DaoyuanLi2816 · Sep 9, 2026 · Sep 9, 2026 · Sep 9, 2026 · Sep 9, 2026 · Sep 9, 2026
diff --git a/.github/workflows/gpu.yml b/.github/workflows/gpu.yml
@@ -200,6 +200,13 @@ jobs:
             --output "$full_dir/v013-ppo.json" \
             --work "$full_work/v013-ppo" \
             "${offline[@]}"
+          "$QUALIFICATION_PYTHON" scripts/run_v014_product_qualification.py \
+            --commit "$GITHUB_SHA" \
+            --wheel "$QUALIFICATION_WHEEL" \
+            --wheel-sha256 "$QUALIFICATION_WHEEL_SHA256" \
+            --output "$full_dir/v014-product.json" \
+            --work "$full_work/v014-product" \
+            "${offline[@]}"
           test -z "$(git status --porcelain)"
           "$QUALIFICATION_PYTHON" scripts/benchmark_rollout_runtime_v2_selected.py \
             --backend hf_cached \
10BC8
@@ -223,7 +230,8 @@ jobs:
             --vllm-runtime "$full_dir/vllm-runtime.json" \
             --hf-reference benchmarks/results/rollout-runtime-v2-hf-reference.json \
             --v012-rl "$full_dir/v012-rl.json" \
-            --v013-ppo "$full_dir/v013-ppo.json"
+            --v013-ppo "$full_dir/v013-ppo.json" \
+            --v014-product "$full_dir/v014-product.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/(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)$'; then
+          if echo "$changed" | grep -Eq '^(\.github/workflows/verl-bridge\.yml|pyproject\.toml|scripts/[^/]*verl_bridge[^/]*|scripts/publish_(verl_compatibility_corpus|rl_product_examples)\.py|benchmarks/compatibility/.*|src/miniverl/(algorithms|bridge|resources)/.*|src/miniverl/losses/verl_(topk|pg)\.py|tests/.*verl_(bridge|opd|pg|rl|compatibility).*|tests/conformance/test_verl_v0(8_(loss|pg_k1)|9_rl)\.py)$'; then
             echo "bridge=true" >> "$GITHUB_OUTPUT"
           else
             echo "bridge=false" >> "$GITHUB_OUTPUT"
@@ -68,7 +68,7 @@ jobs:
       - name: Install miniVERL bridge environment
         run: >-
           python -m pip install --upgrade pip &&
-          python -m pip install ".[bridge,train]" "hydra-core>=1.3,<2" pytest
+          python -m pip install ".[bridge,train]" "hydra-core==1.3.2" pytest
       - name: Install the exact official verl source without its distributed stack
         run: >-
           python -m pip install --no-deps
@@ -85,7 +85,12 @@ jobs:
         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
+      - name: Regenerate the real upstream compatibility corpus
+        run: >-
+          python scripts/publish_verl_compatibility_corpus.py --upstream _verl-v0.9.0 --check
+      - name: Verify packaged PPO and GRPO workflows
+        run: python scripts/publish_rl_product_examples.py --check
+      - name: Materialize and recheck the pure-OPD bundle
         run: >-
           python -m pytest -q tests/conformance/test_verl_v08_materialize.py
           -m verl_conformance

diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -6,6 +6,34 @@ All notable changes to miniVERL are recorded here. The format follows
 
 ## [Unreleased]
 
+## [0.14.0] - 2026-09-08
+
+### Installed local RL workflows
+
+- Package upstream-derived PPO and GRPO examples with exact source/adaptation
+  ledgers, import reports, native recipes and a complete inspect/resume/handoff
+  walkthrough. No checkout is required by their user commands.
+- Add run-directory inspection and RL reports with actor/critic counters,
+  rewards, advantages, losses, memory, local lowerings and checkpoint history.
+- Gate releases on both installed-candidate workflows, exact tensor replay and
+  checked v0.9 handoff, alongside the existing full RTX 4080 qualification.
+
+### Compatibility and recovery
+
+- Add the immutable v3 compiler profile: upstream prompt minibatches expand by
+  rollout count; physical offload hints are recorded; unimplemented reductions
+  and repeated non-PPO epochs reject explicitly. Published v1/v2 rules remain
+  unchanged.
+- Add six upstream-owned scripts and seven Hydra-resolved configuration cases
+  with CI-gated field accounting, pinned source hashes and precise rejections.
+- Bind PPO old values and RL behavior log-probabilities to their source roles;
+  restore RNG on critic OOM retry and reject nonfinite critic gradients.
+- Bind RL checkpoint log prefixes, archive post-checkpoint tails on resume and
+  reject corrupt/inconsistent inspection artifacts. Add interruption coverage
+  across rollout, reward, old-value, actor and critic phases.
+- Foreground the current local RL workflow in English/Chinese documentation;
+  retain frozen scientific results and evidence-scoped hardware claims.
+
 ## [0.13.0] - 2026-09-08
 
