diff --git a/src/backends/cuda/build_script.py b/src/backends/cuda/build_script.py
new file mode 100644
index 0000000..dd5d9b5
--- /dev/null
+++ b/src/backends/cuda/build_script.py
@@ -0,0 +1,227 @@
+#!/usr/bin/env python3
+"""Read a backend build script as text and report what it declares.
+
+``contract.md`` requires ``build.sh`` to accept an output directory and a compute
+capability, and to build exactly the library the manifest names for exactly the
+architectures it names. Checking that means reading the script — never running
+it: CI must not execute arbitrary repository code to decide whether a manifest is
+honest.
+
+The parsing is shell-shaped but deliberately shallow. It evaluates the two forms
+a build script actually uses (``${NAME:-default}`` and ``$NAME``, including
+nesting) and ignores comments, without pretending to be a shell.
+"""
+
+from __future__ import annotations
+
+import os
+import re
+
+# `sm_90a` / `compute_89` / `arch=compute_90` in a build script. Only used to
+# catch a build script that hard-codes one architecture while the manifest
+# claims more; it is not a substitute for reading the script.
+# A `-gencode` flag and its value: `-gencode "arch=compute_89,code=sm_89"`, or with
+# `=`. This is the flag that decides what nvcc actually builds, which is not
+# necessarily what an `arch=` variable says.
+_GENCODE = re.compile(r"-gencode[=\s]+(\"[^\"]*\"|\S+)")
+
+_ARCH_IN_SCRIPT = re.compile(r"\b(?:sm|compute)_(\d+)[af]?\b")
+_BUILD_SCRIPT_ARCH_FIELD = re.compile(r"\b(?:CUDA_COMPUTE_CAP|ARCH|arch)\b")
+_SCRIPT_RESOLVED = re.compile(r"^\s*([A-Za-z_][A-Za-z0-9_]*)\s*=\s*(.+?)\s*$")
+
+_TOKEN = re.compile(r"\$\{[^{}]*\{[^{}]*\}[^{}]*\}|\$\{[^{}]*\}|\$[A-Za-z_]\w*|\$\d")
+
+def _strip_comment(line):
+    """Remove a shell comment, respecting quotes.
+
+    `# defaults to CUDA_COMPUTE_CAP, else 89.` must not be read as an assignment
+    or as a compute capability, so comments are removed before any scanning.
+    """
+    out = []
+    quote = ""
+    for index, char in enumerate(line):
+        if quote:
+            if char == quote:
+                quote = ""
+            out.append(char)
+            continue
+        if char in ("'", '"'):
+            quote = char
+            out.append(char)
+            continue
+        if char == "#" and (index == 0 or line[index - 1].isspace()):
+            break
+        out.append(char)
+    return "".join(out)
+
+
+def _balanced_parameter(value):
+    """Return (name, default) for a value that is exactly one ``${...}``.
+
+    Shell parameter expansions nest (``${2:-${CUDA_COMPUTE_CAP:-89}}``), so the
+    closing brace has to be matched by counting rather than by a regex, which
+    would stop at the inner brace.
+    """
+    if not value.startswith("${"):
+        return None
+    depth = 0
+    for index, char in enumerate(value):
+        if char == "{":
+            depth += 1
+        elif char == "}":
+            depth -= 1
+            if depth == 0:
+                if index != len(value) - 1:
+                    return None  # trailing text, not a lone parameter expansion
+                interior = value[2:index]
+                name, separator, default = interior.partition(":-")
+                return name, (default if separator else "")
+    return None
+
+
+def _resolve_token(token, script_env, depth=0):
+    """Return (value, known) for one shell token inside a larger string.
+
+    ``known`` is False when the token depends on something CI cannot know, such
+    as a positional parameter the release process supplies. An unknown token must
+    not collapse to an empty string: 'we cannot tell' and 'the value is empty'
+    have to stay distinguishable, or an unexpanded path would be read as a
+    filename.
+    """
+    if depth > 8:
+        return "", False
+    value = token.strip().strip('"').strip("'")
+    parameter = _balanced_parameter(value)
+    if parameter is not None:
+        name, default = parameter
+        if name.isdigit() or ":" in name:
+            if default:
+                return _resolve_token(default, script_env, depth + 1)
+            return "", False
+        assigned = script_env.get(name, "").strip()
+        if not assigned:
+            return _resolve_token(default, script_env, depth + 1) if default else ("", False)
+        return _resolve_token(assigned, script_env, depth + 1)
+    for name, assigned in script_env.items():
+        if value == "$" + name:
+            return _resolve_token(assigned, script_env, depth + 1)
+    if value.startswith("$"):
+        return "", False
+    return value, True
+
+
+# A shell expansion inside a larger word. Alternation order matters: the nested
+# form has to be tried before the flat form, and the flat form must stop at the
+# first `}`. `$` is only a parameter start when followed by a name, `{` or a
+# digit, so the literal prefix in `lib_${arch}.so` is not swallowed by `\$\d`.
