gitadmin/foxhunt
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

feat(plan): conviction drift + regime shift in observation (P11, P12)

Browse Source 
conviction_drift = current_conviction - entry_conviction: model sees
its thesis weakening. regime_shift = |stability_now - stability_entry|:
model detects regime change invalidating the plan. Entry regime stored
in ps[29]. Both features zero when no plan active.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
This commit is contained in:
jgrusewski
2026-04-17 09:17:44 +02:00
parent 70fca46227
commit 7651f6f6cc
2 changed files with 26 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

22
crates/ml/src/cuda_pipeline/experience_kernels.cu
View File

@@ -392,7 +392,9 @@ extern "C" __global__ void experience_state_gather(
/* #11 Time-reversal: reverse bar indexing for selected episodes.
* If > 0, episodes with index % time_reversal_mod == 0 are played backwards.
* 5 = 20% of episodes reversed, 3 = 33%, 0 = disabled. */
int time_reversal_mod
int time_reversal_mod,
const float* __restrict__ plan_params_ptr, /* [N, 6] trade plan params. NULL = no plan. */
const float* __restrict__ isv_signals_ptr /* [12] pinned device-mapped ISV signals. NULL = static. */
) {
int i = blockIdx.x * blockDim.x + threadIdx.x;
if (i >= N) return;
@@ -572,7 +574,7 @@ extern "C" __global__ void experience_state_gather(
: 0.0f;
int portfolio_base = market_dim;
if (portfolio_base + 11 < state_dim) {
if (portfolio_base + 13 < state_dim) {
out[portfolio_base + 0] = f_position; /* raw position */
out[portfolio_base + 1] = f_unrealized_pnl / f_equity; /* trade P&L signal */
out[portfolio_base + 2] = f_drawdown; /* risk: how deep are we? */
@@ -597,6 +599,20 @@ extern "C" __global__ void experience_state_gather(
? fminf(-f_unrealized_pnl / (plan_stop * f_equity + 1e-6f), 2.0f) /* pnl vs stop */
: 0.0f;
out[portfolio_base + 11] = plan_conv; /* conviction at entry [0, 1] */
/* P11: Conviction drift — current conviction minus entry conviction.
* Negative drift = model's thesis is weakening → exit signal.
* Only meaningful when a plan is active (ps[23] > 0.5). */
out[portfolio_base + 12] = (plan_params_ptr != NULL && ps[23] > 0.5f)
? (plan_params_ptr[i * 6 + 4] - ps[27]) /* conviction drift */
: 0.0f;
/* P12: Regime shift since entry — |regime_stability_now - regime_stability_at_entry|.
* Large shift = regime changed mid-trade, plan may be invalid.
* Entry regime stored in ps[29] during plan activation. */
out[portfolio_base + 13] = (isv_signals_ptr != NULL && ps[23] > 0.5f)
? fabsf(isv_signals_ptr[11] - ps[29]) /* regime shift */
: 0.0f;
}
/* -- Multi-timeframe features: [market_dim+8 .. market_dim+8+16) --
@@ -1481,6 +1497,8 @@ extern "C" __global__ void experience_env_step(
ps[26] = pp[3]; /* scale_aggression */
ps[27] = pp[4]; /* conviction */
ps[28] = pp[5]; /* asymmetry */
/* Store entry regime_stability for drift detection (P12). */
ps[29] = (isv_signals_ptr != NULL) ? isv_signals_ptr[11] : 1.0f;
/* Apply conviction to position size — scale by readiness for smooth ramp */
position *= fmaxf(ps[27] * readiness, 0.1f);
}

6
crates/ml/src/cuda_pipeline/gpu_experience_collector.rs
View File

@@ -1919,6 +1919,10 @@ impl GpuExperienceCollector {
.map(|b| b.device_ptr(&self.stream).0)
.unwrap_or(0u64);
// Plan params and ISV signals pointers for conviction drift + regime shift features
let plan_ptr = self.plan_params_dev_ptr;
let isv_ptr = self.isv_signals_dev_ptr;
self.stream
.launch_builder(&self.state_gather_kernel)
.arg(market_features_buf)
@@ -1935,6 +1939,8 @@ impl GpuExperienceCollector {
.arg(&vol_norm) // #13 vol normalization
.arg(&mirror_i32) // #10 mirror universe
.arg(&config.time_reversal_mod) // #11 time-reversal
.arg(&plan_ptr) // plan_params_ptr (NULL=no plan)
.arg(&isv_ptr) // isv_signals_ptr (NULL=static)
.launch(launch_cfg)
.map_err(|e| MLError::ModelError(format!(
"experience_state_gather t={t}: {e}"