diff --git a/.github/ISSUE_TEMPLATE/experiment_result.yml b/.github/ISSUE_TEMPLATE/experiment_result.yml
new file mode 100644
index 0000000..35c2761
--- /dev/null
+++ b/.github/ISSUE_TEMPLATE/experiment_result.yml
@@ -0,0 +1,70 @@
+name: Experiment result
+description: Submit a reproducible WAMProbe result for review
+title: "[Result]: "
+labels: ["experiment-result", "reproduction"]
+body:
+  - type: markdown
+    attributes:
+      value: >-
+        Submit structured evidence, not only a screenshot. Do not include tokens,
+        credentials, private datasets, browser sessions, or model weights.
+  - type: input
+    id: wamprobe_revision
+    attributes:
+      label: WAMProbe version and commit
+      placeholder: 0.1.0rc1 / full Git commit SHA
+    validations:
+      required: true
+  - type: textarea
+    id: model_provenance
+    attributes:
+      label: Adapter and model provenance
+      description: Include adapter revision, upstream repository revision, checkpoint ID/hash, license, and preprocessing configuration.
+    validations:
+      required: true
+  - type: textarea
+    id: benchmark
+    attributes:
+      label: Benchmark configuration
+      description: Include benchmark ID, context count/IDs, action set, seed, horizon, split, and any deviations.
+    validations:
+      required: true
+  - type: textarea
+    id: environment
+    attributes:
+      label: Environment and compute
+      description: Include operating system, Python, accelerator/driver, dependency lock or image digest, peak memory, and elapsed time.
+    validations:
+      required: true
+  - type: textarea
+    id: commands
+    attributes:
+      label: Exact commands
+      render: shell
+    validations:
+      required: true
+  - type: textarea
+    id: artifacts
+    attributes:
+      label: Artifacts and SHA256 hashes
+      description: Link or attach summary.json, results.jsonl, report.md/report.html, evidence manifest, and hashes where public redistribution is allowed.
+    validations:
+      required: true
+  - type: textarea
+    id: interpretation
+    attributes:
+      label: Result and limitations
+      description: Report the metric profile, unsupported metrics, uncertainty unit, failures, and the narrow claim supported by this run.
+    validations:
+      required: true
+  - type: checkboxes
+    id: integrity
+    attributes:
+      label: Evidence integrity
+      options:
+        - label: Every action branch uses the same restored initial context.
+          required: true
+        - label: The submitted result was generated by the revisions and commands above.
+          required: true
+        - label: No secret, private data, checkpoint, or unverifiable substituted output is attached.
+          required: true
diff --git a/.github/ISSUE_TEMPLATE/external_reproduction.yml b/.github/ISSUE_TEMPLATE/external_reproduction.yml
new file mode 100644
index 0000000..21757cd
--- /dev/null
+++ b/.github/ISSUE_TEMPLATE/external_reproduction.yml
@@ -0,0 +1,82 @@
+name: External reproduction
+description: Report an independent clean-install WAMProbe smoke reproduction
+title: "[External reproduction]: "
+labels: ["reproduction"]
+body:
+  - type: markdown
+    attributes:
+      value: >-
+        Thank you for independently checking WAMProbe. This form records the
+        evidence needed for the external-smoke release gate. Failed attempts are
+        useful too. Do not post credentials or private machine paths.
+  - type: input
+    id: reproducer
+    attributes:
+      label: Reproducer name or handle
+    validations:
+      required: true
+  - type: input
+    id: date
+    attributes:
+      label: Date and timezone
+      placeholder: 2026-07-16 UTC
+    validations:
+      required: true
+  - type: input
+    id: revision
+    attributes:
+      label: WAMProbe tag or full commit SHA
+    validations:
+      required: true
+  - type: textarea
+    id: environment
+    attributes:
+      label: Clean environment
+      description: Include OS/architecture, Python version, and whether this was a new venv, container, or machine.
+    validations:
+      required: true
+  - type: textarea
+    id: commands
+    attributes:
+      label: Installation and smoke commands
+      render: shell
+    validations:
+      required: true
+  - type: dropdown
+    id: result
+    attributes:
+      label: Result
+      options:
+        - PASS
+        - FAIL
+    validations:
+      required: true
+  - type: input
+    id: exit_status
+    attributes:
+      label: Exit status and elapsed time
+      placeholder: exit 0; 42 seconds
+    validations:
+      required: true
+  - type: input
+    id: output_hash
+    attributes:
+      label: Output summary.json SHA256
+    validations:
+      required: true
+  - type: textarea
+    id: deviations
+    attributes:
+      label: Deviations, errors, or notes
+      description: Write "None" when the documented procedure was followed exactly. For a failure, include the smallest non-sensitive error output.
+    validations:
+      required: true
+  - type: checkboxes
+    id: independence
+    attributes:
+      label: Independence confirmation
+      options:
+        - label: I performed this install and smoke test independently; this is not a maintainer-run substitute for an external user.
+          required: true
+        - label: I reviewed the form and removed tokens, credentials, private paths, and other secrets.