+def _expand_string(text, script_env):
+    """Substitute every shell token in ``text``, keeping the literal tail.
+
+    ``"$out/libqwen3_5_cuda.so"`` -> ``libqwen3_5_cuda.so``: the prefix is
+    unknowable but the filename is not, and the filename is what the manifest is
+    checked against.
+    """
+    out = []
+    position = 0
+    for match in _TOKEN.finditer(text):
+        out.append(text[position:match.start()])
+        value, _known = _resolve_token(match.group(0), script_env)
+        out.append(value)
+        position = match.end()
+    out.append(text[position:])
+    return "".join(out)
+
+
+def _output_name(text, script_env):
+    """The filename a build script writes, from its last ``-o`` argument.
+
+    The last one wins, for the same reason the last ``arch=`` assignment does: a
+    script that builds more than one target ends with the one a plain invocation
+    produces. The argument may be quoted and may mix a literal prefix with an
+    expansion (``"$out/liblaya_cuda.so"``). Expansion is applied to the word
+    first and quotes are trimmed after, because the closing quote of
+    ``-o "$out/libx.so"`` is only trailing relative to the expanded word.
+    """
+    name = ""
+    for line in text.splitlines():
+        for match in re.finditer(r"(?<!\S)-o\s+(\S+)", line):
+            expanded = _expand_string(match.group(1), script_env)
+            candidate = os.path.basename(expanded).strip().strip('"').strip("'")
+            if candidate:
+                name = candidate
+    return name
+
+
+def gencode_targets(lines, initial_env=None):
+    """Architectures that actually reach nvcc, from the ``-gencode`` flags.
+
+    The environment is tracked line by line, so `arch=89 ... nvcc ... arch=90 ...
+    nvcc` resolves each invocation against the value in force at that point. A
+    literal in the flag wins over any variable, because the literal is what nvcc
+    receives: a script that assigns `arch=${2:-89}` and then writes
+    `-gencode arch=compute_90,code=sm_90` builds sm_90, and reading only the
+    assignment would report a target the compiler never sees.
+    """
+    env = dict(initial_env or {})
+    found = []
+    for line in lines:
+        match = _SCRIPT_RESOLVED.match(line)
+        if match and not _GENCODE.search(line):
+            env[match.group(1)] = match.group(2)
+            continue
+        for flag in _GENCODE.finditer(line):
+            expanded = _expand_string(flag.group(1).strip('"'), env)
+            for value in re.findall(r"(?:sm|compute)_(\d+)", expanded):
+                target = int(value)
+                if target not in found:
+                    found.append(target)
+    return sorted(found)
+
+
+def parse_build_script(text):
+    """Extract the declared architectures and output name from a build script.
+
+    The script is read as text: it is never executed, because CI must not run
+    arbitrary repository code to decide whether a manifest is honest.
+    """
+    lines = [_strip_comment(line) for line in text.splitlines()]
+
+    # Architecture names may be reassigned per target (`arch=89` ... `arch=90`),
+    # so every declaration is collected rather than only the final value. A
+    # manifest claim is about what the script can reach, not about its last
+    # assignment. `env` still keeps the final value, which is what the output
+    # path expands against.
+    env = {}
+    architectures = []
+    for line in lines:
+        match = _SCRIPT_RESOLVED.match(line)
+        if not match:
+            continue
+        name, raw = match.group(1), match.group(2)
+        env[name] = raw
+        if not _BUILD_SCRIPT_ARCH_FIELD.search(name):
+            continue
+        resolved = _expand_string(raw, env).strip()
+        if resolved.isdigit():
+            value = int(resolved)
+            if value not in architectures:
+                architectures.append(value)
+
+    output = _output_name("\n".join(lines), env)
+
+    # Architectures appearing literally anywhere in the script. Kept separate
+    # from `gencode_architectures` because it is the weaker signal: it also
+    # matches a target named in a comment that survived, or in a variable that
+    # is never passed to nvcc.
