Spec: docs/superpowers/specs/2026-05-28-state-conditional-q-synthesis.md
§3 Sub-phase 2.1.
Replaces the distributional Bellman bootstrap
(fused_select_and_project_bellman) with the Dueling-DQN scalar
baseline target
y(s_t, a_t) = r_t + γⁿ × V_target(s_{t+1})
advantage_scalar = y − V(s_t)
projected via two-bin linear interpolation onto the action's atom
support. The Q-head's atom logits are now interpreted as the
advantage distribution A_dist(s, a); the CE loss between
softmax(A_dist(s_t, a_t)) and the projected target trains A to
model deviation from the scalar V baseline.
Changes:
- NEW crates/ml-alpha/cuda/bellman_scalar_target_project.cu —
device-side scalar Bellman target + per-action atom projection.
Grid (B, 1, 1), Block (Q_N_ATOMS=21, 1, 1). No atomicAdd, no
cross-thread accumulation: scalar advantage produces a delta on
exactly two atoms so each thread writes its own atom value.
- crates/ml-alpha/build.rs registers the new kernel.
- crates/ml-alpha/src/rl/dqn.rs:
* Loads bellman_scalar_target_project from its own cubin.
* Adds DqnHead::bellman_scalar_target_project wrapper.
* Adds host-side composed_q_value(v_scalar, a_logits,
atom_supports) -> Vec<f32> implementing the Dueling
composition Q = V + A - mean_a(A) for test oracles.
- crates/ml-alpha/src/trainer/integrated.rs:
* Allocates sampled_v_curr_d + sampled_v_target_tp1_d for
V evaluated on PER-sampled hidden states.
* step_synthetic_body Bellman build: replaces forward_target +
fused_select_and_project_bellman with V forwards on sampled
buffers + the new scalar Bellman projection (single source of
truth — same swap applied to dqn_replay_step for k>0 path).
* Test accessors isv_host_slice_mut_for_test,
isv_dev_ptr_for_test, dqn_head_ref so kernel-oracle unit
tests can drive the new kernel without a full mega-graph.
- NEW crates/ml-alpha/tests/q_projection_unit.rs — pre-impl oracle:
projects y=0 to center atom, y=V_MAX to last, y=V_MIN to first,
arbitrary in-range cases recover EV within Δz/2, verifies V_t
enters as a subtraction. Plus pure host-side check of
composed_q_value.
All Phase 2.1 unit tests pass (5 CUDA --ignored + 2 host-side).
Smoke status: BLOCKED on pre-existing NaN abort at step 4 affecting
HEAD (c52282fb4 + f11bab542 + 7e38e46e6) — l_pi=NaN, l_aux=NaN at
step 4 reproduce IDENTICALLY with and without Phase 2.1 changes.
Per feedback_stop_on_anomaly, the pre-existing NaN must be diagnosed
before the Phase 2.1 smoke gate (l_v<2.0, l_q<3.0,
action_entropy>1.5 for 500 steps) can be validated. The Phase 2.1
changes themselves are structurally sound — unit tests confirm
correct projection and the smoke loss trajectories at steps 0-3 are
within ~1% of HEAD (l_q 2.9484 vs 2.9926, l_v 6.3293 vs 6.3291).
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
db4d9a16ff