCH_IDX_INDEX=39] (existing per-fold counter)
|
||||
// - temporal_target ← ISV[KELLY_TEMPORAL_TARGET_INDEX=335] (SP9 EMA target)
|
||||
// - temporal_target ← ISV[KELLY_TEMPORAL_TARGET_INDEX=335] (Fix 37.1 Invariant-1 anchor = 5.0)
|
||||
//
|
||||
// Writes: 1 float to `scratch_buf[scratch_idx]`. Downstream
|
||||
// `apply_pearls_ad_kernel` smooths into ISV[KELLY_WARMUP_FLOOR_INDEX=330].
|
||||
|
||||
@@ -720,17 +720,17 @@ impl StateResetRegistry {
|
||||
RegistryEntry {
|
||||
name: "sp9_kelly_sample_count_target",
|
||||
category: ResetCategory::FoldReset,
|
||||
description: "ISV[KELLY_SAMPLE_COUNT_TARGET_INDEX=333] — SP9 Fix 37 EMA-tracked sample-sufficiency target. Statistical-axis denominator of the OR'd cold-start exit. Produced by `kelly_sample_count_target_ema_kernel` from `ISV[KELLY_SAMPLE_COUNT_INDEX=283]` (Pearl 6's cumulative-via-max trade count); smoothed by Pearl D Wiener-α. The consumer's `kelly_sample_count / sample_count_target` ratio yields a self-normalised `statistical_conf` in [0, 1]. FoldReset sentinel 0 — Pearl A bootstraps from the new fold's first observation.",
|
||||
description: "ISV[KELLY_SAMPLE_COUNT_TARGET_INDEX=333] — SP9 Fix 37.1 (2026-05-03) Invariant-1 numerical anchor: minimum trade samples (100.0) for statistical-confidence release of the OR'd cold-start exit. Originally Fix 37 routed this through `kelly_sample_count_target_ema_kernel` + Pearl A/D smoothing, but smoke-test-wrwkz revealed sentinel-bootstrap saturated the EMA to the first observation in epoch 1, making `current/target = 1.0` and defeating the cold-start floor. Now constructor-written ONCE per `feedback_isv_for_adaptive_bounds.md`. FoldReset rewrites the same Invariant-1 anchor (NOT sentinel 0) since this is a constant, not a stateful EMA. The consumer's `kelly_sample_count / sample_count_target` ratio yields a self-normalised `statistical_conf` in [0, 1] that rises monotonically as samples accumulate.",
|
||||
},
|
||||
RegistryEntry {
|
||||
name: "sp9_kelly_divergence_target",
|
||||
category: ResetCategory::FoldReset,
|
||||
description: "ISV[KELLY_DIVERGENCE_TARGET_INDEX=334] — SP9 Fix 37 EMA-tracked behavioral-gap target. Behavioral-axis denominator of the OR'd cold-start exit. Produced by `kelly_divergence_target_ema_kernel` from `ISV[INTENT_EVAL_DIVERGENCE_INDEX=332]`; smoothed by Pearl D Wiener-α. The consumer's `1 − divergence / divergence_target` ratio yields a self-normalised `behavioral_conf` in [0, 1]. FoldReset sentinel 0 — Pearl A bootstraps from the new fold's first observation.",
|
||||
description: "ISV[KELLY_DIVERGENCE_TARGET_INDEX=334] — SP9 Fix 37.1 (2026-05-03) Invariant-1 numerical anchor: divergence ratio (2.0) above which intent-vs-eval is considered diverged enough to keep the cap closed (behavioral-axis denominator of the OR'd cold-start exit). Originally Fix 37 routed this through `kelly_divergence_target_ema_kernel` + Pearl A/D smoothing, but smoke-test-wrwkz revealed sentinel-bootstrap saturated the EMA to the first observation in epoch 1 (`div_tgt=70 005`), making the consumer's ratio collapse to 1.0 and defeating the cold-start floor. Now constructor-written ONCE per `feedback_isv_for_adaptive_bounds.md`. FoldReset rewrites the same Invariant-1 anchor (NOT sentinel 0). The consumer's `1 − divergence / divergence_target` ratio yields a self-normalised `behavioral_conf` in [0, 1] that drops sharply when divergence exceeds the threshold.",