+          required: true
diff --git a/CONTRIBUTING.md b/CONTRIBUTING.md
index fbff90d..a98b88b 100644
--- a/CONTRIBUTING.md
+++ b/CONTRIBUTING.md
@@ -2,6 +2,27 @@
 
 Thank you for helping make WAM evaluation more causal, reproducible, and control-grounded.
 
+## Choose a contribution path
+
+- **Connect a model:** run the
+  [custom adapter starter](docs/adapters/CUSTOM_ADAPTER.md), then open an
+  [Adapter proposal](https://github.com/myheart521/WAMProbe/issues/new?template=adapter_proposal.yml).
+- **Add a paired diagnostic:** open a
+  [Benchmark proposal](https://github.com/myheart521/WAMProbe/issues/new?template=benchmark_proposal.yml)
+  before generating a large dataset.
+- **Add or change a metric:** open a
+  [Metric proposal](https://github.com/myheart521/WAMProbe/issues/new?template=metric_proposal.yml)
+  and state how the metric can be gamed.
+- **Share a run:** use the
+  [Experiment result](https://github.com/myheart521/WAMProbe/issues/new?template=experiment_result.yml)
+  form with standard artifacts and hashes.
+- **Reproduce independently:** use the
+  [External reproduction](https://github.com/myheart521/WAMProbe/issues/new?template=external_reproduction.yml)
+  form. A maintainer-controlled run cannot satisfy the external-user gate.
+- **Pick a bounded task:** see the
+  [roadmap](docs/NEXT_STEPS.md) and issues labeled
+  [`good first issue`](https://github.com/myheart521/WAMProbe/labels/good%20first%20issue).
+
 ## Development setup
 
 ```bash
@@ -10,6 +31,15 @@ source .venv/bin/activate
 python -m pip install -e '.[dev]'
 ```
 
+Verify the adapter starter before replacing its backend:
+
+```bash
+python -m examples.custom_adapter.run \
+  --output runs/custom-adapter \
+  --contexts 8 \
+  --seed 7
+```
+
 Before opening a pull request, run:
 
 ```bash
@@ -31,11 +61,14 @@ the audit itself.
 - Follow test-driven development for behavior changes.
 - Keep the core package CPU-only and avoid heavy mandatory dependencies.
 - New adapters must publish a capability declaration and pin an upstream revision used by
-  their smoke test.
+  their smoke test. Preserve context/action identities and reject malformed model outputs
+  at the adapter boundary.
 - New benchmarks must document generation, splits, licenses, and shared-state guarantees.
 - New metrics must include a metric card, expected baseline ordering, and at least one
   failure/anti-gaming test.
 - Do not commit datasets, checkpoints, generated runs, tokens, or browser sessions.
+- Result submissions must identify the WAMProbe, adapter, model, checkpoint, preprocessing,
+  and benchmark revisions; include exact commands and public artifact hashes.
 
 Please use focused commits and explain both the scientific reason and user-visible impact
 of a change.
diff --git a/README.md b/README.md
index f072284..a71cbf8 100644
--- a/README.md
+++ b/README.md
@@ -241,6 +241,15 @@ Start with the [scope and capability RFC](docs/rfcs/0001-scope-and-capabilities.
 [counterfactual metrics RFC](docs/rfcs/0002-counterfactual-metrics.md), and the
 [adapter selection record](docs/research/ADAPTER_SELECTION.md).
 
+For a working integration instead of pseudocode, run the tested
+[custom adapter starter kit](examples/custom_adapter/README.md) and follow the
+[step-by-step adapter guide](docs/adapters/CUSTOM_ADAPTER.md):
+
+```bash
+python -m examples.custom_adapter.run \
+  --output runs/custom-adapter --contexts 8 --seed 7
+```
+
 ## Real-model artifacts
 
 Model weights are never committed to Git. The first StarWAM spike requires approximately
@@ -280,11 +289,14 @@ Review the [reproducibility guide](docs/reproducibility/REPRODUCIBILITY.md) and
 | Run the CPU demo | [Quick start](docs/QUICKSTART.md) |
 | Understand the scientific motivation | [WAM/VLA failure-case evidence map](docs/research/WAM_VLA_FAILURE_CASES.md) |
 | Interpret a score correctly | [Core metric cards](docs/metrics/CORE_METRICS.md) |
-| Add a model integration | [Scope and capability RFC](docs/rfcs/0001-scope-and-capabilities.md) |
+| Add a model integration | [Custom adapter starter guide](docs/adapters/CUSTOM_ADAPTER.md) |
+| Submit a structured experiment | [Experiment result form](https://github.com/myheart521/WAMProbe/issues/new?template=experiment_result.yml) |
+| Reproduce independently | [External reproduction form](https://github.com/myheart521/WAMProbe/issues/new?template=external_reproduction.yml) |
 | Reproduce LIBERO-CF-Mini | [Benchmark card](docs/benchmarks/LIBERO_CF_MINI.md) |
 | Inspect the closed-loop protocol | [Experiment card](docs/experiments/TOY_CLOSED_LOOP_V0.1.md) |
 | Audit the release | [Reproducibility guide](docs/reproducibility/REPRODUCIBILITY.md) |
 | Read the full implementation plan | [Detailed Chinese project plan](docs/WAMProbe_PLAN.md) |
+| Choose the next milestone | [Dependency-aware roadmap](docs/NEXT_STEPS.md) |
 | Browse all documentation | [Published documentation site](https://myheart521.github.io/WAMProbe/) |
 
 ## Project status and roadmap
@@ -293,7 +305,8 @@ The `v0.1.0rc1` engineering scope is implemented and published. The remaining fo
 acceptance item is one genuine independent clean-install smoke report in
 [Issue #2](https://github.com/myheart521/WAMProbe/issues/2).
