Add markproof init, and the files a public repo is expected to have

.github/ISSUE_TEMPLATE/config.yml

@@ -0,0 +1,8 @@
+blank_issues_enabled: true
+contact_links:
+  - name: Security vulnerability
+    url: https://github.com/Tippel-AI/markproof/security/advisories/new
+    about: Report privately. Signing keys and report integrity are in scope — see SECURITY.md.
+  - name: Is my system in scope of Article 50?
+    url: https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai
+    about: markproof is a conformance test, not legal advice. That question is for a lawyer.

.github/ISSUE_TEMPLATE/false-verdict.yml

@@ -0,0 +1,50 @@
+name: A verdict looks wrong
+description: markproof reported PASS, FAIL, WARN or SKIP where you expected something else.
+labels: ["bug"]
+body:
+  - type: markdown
+    attributes:
+      value: |
+        This is the most important kind of report this project can get. A tool that
+        decides compliance questions has one job, and a wrong verdict is the whole
+        failure — especially a **PASS that should have been a FAIL**, which is the
+        outcome nothing downstream will catch.
+  - type: input
+    id: rule
+    attributes:
+      label: Which rule
+      placeholder: MPF-D-001
+    validations: {required: true}
+  - type: dropdown
+    id: direction
+    attributes:
+      label: What happened
+      options:
+        - PASS where I expected FAIL
+        - FAIL where I expected PASS
+        - WARN that fires on every target (carries no information)
+        - SKIP where I expected a verdict
+        - Something else
+    validations: {required: true}
+  - type: textarea
+    id: evidence
+    attributes:
+      label: The report, or the console output
+      description: |
+        `report.json` is ideal — it carries the rulepack digest and the probe URLs,
+        so the run can be reconstructed. Redact anything sensitive; the digests are
+        what matter.
+      render: json
+    validations: {required: true}
+  - type: textarea
+    id: expected
+    attributes:
+      label: What you expected, and why
+      description: If an Article 50 reading is involved, cite it. Disagreeing with the rulepack is a legitimate report.
+    validations: {required: true}
+  - type: input
+    id: version
+    attributes:
+      label: Version
+      description: "`markproof --version`, or the commit"
+    validations: {required: true}

.github/ISSUE_TEMPLATE/rulepack.yml

@@ -0,0 +1,30 @@
+name: A rule reads the law wrong
+description: The rulepack's interpretation of Article 50 is mistaken, outdated, or too broad.
+labels: ["rulepack"]
+body:
+  - type: markdown
+    attributes:
+      value: |
+        Rulepacks are published under CC-BY-4.0 and cite the paragraph they rest on,
+        specifically so this argument can be had in the open. `docs/RULES_SOURCES.md`
+        records the reasoning behind each rule, including the obligations deliberately
+        left unimplemented — worth reading first, it may already say why.
+
+        Legal disagreement is welcome. This project does not claim to be right about
+        Article 50; it claims to be explicit about what it checks.
+  - type: input
+    id: rule
+    attributes:
+      label: Which rule (or which one is missing)
+    validations: {required: true}
+  - type: textarea
+    id: reading
+    attributes:
+      label: The reading you think is correct
+      description: Cite the Article, and the Guidelines paragraph if there is one.
+    validations: {required: true}
+  - type: textarea
+    id: consequence
+    attributes:
+      label: What changes if you are right
+      description: Does a verdict flip? Does a rule stop applying? Does a new one become checkable?

.github/PULL_REQUEST_TEMPLATE.md

@@ -0,0 +1,26 @@
+<!--
+Thanks for contributing. Two things this project cares about more than most:
+
+  * A verdict must never claim more than was measured. If a change makes a check
+    stricter or more confident, say what evidence justifies the extra confidence.
+  * Documentation is part of the surface. A README that promises what the code
+    does not do is a defect here, not a nit.
+-->
+
+## What changes, and why
+
+## Does this change any verdict?
+
+<!-- If yes: which rule, in which direction, and the golden diff that shows it.
+     `pytest -m determinism` will fail until the goldens are regenerated, which is
+     deliberate — a golden diff is a change in what the tool calls conformant, and
+     it belongs in the review. -->
+
+- [ ] No verdict changes
+- [ ] Verdicts change, and the golden diff is in this PR
+
+## Checks
+
+- [ ] `pytest` passes
+- [ ] `ruff check .` and `mypy` pass
+- [ ] Docs updated if the change is user-visible