|
||||
},
|
||||
RegistryEntry {
|
||||
name: "sp9_kelly_temporal_target",
|
||||
category: ResetCategory::FoldReset,
|
||||
description: "ISV[KELLY_TEMPORAL_TARGET_INDEX=335] — SP9 Fix 37 EMA-tracked temporal target. Temporal-axis denominator of the OR'd cold-start exit; the safety net per Risk 3 in the SP9 spec (always fires once epoch_idx ≥ target if statistical and behavioral both stay dormant). Produced by `kelly_temporal_target_ema_kernel` from `ISV[EPOCH_IDX_INDEX=39]`; smoothed by Pearl D Wiener-α. FoldReset sentinel 0 — Pearl A bootstraps from the new fold's first observation.",
|
||||
description: "ISV[KELLY_TEMPORAL_TARGET_INDEX=335] — SP9 Fix 37.1 (2026-05-03) Invariant-1 numerical anchor: minimum epochs in fold (5.0) before temporal-confidence release of the OR'd cold-start exit (the safety net per Risk 3 in the SP9 spec — always fires once epoch_idx ≥ target if statistical and behavioral both stay dormant). Originally Fix 37 routed this through `kelly_temporal_target_ema_kernel` + Pearl A/D smoothing, but smoke-test-wrwkz revealed sentinel-bootstrap saturated the EMA to the first observation in epoch 1 (`temp_tgt=1.0`), making temporal_conf trivially `epoch_idx/1.0 ≥ 1` from epoch 1 onwards. Now constructor-written ONCE per `feedback_isv_for_adaptive_bounds.md`. FoldReset rewrites the same Invariant-1 anchor (NOT sentinel 0). The consumer's `epoch_idx_in_fold / temporal_target` ratio yields a self-normalised `temporal_conf` in [0, 1] that rises linearly with fold-time.",
|
||||
},
|
||||
RegistryEntry {
|
||||
name: "sp9_eval_dist_q",
|
||||
|
||||
@@ -6896,21 +6896,38 @@ impl DQNTrainer {
|
||||
}
|
||||
}
|
||||
"sp9_kelly_sample_count_target" => {
|
||||
// SP9 Fix 37.1 (2026-05-03): Invariant-1 numerical anchor —
|
||||
// not a stateful EMA. Rewrite the constructor's value at
|
||||
// fold boundary so the new fold sees the correct
|
||||
// sample-sufficiency threshold (100 trades) for the
|
||||
// statistical-confidence axis. Per
|
||||
// `feedback_isv_for_adaptive_bounds.md`.
|
||||
if let Some(ref fused) = self.fused_ctx {
|
||||
use crate::cuda_pipeline::sp5_isv_slots::KELLY_SAMPLE_COUNT_TARGET_INDEX;
|
||||
fused.trainer().write_isv_signal_at(KELLY_SAMPLE_COUNT_TARGET_INDEX, 0.0);
|
||||
fused.trainer().write_isv_signal_at(KELLY_SAMPLE_COUNT_TARGET_INDEX, 100.0);
|
||||
}
|
||||
}
|
||||
"sp9_kelly_divergence_target" => {
|
||||
// SP9 Fix 37.1 (2026-05-03): Invariant-1 numerical anchor —
|
||||
// not a stateful EMA. Rewrite the constructor's value at
|
||||
// fold boundary so the new fold sees the correct
|
||||
// divergence threshold (2.0) for the behavioral-confidence
|
||||
// axis. Per `feedback_isv_for_adaptive_bounds.md`.
|
||||
if let Some(ref fused) = self.fused_ctx {
|
||||
use crate::cuda_pipeline::sp5_isv_slots::KELLY_DIVERGENCE_TARGET_INDEX;
|
||||
fused.trainer().write_isv_signal_at(KELLY_DIVERGENCE_TARGET_INDEX, 0.0);
|
||||
fused.trainer().write_isv_signal_at(KELLY_DIVERGENCE_TARGET_INDEX, 2.0);
|
||||
}
|
||||
}
|
||||
"sp9_kelly_temporal_target" => {
|
||||
// SP9 Fix 37.1 (2026-05-03): Invariant-1 numerical anchor —
|
||||
// not a stateful EMA. Rewrite the constructor's value at
|
||||
// fold boundary so the new fold sees the correct
|
||||
// fold-temporal threshold (5 epochs) for the
|
||||
// temporal-confidence axis. Per
|
||||
// `feedback_isv_for_adaptive_bounds.md`.