 
-Next research milestones:
+The [dependency-aware roadmap](docs/NEXT_STEPS.md) tracks evidence gates separately from
+repository work. Next research milestones:
 
 1. expand LIBERO initial-state coverage;
 2. evaluate action-conditioned real-WAM futures when an adapter exposes that capability;
@@ -306,6 +319,11 @@ Contributions are welcome—especially adapters, paired benchmark generators, me
 anti-gaming tests, and independent reproductions. New metrics must document the failure
 they detect and include a sanity check against reference baselines.
 
+Start with a labeled
+[`good first issue`](https://github.com/myheart521/WAMProbe/labels/good%20first%20issue),
+an [Adapter proposal](https://github.com/myheart521/WAMProbe/issues/new?template=adapter_proposal.yml),
+or the [Experiment result form](https://github.com/myheart521/WAMProbe/issues/new?template=experiment_result.yml).
+
 ```bash
 python -m pip install -e '.[dev]'
 ruff format --check .
diff --git a/README.zh-CN.md b/README.zh-CN.md
index 96b4ab1..963436b 100644
--- a/README.zh-CN.md
+++ b/README.zh-CN.md
@@ -235,6 +235,15 @@ class MyWorldModel:
 [反事实指标 RFC](docs/rfcs/0002-counterfactual-metrics.md)和
 [适配器选择记录](docs/research/ADAPTER_SELECTION.md)。
 
+如果希望从可运行代码开始，而不是照着伪代码重新搭建，请运行已经测试过的
+[自定义 Adapter Starter Kit](examples/custom_adapter/README.md)，并按照
+[适配器接入指南](docs/adapters/CUSTOM_ADAPTER.md)逐步替换后端：
+
+```bash
+python -m examples.custom_adapter.run \
+  --output runs/custom-adapter --contexts 8 --seed 7
+```
+
 ## 真实模型文件
 
 模型权重永远不会提交到 Git。第一个 StarWAM 验证需要约 46.3 GB 固定版本的 StarWAM 和
@@ -271,11 +280,14 @@ StarWAM 和 LIBERO 的隔离环境、GPU 预检、预处理来源及 smoke 命
 | 运行 CPU 演示 | [快速开始](docs/QUICKSTART.md) |
 | 理解项目的研究动机 | [WAM/VLA 失败案例证据图谱](docs/research/WAM_VLA_FAILURE_CASES.md) |
 | 正确解释指标 | [核心指标卡](docs/metrics/CORE_METRICS.md) |
-| 接入一个新模型 | [范围与能力 RFC](docs/rfcs/0001-scope-and-capabilities.md) |
+| 接入一个新模型 | [自定义适配器接入指南](docs/adapters/CUSTOM_ADAPTER.md) |
+| 提交结构化实验结果 | [实验结果表单](https://github.com/myheart521/WAMProbe/issues/new?template=experiment_result.yml) |
+| 独立复现项目 | [外部复现表单](https://github.com/myheart521/WAMProbe/issues/new?template=external_reproduction.yml) |
 | 复现 LIBERO-CF-Mini | [基准卡](docs/benchmarks/LIBERO_CF_MINI.md) |
 | 检查闭环实验协议 | [实验卡](docs/experiments/TOY_CLOSED_LOOP_V0.1.md) |
 | 审计发行版 | [可复现性指南](docs/reproducibility/REPRODUCIBILITY.md) |
 | 阅读完整实施计划 | [WAMProbe 中文规划文档](docs/WAMProbe_PLAN.md) |
+| 选择下一阶段任务 | [带依赖关系的后续路线图](docs/NEXT_STEPS.md) |
 | 浏览全部文档 | [在线文档站](https://myheart521.github.io/WAMProbe/) |
 
 ## 项目状态与路线图
@@ -283,7 +295,7 @@ StarWAM 和 LIBERO 的隔离环境、GPU 预检、预处理来源及 smoke 命
 `v0.1.0rc1` 工程范围已经实现并发布。正式验收目前只剩一项：由真实独立用户在
 [Issue #2](https://github.com/myheart521/WAMProbe/issues/2) 提交干净环境安装复现报告。
 
-下一阶段研究计划：
+[后续路线图](docs/NEXT_STEPS.md)会把外部证据门槛和仓库内工程任务分开跟踪。下一阶段研究计划：
 
 1. 扩展 LIBERO 初始状态覆盖；
 2. 在适配器真正暴露对应能力后，评测动作条件真实 WAM 未来；
@@ -295,6 +307,11 @@ StarWAM 和 LIBERO 的隔离环境、GPU 预检、预处理来源及 smoke 命
 欢迎提交适配器、配对基准生成器、指标反作弊测试和独立复现报告。新增指标必须说明它要
 发现的失败模式，并包含相对参考基线的 sanity check。
 
+可以先认领带有
+[`good first issue`](https://github.com/myheart521/WAMProbe/labels/good%20first%20issue) 标签的任务，
+提交 [Adapter 提案](https://github.com/myheart521/WAMProbe/issues/new?template=adapter_proposal.yml)，
+或使用[实验结果表单](https://github.com/myheart521/WAMProbe/issues/new?template=experiment_result.yml)。
+
 ```bash
 python -m pip install -e '.[dev]'
 ruff format --check .