 ### PPO and full local RL
@@ -1206,7 +1234,7 @@ 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/v0.13.0...HEAD
+[Unreleased]: https://github.com/DaoyuanLi2816/mini-verl/compare/v0.14.0...HEAD
 [0.13.0]: https://github.com/DaoyuanLi2816/mini-verl/compare/v0.12.0...v0.13.0
 [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

diff --git a/CITATION.cff b/CITATION.cff
@@ -1,46 +1,24 @@
 cff-version: 1.2.0
-title: "miniVERL: A bounded single-GPU runtime for verl-style OPD"
+title: "miniVERL: Run verl experiment semantics on one GPU"
 message: "If you use miniVERL in your work, please cite it as below."
 type: software
-version: 0.13.0
+version: 0.14.0
 date-released: 2026-09-08
 license: Apache-2.0
 repository-code: "https://github.com/DaoyuanLi2816/mini-verl"
 url: "https://github.com/DaoyuanLi2816/mini-verl"
 abstract: >-
-  miniVERL runs a documented subset of verl-style on-policy distillation on one
-  consumer GPU. It is a local single-GPU runtime for a documented subset of verl v0.8
-  on-policy distillation. It consumes typed verl-shaped configuration and
-  Parquet prompts, executes actor rollout, teacher scoring and actor update in
-  one inspectable process, and exports standard PEFT, Parquet and configuration
-  artifacts through a fail-closed pinned bridge. Its native core also supports
-  multi-turn, tool-aware on-policy distillation: a student language
-  model samples its own tool-using trajectories against deterministic local
-  environments, a teacher scores exactly the states the student visited, and
-  the student is updated with token-level distributional supervision on its own
-  generated tokens only. It implements exact full-vocabulary forward KL,
-  reverse KL and Jensen-Shannon divergences as well as a compressed top-k plus
-  tail coarse-graining, enforces per-token provenance so that tool output can
-  never become a training label, and stores teacher targets in a versioned,
-  checksummed, pickle-free cache with policy-version enforcement.
-  Around that core it provides supervised fine-tuning, recorded-provenance
-  preference optimization and knowledge distillation; shared-backbone
-  execution that switches student, teacher and reference
-  roles across adapters on one set of base weights to fit low-memory hardware;
-  padded update batching; deterministic replay; transactional checkpoints and
-  cross-process run locks; and mechanism studies such as RecoveryBench that are
-  published with their negative and mixed results intact. A fail-closed
-  artifact bridge exchanges configs, LoRA adapters, tokenizers and Parquet
-  datasets with one pinned verl profile, verifying what it can and refusing to
-  imply the rest: it does not launch distributed jobs or establish broad
-  algorithmic parity. miniVERL is designed for one personal CUDA GPU,
-  automatically selects bf16 or fp16, and requires neither Ray nor a cluster.
-  Published performance is measured on one RTX 4080; other GPU models use the
  same code path but remain unmeasured. The v0.7 external-alignment study
-  terminated at its preregistered checkpoint-selection gate: two declared
-  lineages and eight candidates all scored 0/64 retained JSONNav utility, so
-  no teacher qualification, continuation method comparison or reserved final
-  test was run.
+  miniVERL compiles supported verl experiment semantics for one NVIDIA GPU.
+  Its versioned profiles consume resolved configuration and Parquet prompts,
+  run PPO with an independent critic or critic-free reward-driven algorithms,
+  and support on-policy distillation with explicit teachers. Typed trajectories,
+  role identities, transactional checkpoints and exact replay make local runs
+  inspectable and resumable. Portable PEFT, critic, Parquet and provenance
+  bundles support handoff toward pinned upstream verl. Installed PPO and GRPO
+  walkthroughs exercise the complete workflow. Runtime qualification is measured
+  on one RTX 4080 and kept separate from task-quality research; distributed
+  execution is outside the local runtime. Frozen studies preserve their original
+  negative and mixed results.
 authors:
   - family-names: Li
     given-names: Daoyuan

diff --git a/PROJECT_STATE.md b/PROJECT_STATE.md
@@ -6,21 +6,39 @@ current product and evidence state rather than repeating release history.
 
 Last updated: 2026-09-08.
 
-Canonical release state: stable `v0.13.0` (`8e7bc3b5ff1145ae502cca893b7fffca3e12b6a8`), development `0.13.1.dev0`.
+Canonical release state: releasing `v0.14.0`.
 
 ## Release state
 
+- Active work: `v0.14.0-p0-productization`, based on verified main `cdc608f`.
+  The next release focuses on upstream configuration coverage, packaged PPO/GRPO
+  walkthroughs, run inspection and system-level recovery. Implementation is
+  settled; release metadata is finalized for exact-candidate qualification.
+- Upstream drift check: GitHub's latest stable release is still `v0.9.0`
+  (published 2026-08-14); existing compiler targets remain appropriate.
+- Implemented: v3 prompt-minibatch units and explicit physical lowerings;
+  six upstream scripts/seven resolved cases with complete field accounting;
+  installed PPO/GRPO templates and adaptation ledgers; strict run inspection;
+  PPO value/behavior identity guards and checkpoint-bound log recovery.
+- Development rehearsal: both wheel-installed workflows completed on RTX 4080;
+  PPO 4 actor/4 critic updates and GRPO 2 actor updates, with exact tensor resume
+  and inspected handoff. This dirty-tree rehearsal is not release evidence.
+- Pending: final exact-commit qualification, publication and state sync.
+- Local regression: 2,655 non-GPU/non-network tests passed at 83% branch
+  coverage; 10 RTX 4080 GPU and 16 network tests passed. Four pinned upstream
+  conformance tests and four-viewport documentation browser checks passed.
+
 - Stable release: `v0.13.0` at
   `8e7bc3b5ff1145ae502cca893b7fffca3e12b6a8`.
-- Current development line: `0.13.1.dev0`.
+- Release candidate line: `0.14.0`.
 - 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
 
-Development `0.13.1.dev0` has a closed typed profile compiler: it retains the v1 critic-free contract and adds
-`verl-rl-v0.9-single-gpu-v2`, a typed resolved PPO subset of official verl
+Release `0.14.0` has a closed typed profile compiler: it retains v1/v2 and adds
+`verl-rl-v0.9-single-gpu-v3`, a typed resolved RL subset of official verl
 `v0.9.0` at commit `483b8a009ba3a97563edee3a19887e4862b8094a`.
 It compiles PPO/GAE into phased one-GPU actor, critic, reference and reward
 roles. The critic is independently trainable and owns its optimizer, schedule,