DaoyuanLi2816 · Sep 8, 2026 · Sep 8, 2026 · Sep 8, 2026
diff --git a/.github/dependabot.yml b/.github/dependabot.yml
@@ -5,7 +5,9 @@ updates:
     directory: "/"
     schedule:
       interval: monthly
-    open-pull-requests-limit: 5
+    # Dependency updates are reviewed and committed on first-party maintainer
+    # branches so automated authors never enter the contributor graph.
+    open-pull-requests-limit: 0
     groups:
       dev-tooling:
         patterns: ["ruff", "mypy", "pytest*", "hypothesis", "build", "twine", "types-*"]
@@ -22,4 +24,4 @@ updates:
     directory: "/"
     schedule:
       interval: monthly
-    open-pull-requests-limit: 3
+    open-pull-requests-limit: 0
diff --git a/.github/workflows/build.yml b/.github/workflows/build.yml
@@ -48,6 +48,7 @@ jobs:
               "environments/known-good-rtx4080-wsl2-cu130.json",
               "environments/known-good-rtx4080-wsl2-cu130.txt",
               "scripts/run_v011_profile_qualification.py",
+              "scripts/run_v012_rl_qualification.py",
               "LICENSE",
               "README.md",
           ):

diff --git a/.github/workflows/docs.yml b/.github/workflows/docs.yml
@@ -34,6 +34,7 @@ jobs:
           python scripts/publish_alignment_lab_artifacts.py --check
           python scripts/publish_verl_bridge_diagrams.py --check
           python scripts/publish_verl_opd_compatibility.py --check
+          python scripts/publish_verl_rl_v09_compatibility.py --check
           python scripts/publish_hardware_record_schema.py --check
           python scripts/publish_gpu_qualification_schema.py --check
           python scripts/check_known_good_environment.py

diff --git a/.github/workflows/gpu.yml b/.github/workflows/gpu.yml
@@ -6,7 +6,7 @@ on:
   workflow_dispatch:
     inputs:
       qualification_level:
-        description: "Release smoke, or the complete v0.11 WSL2 qualification"
+        description: "Release smoke, or the complete v0.12 WSL2 qualification"
         type: choice
         required: true
         default: smoke
@@ -186,6 +186,13 @@ jobs:
             --output "$full_dir/v011-profiles.json" \
             --work "$full_work/v011-profiles" \
             "${offline[@]}"
+          "$QUALIFICATION_PYTHON" scripts/run_v012_rl_qualification.py \
+            --commit "$GITHUB_SHA" \
+            --wheel "$QUALIFICATION_WHEEL" \
+            --wheel-sha256 "$QUALIFICATION_WHEEL_SHA256" \
+            --output "$full_dir/v012-rl.json" \
+            --work "$full_work/v012-rl" \
+            "${offline[@]}"
           test -z "$(git status --porcelain)"
           "$QUALIFICATION_PYTHON" scripts/benchmark_rollout_runtime_v2_selected.py \
             --backend hf_cached \
@@ -207,7 +214,8 @@ jobs:
             --v011-profiles "$full_dir/v011-profiles.json" \
             --hf-cached-runtime "$full_dir/hf-cached-runtime.json" \
             --vllm-runtime "$full_dir/vllm-runtime.json" \
-            --hf-reference benchmarks/results/rollout-runtime-v2-hf-reference.json
+            --hf-reference benchmarks/results/rollout-runtime-v2-hf-reference.json \
+            --v012-rl "$full_dir/v012-rl.json"
       - name: Upload full qualification records
         uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a # v7.0.1
         if: ${{ success() && inputs.qualification_level == 'full' }}

diff --git a/.github/workflows/verl-bridge.yml b/.github/workflows/verl-bridge.yml
@@ -47,7 +47,7 @@ jobs:
           changed=$(git diff --name-only "origin/$BASE_REF"...HEAD)
           echo "changed files:"
           echo "$changed"
-          if echo "$changed" | grep -Eq '^(\.github/workflows/verl-bridge\.yml|pyproject\.toml|scripts/[^/]*verl_bridge[^/]*|src/miniverl/bridge/.*|src/miniverl/losses/verl_(topk|pg)\.py|tests/.*verl_(bridge|opd|pg).*|tests/conformance/test_verl_v08_(loss|pg_k1)\.py)$'; then
+          if echo "$changed" | grep -Eq '^(\.github/workflows/verl-bridge\.yml|pyproject\.toml|scripts/[^/]*verl_bridge[^/]*|src/miniverl/(algorithms|bridge)/.*|src/miniverl/losses/verl_(topk|pg)\.py|tests/.*verl_(bridge|opd|pg).*|tests/conformance/test_verl_v0(8_(loss|pg_k1)|9_rl)\.py)$'; then
             echo "bridge=true" >> "$GITHUB_OUTPUT"
           else
             echo "bridge=false" >> "$GITHUB_OUTPUT"
@@ -77,6 +77,14 @@ jobs:
         run: python -m pytest -q tests/conformance/test_verl_v08_loss.py -m verl_conformance
       - name: Compare sampled k1 policy-loss values, gradients and optimizer step
         run: python -m pytest -q tests/conformance/test_verl_v08_pg_k1.py -m verl_conformance
+      - name: Checkout the pinned verl v0.9.0 algorithm source
+        run: >-
+          git clone --depth 1 --branch v0.9.0
+          https://github.com/volcengine/verl.git _verl-v0.9.0
+      - name: Compare GRPO, Dr.GRPO, RLOO, REINFORCE++, GAE, policy and value math
+        env:
+          MINIVERL_VERL_V09_SOURCE: _verl-v0.9.0
+        run: python -m pytest -q tests/conformance/test_verl_v09_rl.py -m verl_conformance
       - name: Materialize and recheck a tiny launchable pure-OPD bundle
         run: >-
           python -m pytest -q tests/conformance/test_verl_v08_materialize.py

diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -6,6 +6,40 @@ All notable changes to miniVERL are recorded here. The format follows
 
 ## [Unreleased]
 
+## [0.12.0] - 2026-09-07
+
+### Single-GPU verl RL
+
+- Added the versioned `verl-rl-v0.9-single-gpu-v1` compiler for a resolved
+  official verl v0.9 subset. It classifies every source field, lowers resource
+  placement to one process/device, validates runnable recipes, and publishes
+  the recipe/report pair transactionally.
+- Added critic-free GRPO, Dr.GRPO, RLOO and REINFORCE++ execution with complete
+  prompt groups, behavior-policy log-probabilities, vanilla dual-clipped policy
+  loss, outcome rewards and optional fixed reference-policy reward KL.
+- Added upstream-pinned conformance for estimator, reference-KL, policy-loss,
+  GAE and clipped-value mathematics against verl v0.9.0 at commit
+  `483b8a009ba3a97563edee3a19887e4862b8094a`.
+- Unified prompt and agent/tool trajectories through schema-v3 group/sample
+  identity, generated-token-only masks and recorded environment verifier
+  outcomes. Added exact interrupted/resume coverage for the RL path.
+- Added typed batch reward providers and weighted composition for trusted local
+  Python use, plus an explicit deterministic target-length provider, while
+  keeping YAML free of arbitrary imports.
+- Extended the exact-wheel RTX 4080 release chain with a compiled v0.9 GRPO
+  workload that requires within-group reward variance, nonzero advantages, two
+  parameter-changing updates and a bounded-memory evidence record.
+
+### Correctness and product surface
+
+- Fixed full-vocabulary bucketed teacher entropy so `K == V` preserves valid
+  zero-probability terms instead of creating non-finite values (fixes #89).
+- Made the RL compatibility compiler usable from the Torch-free core install
+  and generated a byte-checked v0.9 field report from the shipped example.
+- Rebuilt the README and documentation landing path around verl experiment
+  semantics compiled into a single-GPU plan, with RL and OPD capabilities shown
+  together and detailed boundaries consolidated in compatibility/limitations.
+
 ## [0.11.0] - 2026-08-29
 
 ### Rollout Runtime v2 baseline
@@ -1143,7 +1177,8 @@ Same-tokenizer only; one trajectory per forward pass; `swap` unavailable for
 quantized models; only Qwen3 and Qwen2 architectures tested; single-seed GPU
 results. The full list is in `docs/limitations.md`.
 
-[Unreleased]: https://github.com/DaoyuanLi2816/mini-verl/compare/v0.11.0...HEAD
+[Unreleased]: https://github.com/DaoyuanLi2816/mini-verl/compare/v0.12.0...HEAD
+[0.12.0]: https://github.com/DaoyuanLi2816/mini-verl/compare/v0.11.0...v0.12.0
 [0.11.0]: https://github.com/DaoyuanLi2816/mini-verl/compare/v0.10.1...v0.11.0
 [0.10.1]: https://github.com/DaoyuanLi2816/mini-verl/compare/v0.10.0...v0.10.1
 [0.10.0]: https://github.com/DaoyuanLi2816/mini-verl/compare/v0.9.1...v0.10.0

diff --git a/CITATION.cff b/CITATION.cff
@@ -2,8 +2,8 @@ cff-version: 1.2.0
 title: "miniVERL: A bounded single-GPU runtime for verl-style OPD"
 message: "If you use miniVERL in your work, please cite it as below."
 type: software
-version: 0.11.0
-date-released: 2026-08-29
+version: 0.12.0
+date-released: 2026-09-07
 license: Apache-2.0
 repository-code: "https://github.com/DaoyuanLi2816/mini-verl"
 url: "https://github.com/DaoyuanLi2816/mini-verl"

diff --git a/PROJECT_STATE.md b/PROJECT_STATE.md
@@ -4,25 +4,32 @@ Current maintainer handoff for **miniVERL** (`mini-verl` package, `miniverl`
 CLI). `release-state.yaml` is the canonical version source; this page indexes
 current product and evidence state rather than repeating release history.
 