diff --git a/docs/NEXT_STEPS.md b/docs/NEXT_STEPS.md
new file mode 100644
index 0000000..6ab577c
--- /dev/null
+++ b/docs/NEXT_STEPS.md
@@ -0,0 +1,112 @@
+# Next steps / 后续路线图
+
+This roadmap separates work that can be automated in the repository from gates
+that require independent people or new scientific evidence. A checked code task
+must never be used as a substitute for an unmet evidence gate.
+
+本路线图区分“仓库内可以自动完成的工程任务”和“必须由独立用户或新实验提供证据的验收
+门槛”。代码完成不代表外部证据已经满足。
+
+## Dependency path / 依赖关系
+
+```text
+P0 independent reproduction ──► P1 final v0.1.0
+
+P2 contributor readiness ──► P3 broader paired benchmarks
+                         └──► P4 action-conditioned real WAM
+
+P1 stable release + public evidence ──► P5 DOI and outreach
+```
+
+## Milestones / 里程碑
+
+| Priority | Milestone | Current state | Completion evidence |
+|---|---|---|---|
+| P0 | Independent clean-install reproduction / 独立干净环境复现 | **Blocked by an independent user** | A non-maintainer posts commit, environment, commands, exit status, elapsed time, and output hash in Issue #2 |
+| P1 | Promote `v0.1.0rc1` to `v0.1.0` / 发布正式版 | **Blocked by P0** | P0 accepted, release audit green, signed/tagged release artifacts and provenance available |
+| P2 | Contributor readiness / 贡献者接入能力 | **In progress** | Runnable adapter starter, contribution/result forms, roadmap, labeled starter issues, green CI |
+| P3 | Broader paired diagnostics / 扩展配对诊断 | **Planned** | More LIBERO initial states plus Occluded-Object memory benchmark with restore/order checks |
+| P4 | Real action-conditioned WAM evidence / 真实动作条件 WAM | **Planned; model capability required** | Public model exposes candidate-conditioned futures, adapter passes semantic tests, bounded report is reproducible |
+| P5 | Archival and outreach / 归档与传播 | **After stable release** | Zenodo DOI, updated citation metadata, release notes, public result call |
+
+Tracked work:
+
+- P0: [independent reproduction gate #2](https://github.com/myheart521/WAMProbe/issues/2)
+- P3: [LIBERO initial-state coverage #10](https://github.com/myheart521/WAMProbe/issues/10)
+  and [Occluded-Object RFC #7](https://github.com/myheart521/WAMProbe/issues/7)
+- P4: [public action-conditioned WAM audit #9](https://github.com/myheart521/WAMProbe/issues/9)
+- Starter contributions: [adapter troubleshooting #11](https://github.com/myheart521/WAMProbe/issues/11)
+  and [malformed-output tests #12](https://github.com/myheart521/WAMProbe/issues/12)
+- P5: [Zenodo DOI preparation #8](https://github.com/myheart521/WAMProbe/issues/8)
+
+## P0 — independent reproduction
+
+The maintainer can improve instructions and debug a failed report, but cannot
+perform this acceptance item on the maintainer's own machine. The external user
+should use the **External reproduction** issue form. Issue #2 remains the source
+of truth.
+
+维护者可以完善说明、协助排错，但不能在自己的机器上代替外部用户完成验收。真实独立用户
+应通过 **External reproduction** 表单提交，Issue #2 是唯一验收记录。
+
+## P1 — stable release gate
+
+After P0 is accepted:
+
+1. confirm the candidate commit is clean and all required CI/security checks pass;
+2. run the documented reproducible build and archive audit;
+3. update version, changelog, citation, and release notes consistently;
+4. publish the final tag and artifacts through the release workflow;
+5. verify a clean install from the published package.
+
+Do not publish `v0.1.0` before P0.
+
+## P2 — contributor readiness
+
+Repository-owned deliverables:
+
+- a tested `examples/custom_adapter` integration path;
+- a detailed adapter guide with semantic and provenance checks;
+- structured experiment-result and external-reproduction issue forms;
+- visible, labeled, bounded issues suitable for first-time contributors;
+- README and documentation-site links to each contribution path.
+
+## P3 — benchmark expansion
+
+Two bounded tracks should remain separate:
+
+- **LIBERO-CF-Mini coverage:** add initial states and perturbations while preserving
+  snapshot restore, shared-context pairing, deterministic action order reversal,
+  asset licensing, and split/leakage documentation.
+- **Occluded-Object diagnostic:** test whether action-conditioned futures retain a
+  task-relevant object after it becomes unobserved. Define oracle, copy-last,
+  memoryless, and wrong-memory baselines before adding a learned model.
+
+Every addition needs a benchmark card, generation or conversion code, hashes,
+and expected-ordering tests.
