Cluster smoke `alpha-rl-tcr5r` confirmed that the grad-norm-driven
multiplicative LR controller — even with the per-step rate cap +
per-head TARGET_GRAD_NORM fixes — could not avoid closed-loop
oscillation when stacked on Adam:
step 5000 : ALL lrs at MAX (1e-2)
step 15000: ALL lrs at MIN (1e-5)
step 25000: lr_pi MAX again
...
Result: l_pi max = 1.28e19, l_v max = 701k, l_total mean = 1.15e14.
Q-head benefited (l_q mean 3.24) but π and V destabilised
catastrophically.
## Why the prior design was fundamentally broken
The grad-norm signal that drives the LR controller is itself
*produced* by the LR being applied (via Adam → weights → grads →
norms). When the LR controller reduces lr_pi because grad-norm
spiked, the next-step grad-norm shrinks → controller raises LR →
grad-norm spikes again. Classic two-loop instability when stacked
on Adam (which already does per-parameter LR adaptation via its 2nd
moment). No amount of per-step rate capping breaks the cycle; it
just slows it.
## New design: ReduceLROnPlateau-style monotone decay
The controller now:
1. Maintains a SLOW loss EMA per head (α = 0.05, half-life ≈ 13
steps — well below the canonical Wiener 0.4 floor used by the
per-step EMAs because plateau detection needs smoothness, not
responsiveness).
2. Tracks `best_loss_ema` per head — lowest EMA value ever seen.
3. Per step: if current EMA improves on best by ≥ 1%
(IMPROVEMENT_THRESHOLD = 0.99), update best + reset counter.
Otherwise increment counter.
4. When counter exceeds PLATEAU_PATIENCE (1000 steps ≈ 7 sec at
145 steps/sec), halve LR (DECAY_FACTOR = 0.5), reset counter,
keep best.
5. LR can ONLY decrease — never grows. Bottoms out at LR_MIN = 1e-5.
Closed-loop oscillation is impossible by construction: monotone
decay can't drive LR up in response to its own induced gradient
changes. Worst case: LR decays to MIN and stays there (interpretable
as "model has stopped learning at any LR scale" — meaningful signal,
not a control failure).
## State storage
9 new ISV slots (3 per head — Q, π, V):
* RL_LR_Q_LOSS_EMA_INDEX = 427
* RL_LR_Q_BEST_LOSS_INDEX = 428
* RL_LR_Q_STEPS_SINCE_BEST_INDEX = 429
* RL_LR_PI_LOSS_EMA_INDEX = 430
* RL_LR_PI_BEST_LOSS_INDEX = 431
* RL_LR_PI_STEPS_SINCE_BEST_INDEX = 432
* RL_LR_V_LOSS_EMA_INDEX = 433
* RL_LR_V_BEST_LOSS_INDEX = 434
* RL_LR_V_STEPS_SINCE_BEST_INDEX = 435
* RL_SLOTS_END = 436 (was 427)
Counters stored as f32 — mantissa precision to 16M is well beyond
any plausible patience threshold.
## Kernel signature change
```cuda
extern "C" __global__ void rl_lr_controller(
float* isv,
float observed_loss_bce, // unused (perception-owned)
float observed_loss_q, // host scalar from prior step's Q backward
float observed_loss_pi, // host scalar from prior step's PPO surrogate
float observed_loss_v, // host scalar from prior step's V backward
float observed_loss_aux, // unused
int q_loss_ema_slot, int q_best_slot, int q_counter_slot,
int pi_loss_ema_slot, int pi_best_slot, int pi_counter_slot,
int v_loss_ema_slot, int v_best_slot, int v_counter_slot
);
```
Grad-norm EMA producers (commit 383b1ad83) remain wired — they're
still useful diagnostics in the JSONL, just not consumed by the
LR controller anymore.
## Trainer wiring
New trainer fields `last_q_loss` + `last_v_loss` mirror per-step
loss scalars (same pattern as the existing `last_pi_loss` from
PPO surrogate forward). Populated at the end of step_synthetic's
backward chain; consumed at the start of NEXT step_synthetic's
launch_rl_lr_controller call. One-step lag is acceptable —
plateau detection operates on 1000-step windows so a 1-step shift
in observations is negligible.
## Verified gates (local sm_86)
G1, G3, G4, G6, smoke: all ✅
## Expected effect on next 50k smoke
* lr_q starts at 1e-3, decays monotonically toward 1e-5 if l_q
plateaus.
* lr_pi same — but π loss is much noisier, so plateau detection
may fire more often → faster decay.
* lr_v starts at 1e-3, decays as l_v approaches its asymptote
(V regression of ~0 for sparse rewards).
* NO l_pi explosions (controller can't drive LR up).
* Final losses should be similar to or better than fixed-LR's
baseline (mean 2.97 for l_q on `nqd68`; this design's monotone
decay should produce stable equilibrium at some LR ≤ 1e-3).
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
042de99e67