CHANGELOG.md

@@ -33,6 +33,17 @@ explicitly: the **rulepack format** — what a rulepack file may contain — and
 
 ### Added
 
+- **`markproof init`** writes a starting `markproof.yaml`. The CLI's own docstring
+  had promised this command since M4 and the build did not have it. The scaffold
+  configures one chat probe and leaves media, UI and text marking commented out
+  with the reason each is optional — a config full of settings nobody chose is how
+  people end up running checks they cannot interpret.
+
+- **PDF output is reachable** (#21). `report.formats` accepts `pdf` and
+  `pdf-html`; both renderers existed, were tested, and could not be produced by any
+  invocation of the shipped tool. `report.sign_key` and `report.output_dir` are
+  honoured too — they were validated and then ignored.
+
 - **Targets can declare which Article 50 obligations bind them** (#14). A new
   `applicability` block in `markproof.yaml` maps obligations —
   `ai-interaction`, `synthetic-media-marking`, `synthetic-text-marking`,

CITATION.cff

@@ -0,0 +1,26 @@
+cff-version: 1.2.0
+title: "markproof: deterministic CI checks for EU AI Act Article 50"
+message: "If you cite this tool or its rulepacks, please cite it as below."
+type: software
+authors:
+  - family-names: Friedrich
+    given-names: Lukas
+    affiliation: Tippel
+repository-code: "https://github.com/Tippel-AI/markproof"
+license: Apache-2.0
+abstract: >-
+  markproof calls a running GenAI endpoint the way a user would and checks what
+  arrives: whether an interactive system discloses its artificial nature, whether
+  delivered media still carry a valid C2PA manifest declaring an AI source type,
+  and whether generated text still carries the operator's own watermark. The
+  evaluation path contains no language model, so identical evidence produces
+  identical verdicts, and every run emits a canonically serialised, Ed25519-signed
+  report that a third party can verify offline. Rulepacks are versioned, cite the
+  paragraph they implement, and ship under CC-BY-4.0.
+keywords:
+  - EU AI Act
+  - Article 50
+  - transparency
+  - C2PA
+  - SynthID
+  - conformance testing

CODE_OF_CONDUCT.md

@@ -0,0 +1,60 @@
+<!--
+SPDX-FileCopyrightText: 2026 Lukas Friedrich / Tippel
+SPDX-License-Identifier: Apache-2.0
+-->
+
+# Code of conduct
+
+## The short version
+
+Be straightforward with people and rigorous about claims. Those are different
+things, and this project needs both.
+
+## What is expected
+
+**Argue with the work, not the person.** Disagreement about whether a rule reads
+Article 50 correctly is the most valuable contribution this project can receive,
+and it is welcome sharply. "This rule is wrong because paragraph 117 says X" is
+exactly right. "Whoever wrote this has not read the Act" is not, and it is also
+weaker — it gives the reader nothing to check.
+
+**Say what you actually verified.** This is a tool about not overclaiming, so the
+same standard applies to the conversation around it. "I ran it and it failed" and
+"I think it would fail" are different sentences, and mixing them up costs
+everyone's time.
+
+**Assume the other person read the Act too, and reached a different conclusion.**
+Article 50 is genuinely ambiguous in places. `docs/RULES_SOURCES.md` records where,
+and what this project decided. If you think a decision is wrong, the reasoning is
+written down so you can argue with it specifically.
+
+**Take a correction well, and give one plainly.** Being wrong in public about a
+compliance question is not a character flaw; leaving a wrong claim standing
+because it is uncomfortable to withdraw is worse.
+
+## What is not acceptable
+
+Harassment, personal attacks, demeaning comments about anyone's identity, or
+sustained disruption. Also: knowingly asserting something you have not checked as
+though you had, which in a project like this is a form of harm rather than a
+faux pas.
+
+## Scope
+
+Anywhere the project happens — issues, pull requests, discussions, and in
+communication about the project elsewhere.
+
+## Reporting
+
+Email **lukas.friedrich@tippel.ai**. Reports are read by one person and stay
+confidential. You will get an acknowledgement within a few days, and I will tell
+you what I decided.
+
+Maintainers may remove comments, close threads, or block accounts. That is a
+judgement call by a person, not a process, and I will say why when I make one.
+
+## A note on the maintainer
+
+This project is maintained by one person alongside other work. Response times will
+sometimes be slow, and some good contributions will not be merged because keeping
+this small is itself a design decision. Neither is a slight.

