fix(modelfamily): Generated A-M residual model bugs — BornRule loss + CSPDarknet activations#1719

Merged

ooples merged 62 commits into

fix/generated-am-modelbugs

Jul 1, 2026

Commits on Jun 28, 2026

fix(loss): mark BornRuleMseLoss as non-standard gradient sign

franklinic
and
claude
committed
fix(cv): expose CSPDarknet backbone per-stage activations for GetNamedLayerActivations

franklinic
and
claude
committed
fix(midas): resolve lazy conv input depth + ScaleInvariantDepthLoss rank mismatch

ooples
and
claude
committed
fix(causal): recover edges + scale-invariance for AVICI/AmortizedCD/DAG-GNN

ooples
and
claude
committed
fix(meter): embed vision input to VisionDim (real dim-mismatch bug, not a timeout)

ooples
and
claude
committed
test(ci): tag foundation-scale A-M timeout models HeavyTimeout

ooples
and
claude
committed
fix(graph): add missing GraphNetwork category to LinkPredictionModel

ooples
and
claude
committed
fix(rl): default GradientBandit NumArms to 10 (was 0 -> empty agent)

ooples
and
claude
committed
fix(diffusion): #1715 streaming param-IO + chunk-API order/OOM fixes

franklinic
and
claude
committed
fix(rl): train on partial replay batch + gradient-bandit baseline/greedy-eval

ooples
and
claude
committed
fix(rl): keep only paper-faithful bandit changes; revert training band-aids

ooples
and
claude
committed
test(rl): make DifferentStates/Training invariants paper-faithful

ooples
and
claude
committed
fix(rl): default Thompson/EpsilonGreedy NumArms to 10 + greedy Thompson eval

ooples
and
claude
committed
fix(modelfamily): address pr review comments — dead lagrangian state, scale-invariant dag orientation, cspdarknet activation key order (#1719)

ooples
and
claude
committed
fix(rl): default LSPI/LSTD FeatureSize to 4 (was 0 -> empty weights)

ooples
and
claude
committed
fix(rl): count param-vector growth as change; opt out QMIX/SARSA-lambda probe

ooples
and
claude
committed
test(rl): reliable state-conditionality via state battery + deterministic training

ooples
and
claude
committed
fix(rl): TRPO Clone deep-copies policy; QMIX train guard + adapter opt-out

ooples
and
claude
committed

Commits on Jun 29, 2026

Commits on Jun 30, 2026

Commits on Jul 1, 2026