|
||||
if let Some(ref fused) = self.fused_ctx {
|
||||
use crate::cuda_pipeline::sp5_isv_slots::KELLY_TEMPORAL_TARGET_INDEX;
|
||||
fused.trainer().write_isv_signal_at(KELLY_TEMPORAL_TARGET_INDEX, 0.0);
|
||||
fused.trainer().write_isv_signal_at(KELLY_TEMPORAL_TARGET_INDEX, 5.0);
|
||||
}
|
||||
}
|
||||
"sp9_eval_dist_q" => {
|
||||
|
||||
@@ -4679,3 +4679,170 @@ fmaxf chain. Per-fold persistence vs cross-fold persistence is the
|
||||
existing carve-out — Pearl 6's KELLY_F_SMOOTH stays cross-fold; SP9's
|
||||
warmup_floor is per-fold so each fold's cold-start protection re-engages
|
||||
cleanly.
|
||||
|
||||
## Fix 37.1 — SP9 Kelly cold-start follow-up: targets as Invariant-1 anchors + divergence sentinel handling (2026-05-03)
|
||||
|
||||
**Symptom:** smoke-test-wrwkz (5-epoch L40S smoke on Fix 37 commit
|
||||
`48a8b9ee7`) ep1 HEALTH_DIAG line:
|
||||
|
||||
```
|
||||
sp9_kelly_warmup [floor=0.0000 divergence=70005.93 q_var_mag_ema=6.85e-2
|
||||
conf [stat=1.000 bhv=0.333 tmp=1.000]
|
||||
targets [stat_count_tgt=419 div_tgt=105002 temp_tgt=1.0]]
|
||||
```
|
||||
|
||||
`floor=0` and `conf=1` from epoch 1 onward — the cold-start mechanism is
|
||||
effectively never engaged.
|
||||
|
||||
**Root cause (two distinct quirks):**
|
||||
|
||||
*Quirk 1 — EMA target saturation:* The 3 target ISV slots
|
||||
(`KELLY_SAMPLE_COUNT_TARGET_INDEX=333`, `KELLY_DIVERGENCE_TARGET_INDEX=334`,
|
||||
`KELLY_TEMPORAL_TARGET_INDEX=335`) were originally Fix 37 routed through
|
||||
`kelly_*_target_ema_kernel.cu` + `apply_pearls_ad_kernel`. Pearl A
|
||||
sentinel-bootstrap on the EMA path causes the FIRST observation to
|
||||
replace the sentinel directly (per
|
||||
`pearl_first_observation_bootstrap.md`). So `target = first_obs`, ratio
|
||||
`current/target = 1.0`, confidence `clamp(ratio, 0, 1) = 1.0`,
|
||||
`floor = base × (1 − 1.0) = 0`. The cold-start mechanism is defeated
|
||||
because the EMA target tracks the running observation rather than a
|
||||
fixed threshold.
|
||||
|
||||
The principled fix per `feedback_isv_for_adaptive_bounds.md`:
|
||||
"sufficient" is defined in absolute terms via Invariant-1 numerical
|
||||
anchors. The 3 target slots become constructor-written constants
|
||||
(written ONCE in the trainer constructor, similar to existing
|
||||
`config.cql_alpha → ISV[CQL_ALPHA_INDEX=48]`), and the target-relative
|
||||
ratio `current_observation / fixed_anchor` then yields a self-normalised
|
||||
confidence in [0, 1] that rises monotonically as samples /
|
||||
divergence-stability / fold-time accumulate.
|
||||
|
||||
- `ISV[KELLY_SAMPLE_COUNT_TARGET_INDEX=333] = 100.0f` — minimum trade
|
||||
samples for statistical confidence to release the cap
|
||||
- `ISV[KELLY_DIVERGENCE_TARGET_INDEX=334] = 2.0f` — divergence ratio
|
||||
above which intent-vs-eval is considered diverged enough to keep the
|
||||
cap closed
|
||||
- `ISV[KELLY_TEMPORAL_TARGET_INDEX=335] = 5.0f` — minimum epochs in
|
||||
fold for temporal confidence to release the cap (safety net per
|
||||
Risk 3 in spec)
|
||||
|
||||
*Quirk 2 — divergence ratio explosion:* The
|
||||
`intent_eval_divergence_compute_kernel.cu` previously computed
|
||||
`divergence = max(intent_f, EPS_DIV) / max(eval_f, EPS_DIV)`. Before the
|
||||
first val-window populates `ISV[EVAL_DIST_F_INDEX=338]`, the slot is at
|
||||
Pearl A sentinel 0. With `intent_f ≈ 0.07` and `eval_f` at sentinel,
|
||||
flooring at `EPS_DIV=1e-6` produced a synthetic
|
||||
`divergence ≈ 0.07 / 1e-6 ≈ 70 000` (matched in smoke-test-wrwkz
|
||||
exactly). Algebraically the warmup-floor kernel still derived
|
||||
`behavioral_conf = clamp(1 − 70 000/2, 0, 1) = 0` (correct semantic
|
||||
"no behavioral signal yet"), but the synthetic 70 000 in HEALTH_DIAG
|
||||
masked the actual signal-flow and was easily misread as a runaway.