+
+## P4 — real WAM integration
+
+Select a public model only when its actual interface can produce a future for
+each proposed action from the same context. A model that emits only an action
+chunk may be useful for another WAMProbe capability tier, but it does not provide
+candidate-conditioned future evidence.
+
+The first integration should be intentionally small: one pinned checkpoint, one
+configuration, one paired context group, a documented GPU budget, cached
+prediction provenance, and an explicit list of unavailable metrics. Expand only
+after the semantic smoke test passes.
+
+## P5 — DOI and outreach
+
+Prepare Zenodo metadata and communication material after the stable release so
+the DOI points to immutable public artifacts. Outreach should ask for independent
+reproductions and comparable structured reports, not only stars or screenshots.
+
+## Maintainer decision rules / 维护规则
+
+- Never mark P0 complete from a maintainer-controlled machine.
+- Never release a metric without an anti-gaming or failure-baseline test.
+- Never label a proxy as a capability the model does not expose.
+- Never commit checkpoints, datasets, credentials, or generated private evidence.
+- Prefer small reproducible issues with an explicit acceptance checklist.
diff --git a/docs/adapters/CUSTOM_ADAPTER.md b/docs/adapters/CUSTOM_ADAPTER.md
new file mode 100644
index 0000000..af7ad55
--- /dev/null
+++ b/docs/adapters/CUSTOM_ADAPTER.md
@@ -0,0 +1,132 @@
+# Connect a custom state-future model
+
+This guide turns a model-specific inference function into WAMProbe's typed
+`WAMAdapter` contract. Start from the runnable code in
+[`examples/custom_adapter`](https://github.com/myheart521/WAMProbe/tree/main/examples/custom_adapter)
+rather than
+building the evaluation loop yourself.
+
+## What the starter kit proves
+
+The example deliberately separates model inference from evaluation:
+
+```text
+shared Context2D + candidate Action2D
+                  │
+                  ▼
+          PredictionRequest
+                  │
+                  ▼
+      your StateFutureBackend
+                  │
+                  ▼
+validated horizon-length state future
+                  │
+                  ▼
+ WAMProbe metrics and standard report artifacts
+```
+
+The included `LinearStateBackend` is only a wiring oracle. It uses the action
+delta directly and should achieve zero Top-1 Regret on PointMass2D. It is not a
+learned model and its result is not research evidence.
+
+## Run the unchanged example
+
+From a development checkout:
+
+```bash
+python -m pip install -e '.[dev]'
+python -m examples.custom_adapter.run \
+  --output runs/custom-adapter \
+  --contexts 8 \
+  --seed 7
+```
+
+The output directory contains:
+
+```text
+runs/custom-adapter/
+├── summary.json   # aggregate result, uncertainty, and schema version
+├── results.jsonl  # stable context-level records
+├── report.md      # reviewable metric table
+└── report.html    # standalone human-readable report
+```
+
+## Implement the backend boundary
+
+Copy `examples/custom_adapter` into a model-specific module and replace
+`LinearStateBackend` with a class implementing this narrow protocol:
+
+```python
+class MyBackend:
+    def predict_positions(
+        self,
+        request: PredictionRequest,
+    ) -> tuple[tuple[float, float], ...]:
+        model_input = preprocess(request)
+        prediction = self.model.generate(model_input, seed=request.seed)
+        return decode_positions(prediction, horizon=request.horizon)
+```
+
+Every request field has evaluation semantics and must be mapped intentionally:
+
+| Field | Meaning | Integration check |
+|---|---|---|
+| `context_id` | Identity of the shared initial state | Equal across every action branch in a group |
+| `action_name` | Stable candidate-action identity | Preserved in the returned trajectory |
+| `position_xy` | Current observed state | Never replaced with a post-action state |
+| `goal_xy` | Task goal used by the toy scorer | Keep separate from the action input |
+| `action_delta_xy` | Counterfactual intervention | Must affect the model input or conditioning path |
+| `horizon` | Required number of future states | Return exactly this many states |
+| `seed` | Reproduction seed | Forward to every stochastic sampler |
+
+The adapter rejects non-positive horizons, missing or extra time steps, malformed
+coordinates, and non-finite values before metrics see them. Keep these checks when
+replacing the backend.
+
+## Declare only real capabilities
+
+`ModelCapabilities` controls which evidence can be interpreted. The starter
+declares deterministic state futures. If repeated calls with the same seed are
+not deterministic, set `deterministic_seed=False`. Set `stochastic=True` when the
+model represents a distribution or samples futures. Do not claim pixels,
+latents, action scores, or action prediction unless the adapter exposes those
+outputs directly.
+
+The current starter targets models that can decode a two-dimensional state
+future. Pixel- or latent-only models require a capability-specific adapter and
+metrics; do not silently convert an unrelated proxy into a state trajectory.
+
+## Required model-specific tests
+
+Before proposing an adapter, add tests that demonstrate:
+
+1. two different actions from one context reach the model as different inputs;
+2. context and action identities survive preprocessing and decoding;
+3. the same seed reproduces the same result, or nondeterminism is declared;
+4. wrong-length, NaN, and infinite outputs fail loudly;
+5. the smallest public checkpoint/configuration completes one bounded smoke run;
+6. expected oracle and broken-baseline ordering remains intact.