README.md

@@ -11,6 +11,7 @@ markproof calls your *running* AI endpoint the way a user would, and checks what
 
 ```bash
 pipx install git+https://github.com/Tippel-AI/markproof
+markproof init                       # writes a starting markproof.yaml
 markproof run --config markproof.yaml
 ```
 

src/markproof/cli.py

@@ -8,7 +8,13 @@
 ``rules list`` / ``rules schema``
     List the rules of a rulepack; export the pydantic schemas as JSON Schema.
 
-``verify-report``, ``keygen`` and ``init`` arrive with the signed report in M4.
+``verify-report``
+    Check a report's signature. The auditor's command: it needs the report, a
+    public key and this tool, on a machine that never saw the system under test.
+``keygen``
+    Generate an Ed25519 key pair for signing.
+``init``
+    Write a starting ``markproof.yaml``.
 
 The module-level object ``app`` is the console-script target declared in
 ``pyproject.toml`` (``markproof = "markproof.cli:app"``).
@@ -541,6 +547,90 @@ def verify_report_command(
     raise typer.Exit(code=0 if valid else 1)
 
 
+_SCAFFOLD = """\
+# SPDX-License-Identifier: Apache-2.0
+#
+# Written by `markproof init`. Point it at the endpoint your users actually
+# reach — a staging copy proves nothing about production.
+version: 1
+
+target:
+  name: {name}
+  probes:
+    - id: chat
+      type: http-chat
+      url: {url}
+      dialect: openai-chat        # or generic-json + response_path
+      lang: de                    # de | en
+      # auth: {{header: Authorization, env: MARKPROOF_TOKEN}}
+
+# Which Article 50 obligations bind this target. Everything not listed is
+# checked, so silence never removes a rule. A `false` here is a claim, not a mute
+# switch: it travels into the signed report, so a green run states its own scope.
+#
+#   ai-interaction · synthetic-media-marking · synthetic-text-marking
+#   emotion-recognition · deepfake-labelling · public-interest-text
+# applicability:
+#   deepfake-labelling: false
+
+# Verifying text marking needs the parameters you generate with. Without them
+# the text rule skips and says so — it never guesses.
+# text_marking:
+#   method: synthid
+#   watermark_config: secrets/watermark_config.json
+
+rulepack: art50-eu-2026.07
+
+report:
+  formats: [json, summary]        # add `pdf` for the artefact an auditor reads
+  # sign_key: env:MARKPROOF_SIGNING_KEY
+"""
+
+
+@app.command()
+def init(
+    path: Annotated[Path, typer.Option("--config", "-c", help="Where to write the config.")] = Path(
+        "markproof.yaml"
+    ),
+    url: Annotated[
+        str,
+        typer.Option("--url", help="The chat endpoint to probe."),
+    ] = "https://api.example.com/v1/chat/completions",
+    name: Annotated[str, typer.Option("--name", help="A name for the target.")] = "my-service",
+    force: Annotated[bool, typer.Option("--force", help="Overwrite an existing file.")] = False,
+) -> None:
+    """Write a starting markproof.yaml.
+
+    Deliberately small. The scaffold configures one chat probe and leaves the
+    media, UI and text-marking blocks commented out with the reason each is
+    optional, because a config full of settings nobody chose is how people end up
+    running checks they cannot interpret.
+    """
+    if path.exists() and not force:
+        err_console.print(
+            f"[bold red]error:[/] {path} already exists. Pass --force to overwrite it."
+        )
+        raise typer.Exit(code=2)
+
+    try:
+        path.write_text(_SCAFFOLD.format(name=name, url=url), encoding="utf-8")
+    except OSError as exc:
+        err_console.print(f"[bold red]error:[/] could not write {path}: {exc}")
+        raise typer.Exit(code=2) from exc
+
+    console.print()
+    console.print(f"  wrote [cyan]{path}[/cyan]")
+    console.print()
+    console.print("  Next:")
+    console.print(f"    1. set the url in {path} to the endpoint your users reach")
+    console.print(f"    2. [cyan]markproof run --config {path}[/cyan]")
+    console.print()
+    console.print(
+        "  [dim]No endpoint yet? examples/demo-bot is a deliberately non-conformant one.[/dim]"
+    )
+    console.print()
+
+
 @app.command()
 def keygen(
     out_dir: Annotated[