Wave 21: adversarial-review fixes — all 9 verified findings closed

Wave 21: Stage-0 dataset pipeline — swesmith engine, rollout harness, gates, contract

Wave 20: Tier-0 fidelity fixes — k1-in-reward KL + Composer-2 behavior rewards

Wave 20: close F4 s3:// gap — live-S3 DiLoCo allreduce smoke (AWS_SMOKE-gated)

R14: harden live-Docker tests against transient daemon contention

Wave 3 cleanup: close deferred-LOW review items R5/R6/R11

Wave 3: close the HIGH review findings (kill-switch wiring, HeldoutSplit, EKS entrypoint bug)

Wave 2: 4 new modules (kill-switch, EKS/SageMaker executors, DockerSandbox) + B4/B7 completion

Wave 1: fix 8 failing tests + unblock Docker E2E + dep/doc debt

feat(trainer): policy-optimization objective MENU (ADR-014)

fix(phase-8): close all 5 cross-family final-verify findings + regression tests

feat(b4-gpu+b6): GPU train-proof on A10G + docker-gated substrate E2E test

feat(wave-b): ADR-013 LMA integration + B4 end-to-end SDPO-fires proof + doc refresh

feat(wave-a): close ADR-011 (SDPO alignment indices) + ADR-012 (review findings)

review+fix: cross-family adversarial ADR review (owed item) + remediation

feat(datagen): ADR-010 FeatureDeletionEnv synthetic-data subsystem; accepted

feat(hints): ADR-009 layered HintGenerator; accepted

feat(trainer): ADR-008 Dr.GRPO config + SDPO strict-alignment guard

Wave 21c: verify PRIME-RL adapter parity against upstream source (byte-for-byte)

Wave 21b: skip zero-signal SDPO on empty-recovery error turns + real-trace validation

Wave 21: close both Wave 20 debt items — chat-template alignment + structural is_error

Wave 20: ModalSpawnExecutor — finish the Modal-backed serverless executor

Wave 19: production-grade SDPO via ComposerDataCollator + adapter + collator fixes

Wave 18: 14 backlog items closed + 3-reviewer cross-family review

Wave 17: close all 5 audit FLAGs + SDPO context alignment + serverless re-exports

Wave 16: install ergonomics + gradient evidence + SDPO end-to-end example

Wave 15: 4-angle multi-model self-critique caught 2 math BLOCKERs in primary loss kernels; fixed against upstream byte-for-byte + GSM8K example + ergonomics

Wave 14: close every Wave 13 review finding + 4 documentation files; Wave 14b: real PRIME-RL parity + multi-process DiLoCo convergence

Wave 13: serverless DiLoCo + replaysim normalization + 3 distillation losses + PRIME-RL + Monarch

Wave 12: close V1-V8 brief — GPU smoke, SDPO firing, real-trace e2e

Wave 11: cross-model adversarial review + honest down-revision

Wave 10 — packaging: composer_replication is now pip-installable