+
+Run the repository checks listed in
+[`CONTRIBUTING.md`](https://github.com/myheart521/WAMProbe/blob/main/CONTRIBUTING.md),
+then open an **Adapter proposal** before a large integration. Pin the upstream
+repository revision, checkpoint hash, preprocessing version, license, hardware,
+and exact smoke command.
+
+## Submit a result responsibly
+
+Use the **Experiment result** issue form and attach or link the standard report
+artifacts. A valid result identifies the WAMProbe commit, adapter and upstream
+revisions, benchmark configuration, hardware/software environment, exact command,
+artifact hashes, and unsupported metrics. Generated evidence must not contain
+tokens, private data, browser sessions, local credentials, or model weights.
+
+## 中文速览
+
+这个 Starter Kit 的作用，是把模型自己的推理函数接到 WAMProbe 的类型化评测接口上。
+你通常只需要复制 `examples/custom_adapter`，替换 `LinearStateBackend`，并把
+`PredictionRequest` 中的上下文、动作、预测长度和随机种子完整传给模型。后端必须返回与
+`horizon` 等长的有限二维状态序列；不要丢弃动作条件，也不要把像素或潜变量偷偷换成
+不相关的状态代理。提交适配器前，需要固定上游版本、权重哈希、预处理规则、环境和 smoke
+命令，并为动作条件、确定性和异常输出补测试。
diff --git a/docs/index.md b/docs/index.md
index 9cd1547..ab5c246 100644
--- a/docs/index.md
+++ b/docs/index.md
@@ -21,10 +21,12 @@ It intentionally does not publish a single composite score.
 ## Start here
 
 - Follow the [15-minute quick start](QUICKSTART.md) for dependency-free CPU experiments.
+- Connect a model with the tested [custom adapter starter guide](adapters/CUSTOM_ADAPTER.md).
 - Read the [scope and capability RFC](rfcs/0001-scope-and-capabilities.md) before adding an adapter.
 - Use the [core metric cards](metrics/CORE_METRICS.md) to select capability-compatible metrics.
 - Review the [reproducibility guide](reproducibility/REPRODUCIBILITY.md) before comparing results.
 - Inspect the [StarWAM model card](models/STARWAM.md) for the first real-model integration.
+- Check [Next steps](NEXT_STEPS.md) for milestone dependencies and contribution-sized tasks.
 
 Install the dependency-free core with `pip install wamprobe`. Committed example reports
 and audited release artifacts remain available on
diff --git a/examples/__init__.py b/examples/__init__.py
new file mode 100644
index 0000000..5d95b54
--- /dev/null
+++ b/examples/__init__.py
@@ -0,0 +1 @@
+"""Runnable examples shipped with the WAMProbe source distribution."""
diff --git a/examples/custom_adapter/README.md b/examples/custom_adapter/README.md
new file mode 100644
index 0000000..b7f467f
--- /dev/null
+++ b/examples/custom_adapter/README.md
@@ -0,0 +1,33 @@
+# Custom adapter starter kit
+
+This runnable example shows the smallest safe bridge between a custom
+state-future model and WAMProbe. The adapter preserves the context ID and action
+name, passes an explicit action intervention to the backend, validates the
+prediction horizon and finite coordinates, and declares machine-readable model
+capabilities.
+
+Run it from the repository root:
+
+```bash
+python -m examples.custom_adapter.run \
+  --output runs/custom-adapter \
+  --contexts 8 \
+  --seed 0
+```
+
+The command writes `summary.json`, `results.jsonl`, `report.md`, and
+`report.html`. The included linear backend should achieve zero Top-1 Regret on
+PointMass2D; this is a wiring check, not a research result.
+
+To connect a model:
+
+1. Implement `StateFutureBackend.predict_positions()` in `adapter.py`.
+2. Map every field in `PredictionRequest` to your model input without dropping
+   the action intervention or context identity.
+3. Return exactly one `(x, y)` state per requested horizon step.
+4. Update `ModelCapabilities` if your backend is stochastic or has different
+   behavior.
+5. Run this example, then add model-specific tests before submitting results.
+
+See [`docs/adapters/CUSTOM_ADAPTER.md`](../../docs/adapters/CUSTOM_ADAPTER.md)
+for the full integration and submission guide.
diff --git a/examples/custom_adapter/__init__.py b/examples/custom_adapter/__init__.py
new file mode 100644
index 0000000..3091376
--- /dev/null
+++ b/examples/custom_adapter/__init__.py
@@ -0,0 +1,9 @@
+"""Minimal, runnable state-future adapter starter kit."""
+
+from examples.custom_adapter.adapter import (
+    LinearStateBackend,
+    PredictionRequest,
+    StarterWAMAdapter,
+)
+
+__all__ = ["LinearStateBackend", "PredictionRequest", "StarterWAMAdapter"]
diff --git a/examples/custom_adapter/adapter.py b/examples/custom_adapter/adapter.py
new file mode 100644
index 0000000..9467984
--- /dev/null
+++ b/examples/custom_adapter/adapter.py
@@ -0,0 +1,113 @@
+"""A small reference bridge from an arbitrary backend to ``WAMAdapter``.