|
||||
|
||||
The principled fix: explicitly detect the sentinel via threshold
|
||||
`SENTINEL_THRESHOLD = 1e-5f` (well below any realisable `eval_f` after
|
||||
the Pearl A first-observation replacement) and emit
|
||||
`SENTINEL_DIVERGENCE = 1e6f`. Same `behavioral_conf = 0` outcome but
|
||||
with explicit provenance — HEALTH_DIAG now reads `divergence=1e6` until
|
||||
the first val window populates `eval_f`, at which point the real ratio
|
||||
takes over.
|
||||
|
||||
**Fix structure (atomic commit per `feedback_no_partial_refactor`):**
|
||||
|
||||
- DELETE 3 EMA target updater kernels (`.cu` files):
|
||||
- `kelly_sample_count_target_ema_kernel.cu`
|
||||
- `kelly_divergence_target_ema_kernel.cu`
|
||||
- `kelly_temporal_target_ema_kernel.cu`
|
||||
- DELETE 3 entries from `crates/ml/build.rs::kernels_with_common`
|
||||
- DELETE 3 cubin static byte arrays + 3 struct fields + 3 cubin loaders
|
||||
+ 3 fields in trainer construction tuple in `gpu_dqn_trainer.rs`
|
||||
- DELETE 3 launches in `launch_sp9_kelly_warmup_floor` (chain shrinks
|
||||
from 5 → 2 launches: q_var_mag_ema + main warmup-floor kernel)
|
||||
- DELETE 3 scratch slots (`SCRATCH_SP9_SAMPLE_COUNT_TARGET=262`,
|
||||
`SCRATCH_SP9_DIVERGENCE_TARGET=263`, `SCRATCH_SP9_TEMPORAL_TARGET=264`);
|
||||
`SP5_SCRATCH_TOTAL` 268 → 265; `SCRATCH_SP9_EVAL_DIST_BASE` slides
|
||||
265 → 262 (3 floats reclaimed)
|
||||
- ADD constructor-write of 3 Invariant-1 anchors (100.0, 2.0, 5.0) at
|
||||
`gpu_dqn_trainer.rs` constructor-time, immediately before the layout-
|
||||
fingerprint write — same pattern as existing
|
||||
`*sig_ptr.add(CQL_ALPHA_INDEX) = config.cql_alpha`
|
||||
- UPDATE 3 dispatch arms in `reset_named_state` to rewrite the same
|
||||
Invariant-1 anchors at fold boundary (NOT sentinel 0) — these are
|
||||
constants, not stateful EMAs; they must never reach 0 between folds
|
||||
- UPDATE 3 registry descriptions in `state_reset_registry.rs` to reflect
|
||||
the Invariant-1 anchor semantic + the smoke-test-wrwkz evidence
|
||||
- UPDATE `intent_eval_divergence_compute_kernel.cu` with
|
||||
`SENTINEL_THRESHOLD=1e-5f` / `SENTINEL_DIVERGENCE=1e6f` sentinel-detect
|
||||
branch
|
||||
- ISV slot layout UNCHANGED — the 3 target slots @ ISV[333..336)
|
||||
remain. Wiener-buffer linear span (`SP5_PRODUCER_COUNT=165`) UNCHANGED
|
||||
— the 3 wiener triples for the deleted producers become "reserved-
|
||||
unused" similar to Pearl 6's [525..543) carve-out (no
|
||||
`launch_apply_pearls` consumer fires for those slots, so the wiener
|
||||
state stays zero-initialised)
|
||||
|
||||
**Files touched:**
|
||||
|
||||
- `crates/ml/src/cuda_pipeline/kelly_sample_count_target_ema_kernel.cu` (DELETED)
|
||||
- `crates/ml/src/cuda_pipeline/kelly_divergence_target_ema_kernel.cu` (DELETED)
|
||||
- `crates/ml/src/cuda_pipeline/kelly_temporal_target_ema_kernel.cu` (DELETED)
|
||||
- `crates/ml/build.rs` — 3 entries removed; comment updated to "Four
|
||||
producer kernels"
|
||||
- `crates/ml/src/cuda_pipeline/intent_eval_divergence_compute_kernel.cu` —
|
||||
sentinel-detect branch for uninitialised `eval_f`
|
||||
- `crates/ml/src/cuda_pipeline/gpu_dqn_trainer.rs` — 3 cubin statics +
|
||||
3 struct fields + 3 cubin loaders + 3 tuple entries + 3 launches
|
||||
removed; constructor adds 3 Invariant-1 anchor writes;
|
||||
`SP5_SCRATCH_TOTAL` 268 → 265; `SCRATCH_SP9_EVAL_DIST_BASE` 265 → 262;
|
||||
3 obsolete `SCRATCH_SP9_*_TARGET` constants removed
|
||||
- `crates/ml/src/trainers/dqn/state_reset_registry.rs` — 3 descriptions
|
||||
updated to reflect Invariant-1 anchor semantic
|
||||
- `crates/ml/src/trainers/dqn/trainer/training_loop.rs` — 3 dispatch
|
||||
arms updated to write 100.0 / 2.0 / 5.0 (NOT 0.0)
|
||||
- `docs/dqn-wire-up-audit.md` (this entry)
|
||||
|
||||
**Verification:**
|
||||
|
||||
- `SQLX_OFFLINE=true cargo check -p ml` — clean (only pre-existing 18
|
||||
warnings; no new errors or warnings).
