DaoyuanLi2816 · Sep 10, 2026 · Sep 10, 2026 · Sep 10, 2026 · Sep 10, 2026
diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
@@ -81,6 +81,9 @@ jobs:
           python scripts/release_state.py --check
           python scripts/check_known_good_environment.py
           python scripts/publish_gpu_qualification_schema.py --check
+          python scripts/publish_verl_direct_schema.py --check
+          python scripts/publish_direct_examples.py --check
+          python scripts/publish_direct_compatibility.py --check
           python scripts/check_markdown_links.py
           python scripts/check_text_integrity.py
           miniverl --help

diff --git a/.github/workflows/gpu.yml b/.github/workflows/gpu.yml
@@ -207,6 +207,13 @@ jobs:
             --output "$full_dir/v014-product.json" \
             --work "$full_work/v014-product" \
             "${offline[@]}"
+          "$QUALIFICATION_PYTHON" scripts/run_v015_direct_qualification.py \
+            --commit "$GITHUB_SHA" \
+            --wheel "$QUALIFICATION_WHEEL" \
+            --wheel-sha256 "$QUALIFICATION_WHEEL_SHA256" \
+            --output "$full_dir/v015-direct.json" \
+            --work "$full_work/v015-direct" \
+            "${offline[@]}"
           test -z "$(git status --porcelain)"
           "$QUALIFICATION_PYTHON" scripts/benchmark_rollout_runtime_v2_selected.py \
             --backend hf_cached \
@@ -231,7 +238,8 @@ jobs:
             --hf-reference benchmarks/results/rollout-runtime-v2-hf-reference.json \
             --v012-rl "$full_dir/v012-rl.json" \
             --v013-ppo "$full_dir/v013-ppo.json" \
-            --v014-product "$full_dir/v014-product.json"
+            --v014-product "$full_dir/v014-product.json" \
+            --v015-direct "$full_dir/v015-direct.json"
       - name: Upload full qualification records
         uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a # v7.0.1
         if: ${{ success() && inputs.qualification_level == 'full' }}

diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -6,6 +6,40 @@ All notable changes to miniVERL are recorded here. The format follows
 
 ## [Unreleased]
 
+## [0.15.0] - 2026-09-10
+
+### Direct verl configurations
+
+- Run resolved upstream PPO/GRPO-style YAML directly with `miniverl run FILE`.
+  The immutable v4 compiler separates semantic acceptance, missing local
+  bindings and exact-model hardware feasibility. Preserve v1/v2/v3 identities.
+- Classify all leaves of seven complete pinned upstream cases; add five stress
+  variants, including explicit unsupported sample-selection behavior.
+- Add typed model/data/reward bindings, checksum-approved Python reward code,
+  snapshot resolution, physical microbatch one and temporal role placement.
+- Retain source bytes, field decisions, bindings, local hardware plan and IR
+  identity through execution and checkpoint resume.
+
+### Experiment semantics
+
+- Support upstream token/sequence reductions, with scalar, gradient and
+  optimizer conformance; retain logical denominators across physical batches.
+- Filter overlong prompts before sampling, preserve drop-last traversal, derive
+  epoch schedules from retained rows, and execute scheduled reward validation.
+- Support repeated critic-free actor epochs, independent critic snapshots and
+  minibatches, and rollout-iteration AdamW learning-rate schedules.
+- Add an independent frozen-base KL reference and the upstream discriminative
+  classifier path using chat rendering and raw logits.
+- Preserve new reduction, validation and critic settings in fail-closed v4
+  handoff templates. Report unmeasured success as N/A and identify RL as on-policy.
+
+### Release integrity
+
+- Gate publication on installed-candidate direct PPO and GRPO runs, exact
+  actor/critic/optimizer replay, source-byte retention and inspected handoff.
+- Keep all frozen scientific JSON/JSONL, published tags and model revisions
+  unchanged. This release adds systems compatibility, not a new quality result.
+
 ## [0.14.0] - 2026-09-09
 
 ### Installed local RL workflows
@@ -1234,7 +1268,9 @@ 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/v0.14.0...HEAD
+[Unreleased]: https://github.com/DaoyuanLi2816/mini-verl/compare/v0.15.0...HEAD
+[0.15.0]: https://github.com/DaoyuanLi2816/mini-verl/compare/v0.14.0...v0.15.0
+[0.14.0]: https://github.com/DaoyuanLi2816/mini-verl/compare/v0.13.0...v0.14.0
 [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
@@ -2,8 +2,8 @@ cff-version: 1.2.0
 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: 0.14.0
-date-released: 2026-09-09
+version: 0.15.0
+date-released: 2026-09-10
 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
@@ -4,12 +4,25 @@ 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: 2026-09-09.
+Last updated: 2026-09-10.
 
-Canonical release state: stable `v0.14.0` (`60afc63da1257e3e22aa6056607a63a90d4d90f1`), development `0.14.1.dev0`.
+Canonical release state: releasing `v0.15.0`.
 
 ## Release state
 
+- v0.15.0 is preparing the direct-config release. Implementation and installed
+  development-wheel PPO/GRPO rehearsals pass; formal candidate qualification
+  and OIDC publication are pending.
+- v4 semantically accepts all seven complete pinned upstream configurations.
+  Five derived stress cases add two accepted and three explicit rejections.
+  Exact production-model execution is not claimed by static compilation.
+- The installed direct rehearsal completed four actor updates in both cases,
+  four PPO critic updates, reward validation, exact tensor replay and handoff.
+- Local full regression: 2,725 passed, 82.01% branch coverage, plus 18 pinned
+  upstream conformance cases and four documentation viewports.
+
+### Prior stable release (historical)
+
 - v0.14.0 productization is published: upstream configuration coverage,
   packaged PPO/GRPO walkthroughs, strict run inspection and system-level recovery.
 - Upstream drift check: GitHub's latest stable release is still `v0.9.0`
@@ -35,8 +48,8 @@ Canonical release state: stable `v0.14.0` (`60afc63da1257e3e22aa6056607a63a90d4d
 
 ## Current product boundary
 
-Development `0.14.1.dev0` 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
+Release `0.15.0` has a closed typed profile compiler: it retains v1/v2/v3 and adds
+`verl-rl-v0.9-single-gpu-v4`, direct execution of resolved RL inputs from 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,