+
+Replace ``LinearStateBackend`` with a wrapper around your model. Keep the
+``StarterWAMAdapter`` validation boundary so benchmark inputs and model outputs
+remain semantically aligned.
+"""
+
+from __future__ import annotations
+
+import math
+from dataclasses import dataclass
+from typing import Protocol
+
+from wamprobe.api.capabilities import FutureRepresentation, ModelCapabilities
+from wamprobe.api.types import Action2D, Context2D, Trajectory2D, Vec2
+
+
+@dataclass(frozen=True, slots=True)
+class PredictionRequest:
+    """Backend-neutral prediction request with explicit intervention semantics."""
+
+    context_id: str
+    action_name: str
+    position_xy: tuple[float, float]
+    goal_xy: tuple[float, float]
+    action_delta_xy: tuple[float, float]
+    horizon: int
+    seed: int
+
+
+class StateFutureBackend(Protocol):
+    """Narrow interface that a custom state-future model must implement."""
+
+    def predict_positions(
+        self,
+        request: PredictionRequest,
+    ) -> tuple[tuple[float, float], ...]:
+        """Return one predicted ``(x, y)`` state for every requested step."""
+
+        ...
+
+
+class LinearStateBackend:
+    """Dependency-free example backend using the supplied action intervention."""
+
+    def predict_positions(
+        self,
+        request: PredictionRequest,
+    ) -> tuple[tuple[float, float], ...]:
+        x, y = request.position_xy
+        dx, dy = request.action_delta_xy
+        return tuple((x + dx * step, y + dy * step) for step in range(1, request.horizon + 1))
+
+
+class StarterWAMAdapter:
+    """Validate and expose a custom state-future backend as a WAMProbe adapter."""
+
+    def __init__(
+        self,
+        backend: StateFutureBackend,
+        *,
+        model_id: str = "starter-linear-state",
+    ) -> None:
+        self._backend = backend
+        self._capabilities = ModelCapabilities(
+            model_id=model_id,
+            future_representation=FutureRepresentation.STATES,
+            deterministic_seed=True,
+        )
+
+    @property
+    def capabilities(self) -> ModelCapabilities:
+        """Declare the output representation and deterministic seed behavior."""
+
+        return self._capabilities
+
+    def predict_future(
+        self,
+        context: Context2D,
+        action: Action2D,
+        *,
+        horizon: int,
+        seed: int,
+    ) -> Trajectory2D:
+        """Translate typed inputs, call the backend, and validate its state future."""
+
+        if horizon <= 0:
+            raise ValueError("horizon must be positive")
+        request = PredictionRequest(
+            context_id=context.context_id,
+            action_name=action.name,
+            position_xy=(context.position.x, context.position.y),
+            goal_xy=(context.goal.x, context.goal.y),
+            action_delta_xy=(action.delta.x, action.delta.y),
+            horizon=horizon,
+            seed=seed,
+        )
+        positions = self._backend.predict_positions(request)
+        if len(positions) != horizon:
+            raise ValueError(f"backend must return exactly {horizon} states")
+
+        states: list[Vec2] = []
+        for index, position in enumerate(positions):
+            if len(position) != 2:
+                raise ValueError(f"backend state {index} must contain exactly two coordinates")
+            x, y = position
+            if not math.isfinite(x) or not math.isfinite(y):
+                raise ValueError(f"backend state {index} coordinates must be finite")
+            states.append(Vec2(float(x), float(y)))
+        return Trajectory2D(context.context_id, action.name, tuple(states))
+
+    def close(self) -> None:
+        """Release resources; the starter backend owns none."""
diff --git a/examples/custom_adapter/run.py b/examples/custom_adapter/run.py
new file mode 100644
index 0000000..9f0d028
--- /dev/null
+++ b/examples/custom_adapter/run.py
@@ -0,0 +1,43 @@
+"""Run the custom-adapter starter kit against PointMass2D."""
+
+from __future__ import annotations
+
+import argparse
+from pathlib import Path
+
+from examples.custom_adapter.adapter import LinearStateBackend, StarterWAMAdapter
+from wamprobe.benchmarks.pointmass import PointMass2D
+from wamprobe.evaluation import evaluate
+from wamprobe.reporting import write_reports
+
+
+def run(output: Path, *, contexts: int, seed: int) -> tuple[Path, Path, Path]:
+    """Evaluate the starter adapter and write WAMProbe's standard artifacts."""
+
+    benchmark = PointMass2D()
+    suite = benchmark.make_suite(contexts=contexts, seed=seed)
+    adapter = StarterWAMAdapter(LinearStateBackend())
+    result = evaluate(adapter, benchmark, suite, seed=seed)
+    return write_reports(
+        output,
+        benchmark=benchmark.benchmark_id,
+        results=[result],
+        seed=seed,
+    )
+
+
+def main() -> None:
+    """Parse CLI arguments and run the starter evaluation."""