-Last updated: 2026-08-29.
+Last updated: 2026-09-07.
 
-Canonical release state: stable `v0.11.0` (`479695c4980150bc6f7a51bddec2b90c08c8bb98`), development `0.11.1.dev0`.
+Canonical release state: releasing `v0.12.0`.
 
 ## Release state
 
-- Stable release: `v0.11.0` at `479695c4980150bc6f7a51bddec2b90c08c8bb98`.
-- Development release: `0.11.1.dev0`.
+- Release candidate: `v0.12.0`; publication is gated on exact-SHA RTX 4080 qualification.
+- Previous stable release: `v0.11.0` at `479695c4980150bc6f7a51bddec2b90c08c8bb98`.
 - Stable docs: <https://daoyuanli2816.github.io/mini-verl/>.
 - Development docs: <https://daoyuanli2816.github.io/mini-verl/dev/>.
 - Historical build log: [v0.1-v0.9 archive](docs/history/project-state-v0.1-v0.9.md).
 
 ## Current product boundary
 
-The primary product is `verl-opd-v0.8-single-gpu-v1`: a typed, resolved subset
-of official verl `v0.8.0` at commit
-`7aed6b230776f963fa09509c10d9c3a767d1102c`. It runs one actor, one teacher,
-`n=1`, reward-free direct GKD `forward_kl_topk` and token-mean reduction in one
-local process on one NVIDIA CUDA GPU.
+Release `0.12.0` has a closed typed profile compiler and adds
+`verl-rl-v0.9-single-gpu-v1`, a typed resolved
+subset of official verl `v0.9.0` at commit
+`483b8a009ba3a97563edee3a19887e4862b8094a`. It compiles GRPO, Dr.GRPO, RLOO
+and REINFORCE++ experiments into grouped one-GPU actor/reward/reference phases,
+with behavior-policy identity, fixed reference reward KL, transactional resume
+and a machine-readable field report.
+
+The established `verl-opd-v0.8-single-gpu-v1` family remains the teacher-based
+distillation path. It targets official verl `v0.8.0` at
+`7aed6b230776f963fa09509c10d9c3a767d1102c` and provides direct GKD,
+sampled-k1, grouped and rewarded profiles.
 
 The stable v0.9.1 repair makes dtype, quantization, attention, student
 adapter input, logical/physical batch limits and placement legality explicit.
@@ -31,10 +38,11 @@ Executable compatibility claims are mutation-tested and recorded in
 unknown-size quantized roles require proof instead of receiving an executable
 plan.
 
-Development `0.11.1.dev0` has a closed typed profile registry and torch-free
-compatibility introspection. Profile-scoped plans, caches, checkpoints and
-exports bind an independent identity. The direct-GKD and sampled-k1 vanilla
-policy-loss profiles both have pinned conformance and measured RTX 4080 paths.
+Profile-scoped recipes, plans, caches, checkpoints and exports bind an
+independent identity. The v0.9 compiler is Torch-free; the RL math and runtime
+remain inside `[train]`. The new estimator, policy, value and KL primitives are
+compared against the exact upstream source. The direct-GKD and sampled-k1
+profiles retain their pinned conformance and measured RTX 4080 paths.
 Portable hardware records preserve measured, estimated and unknown states;
 schema-valid community submissions remain unreviewed until maintainer validation.
 The release chain builds one hosted-runner candidate and binds the maintainer's
@@ -44,8 +52,11 @@ Releases use a canonical, explicit evidence layout with four principal JSON
 records and one deterministic subordinate-evidence archive; historical v0.10.1
 asset names remain immutable.
 
-Arbitrary verl YAML, other policy-gradient modes, rewards, PPO/GRPO, Ray, FSDP,
-Megatron, multi-GPU and distributed execution remain unsupported. The legacy
+PPO/GAE execution remains unimplemented because the runtime has no trainable
+critic model/optimizer/checkpoint lifecycle; its mathematical primitives are
+conformance-tested. Ray, FSDP/FSDP2, Megatron, multi-GPU and multi-node
+execution are distributed-only. Arbitrary unresolved verl YAML and unknown
+algorithm-changing fields remain outside the versioned compiler. The legacy
 environment/PPO artifact bridge is migration-only.
 
 ## Current evidence
@@ -66,6 +77,7 @@ calculator schema-v2 source stays at SHA-256
 ## Maintainer entry points
 
 - [Current local runtime](docs/verl-opd-runtime.md)
+- [Single-GPU verl RL](docs/verl-rl-runtime.md)
 - [Current scale-out contract](docs/verl-opd-scaleout.md)
 - [Compatibility policy](docs/compatibility.md)
 - [Measured workload](docs/verl-opd-reference-workload.md)