|
||||
- `SQLX_OFFLINE=true cargo test -p ml --lib state_reset` — all 4 tests
|
||||
pass (contract test `every_fold_and_soft_reset_entry_has_dispatch_arm`
|
||||
still covers all 9 SP9 entries; their dispatch arms now write
|
||||
Invariant-1 anchors instead of 0).
|
||||
- `SQLX_OFFLINE=true cargo test -p ml --lib sp5_isv_slots` — all 9
|
||||
tests pass (slot layout unchanged; producer count unchanged).
|
||||
|
||||
**Expected smoke-test signature post-fix (5-epoch L40S):**
|
||||
|
||||
- `floor`: starts at base_floor × 1.0 in epoch 1 (Pearl D Wiener-α
|
||||
smoothing of base_floor × (1 − combined_confidence) where confidence
|
||||
is genuinely low), decays as confidence accumulates over epochs
|
||||
- `divergence`: `1e6` (SENTINEL_DIVERGENCE) until the first val window
|
||||
populates `ISV[EVAL_DIST_F_INDEX=338]`; small (≤ 5) afterward
|
||||
- `conf [stat=X bhv=Y tmp=Z]`: `stat = sample_count/100` rises with
|
||||
trade count; `bhv` stays 0 until eval_dist matures; `tmp = epoch_idx/5`
|
||||
rises linearly with fold-time
|
||||
- combined_conf reaches 1.0 only when at least one axis genuinely
|
||||
matures — typically temporal axis at epoch 5 in fold 0
|
||||
|
||||
**Per-pearl provenance:**
|
||||
|
||||
- `feedback_isv_for_adaptive_bounds.md` — Invariant-1 numerical anchors
|
||||
are the survivable case for hardcoded constants (100.0, 2.0, 5.0
|
||||
define what "sufficient" means in absolute terms).
|
||||
- `pearl_controller_anchors_isv_driven.md` — caveat: not every
|
||||
anchor / target needs to be EMA-tracked. Cold-start exit thresholds
|
||||
are the canonical case where a fixed threshold is the principled
|
||||
design (matching test for "regime change") — the EMA anti-pattern
|
||||
(target = current_obs) was a misapplication.
|
||||
- `pearl_first_observation_bootstrap.md` — sentinel-detect branch in
|
||||
divergence kernel makes "before first observation" explicit; same
|
||||
semantic as Pearl A but with cleaner provenance in HEALTH_DIAG.
|
||||
- `feedback_no_partial_refactor.md` — atomic commit (kernel deletions
|
||||
+ Rust deletions + constructor-writes + dispatch-arm updates + audit
|
||||
entry, all in one commit).
|
||||
- `feedback_no_stubs.md` — Invariant-1 anchors are real values defining
|
||||
"sufficient" in absolute terms (100 trades / 2.0 ratio / 5 epochs),
|
||||
not stubs.
|
||||
- `feedback_no_cpu_compute_strict.md` — constructor-writes are GPU-side
|
||||
ISV writes via mapped-pinned `sig_ptr`; same path as all existing
|
||||
constructor-time ISV initialization (CQL_ALPHA, GAMMA_DIR, etc.).
|
||||
|
||||
Reference in New Issue
Block a user