+
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--output", type=Path, default=Path("runs/custom-adapter"))
+    parser.add_argument("--contexts", type=int, default=8)
+    parser.add_argument("--seed", type=int, default=0)
+    args = parser.parse_args()
+    paths = run(args.output, contexts=args.contexts, seed=args.seed)
+    for path in paths:
+        print(path)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/mkdocs.yml b/mkdocs.yml
index 3f0857c..e36c8c7 100644
--- a/mkdocs.yml
+++ b/mkdocs.yml
@@ -25,6 +25,8 @@ nav:
   - Research:
       - WAM/VLA failure cases: research/WAM_VLA_FAILURE_CASES.md
       - Adapter selection: research/ADAPTER_SELECTION.md
+  - Adapters:
+      - Connect a custom model: adapters/CUSTOM_ADAPTER.md
   - Benchmarks:
       - Toy benchmarks: benchmarks/TOY_BENCHMARKS.md
       - LIBERO-CF-Mini: benchmarks/LIBERO_CF_MINI.md
@@ -38,3 +40,4 @@ nav:
       - Guide: reproducibility/REPRODUCIBILITY.md
       - External report template: reproducibility/EXTERNAL_REPRODUCTION_TEMPLATE.md
       - GPU nightly: reproducibility/GPU_NIGHTLY.md
+  - Next steps: NEXT_STEPS.md
diff --git a/pyproject.toml b/pyproject.toml
index 85b3e68..fa28503 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -38,10 +38,16 @@ packages = ["src/wamprobe"]
 include = [
   "/CHANGELOG.md",
   "/CITATION.cff",
+  "/CONTRIBUTING.md",
   "/LICENSE",
   "/README.md",
   "/README.zh-CN.md",
+  "/docs/**/*.md",
   "/docs/assets/*.png",
+  "/examples/__init__.py",
+  "/examples/custom_adapter/*.md",
+  "/examples/custom_adapter/*.py",
+  "/mkdocs.yml",
   "/pyproject.toml",
   "/src/wamprobe",
 ]
@@ -62,6 +68,7 @@ exclude = [
 
 [tool.pytest.ini_options]
 testpaths = ["tests"]
+pythonpath = ["."]
 addopts = [
   "--strict-markers",
 ]
diff --git a/tests/test_custom_adapter_example.py b/tests/test_custom_adapter_example.py
new file mode 100644
index 0000000..9c5ed09
--- /dev/null
+++ b/tests/test_custom_adapter_example.py
@@ -0,0 +1,70 @@
+import json
+from pathlib import Path
+
+import pytest
+
+from examples.custom_adapter.adapter import (
+    LinearStateBackend,
+    PredictionRequest,
+    StarterWAMAdapter,
+)
+from examples.custom_adapter.run import run
+from wamprobe.api.capabilities import FutureRepresentation
+from wamprobe.api.model import WAMAdapter
+from wamprobe.api.types import Action2D, Context2D, Vec2
+
+
+def test_starter_adapter_is_typed_seeded_and_action_conditioned() -> None:
+    adapter = StarterWAMAdapter(LinearStateBackend(), model_id="starter-test")
+    context = Context2D("context-0", position=Vec2(0.0, 0.0), goal=Vec2(1.0, 0.0))
+    action = Action2D("right", Vec2(0.25, 0.0))
+
+    first = adapter.predict_future(context, action, horizon=4, seed=7)
+    second = adapter.predict_future(context, action, horizon=4, seed=7)
+
+    assert isinstance(adapter, WAMAdapter)
+    assert adapter.capabilities.model_id == "starter-test"
+    assert adapter.capabilities.future_representation is FutureRepresentation.STATES
+    assert adapter.capabilities.deterministic_seed is True
+    assert first == second
+    assert first.context_id == context.context_id
+    assert first.action_name == action.name
+    assert first.states == (
+        Vec2(0.25, 0.0),
+        Vec2(0.5, 0.0),
+        Vec2(0.75, 0.0),
+        Vec2(1.0, 0.0),
+    )
+
+
+def test_starter_adapter_rejects_a_backend_with_the_wrong_horizon() -> None:
+    class ShortBackend:
+        def predict_positions(
+            self,
+            request: PredictionRequest,
+        ) -> tuple[tuple[float, float], ...]:
+            return ((request.position_xy[0], request.position_xy[1]),)
+
+    adapter = StarterWAMAdapter(ShortBackend())
+    context = Context2D("context-0", position=Vec2(0.0, 0.0), goal=Vec2(1.0, 0.0))
+    action = Action2D("right", Vec2(0.25, 0.0))
+
+    with pytest.raises(ValueError, match="exactly 4 states"):
+        adapter.predict_future(context, action, horizon=4, seed=7)
+
+
+def test_starter_runner_writes_the_standard_report_contract(tmp_path: Path) -> None:
+    paths = run(tmp_path, contexts=4, seed=7)
+
+    assert {path.name for path in paths} == {"summary.json", "report.md", "report.html"}
+    assert {path.name for path in tmp_path.iterdir()} == {
+        "summary.json",
+        "results.jsonl",
+        "report.md",
+        "report.html",
+    }
+    summary = json.loads((tmp_path / "summary.json").read_text(encoding="utf-8"))
+    assert summary["benchmark"] == "pointmass-2d-v0.1"
+    assert summary["results"][0]["model_id"] == "starter-linear-state"
+    assert summary["results"][0]["contexts"] == 4
+    assert summary["results"][0]["metrics"]["top1_regret"] == 0.0