+    literal = []
+    for match in _ARCH_IN_SCRIPT.finditer("\n".join(lines)):
+        value = int(match.group(1))
+        if value not in literal:
+            literal.append(value)
+
+    gencode = gencode_targets(lines)
+
+    return {"architectures": architectures, "literal_architectures": literal,
+            "gencode_architectures": gencode,
+            "output": output, "env": env}
+
diff --git a/src/backends/cuda/check_contract.py b/src/backends/cuda/check_contract.py
new file mode 100755
index 0000000..c04ee87
--- /dev/null
+++ b/src/backends/cuda/check_contract.py
@@ -0,0 +1,469 @@
+#!/usr/bin/env python3
+"""Tier-1 contract check for CUDA backend libraries. No GPU, no nvcc required.
+
+Discovers every ``src/backends/cuda/<name>/<name>.backend.json`` manifest and
+checks the parts of ``contract.md`` that do not need hardware:
+
+  * the manifest schema,
+  * that every declared source file exists,
+  * that the build script's declared output and architectures match the manifest,
+  * that the ABI version is consistent across backends,
+  * that a ``validated`` backend declares a tolerance and a reference entrypoint,
+  * that a kernel requiring a newer compute capability than the build declares is
+    flagged.
+
+It does not compile CUDA and does not prove numerics. Tier 2 (``--gpu`` in CI)
+runs the reference entrypoint on a self-hosted GPU runner.
+
+Usage:
+    python3 src/backends/cuda/check_contract.py [--repo-root PATH] [--json]
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+import os
+import re
+import sys
+
+from build_script import parse_build_script
+
+MANIFEST_SUFFIX = ".backend.json"
+BACKENDS_DIR = os.path.join("src", "backends", "cuda")
+STATUSES = ("planned", "experimental", "validated")
+# Every key check_manifest reads directly. The missing-key report and the guard
+# that stops after it are both derived from this tuple.
+REQUIRED_KEYS = ("name", "abi_version", "status", "sources", "build")
+CONTRACT_ABI_VERSION = 1
+# `<PREFIX>_ABI_VERSION N` in a backend header. #19 uses `CS1_ABI_VERSION`,
+# bumped whenever the C interface changes.
+_ABI_MACRO = re.compile(r"^[ \t]*#[ \t]*define[ \t]+(\w*ABI_VERSION)[ \t]+(\d+)[ \t]*$", re.M)
+
+
+
+class Issue(object):
+    """One contract violation or warning for a backend."""
+
+    def __init__(self, level, backend, message):
+        self.level = level  # "error" or "warning"
+        self.backend = backend
+        self.message = message
+
+    def __str__(self):
+        return "[%s] %s: %s" % (self.level, self.backend, self.message)
+
+    def as_dict(self):
+        return {"level": self.level, "backend": self.backend, "message": self.message}
+
+
+
+def load_manifests(repo_root):
+    """Return (manifests, issues) for every ``*.backend.json`` under backends/cuda."""
+    issues = []
+    manifests = []
+    root = os.path.join(repo_root, BACKENDS_DIR)
+    if not os.path.isdir(root):
+        issues.append(Issue("error", "-", "%s does not exist" % BACKENDS_DIR))
+        return manifests, issues
+
+    # A backend may be a subdirectory named after itself, or the cuda directory
+    # itself when its files sit directly under it (`kernels/`, `tools/`). The
+    # layout is the model author's call; only the manifest's contents are fixed.
+    found = []
+    for entry in sorted(os.listdir(root)):
+        directory = os.path.join(root, entry)
+        if not os.path.isdir(directory) or entry.startswith("."):
+            continue
+        expected = os.path.join(directory, entry + MANIFEST_SUFFIX)
+        if os.path.isfile(expected):
+            found.append((entry, directory, expected))
+    for filename in sorted(os.listdir(root)):
+        if filename.endswith(MANIFEST_SUFFIX) and os.path.isfile(os.path.join(root, filename)):
+            found.append((filename[: -len(MANIFEST_SUFFIX)], root,
+                          os.path.join(root, filename)))
+
+    # A directory holding `.cu` files with no manifest either way is a backend
+    # someone forgot to declare. Directories that belong to a declared backend
+    # (its `kernels/`, `tools/`) or hold a manifest of their own are not.
+    declared_roots = {directory for _entry, directory, _path in found}
+    for entry in sorted(os.listdir(root)):
+        directory = os.path.join(root, entry)
+        if not os.path.isdir(directory) or entry.startswith("."):
+            continue
+        if any(directory == root_of or directory.startswith(root_of + os.sep)
+               for root_of in declared_roots):
+            continue
+        contents = [f for f in os.listdir(directory) if not f.startswith(".")]
+        present = sorted(f for f in contents if f.endswith(MANIFEST_SUFFIX))
+        if present:
+            # A manifest is here but not under the name discovery looks for, so
+            # nothing loaded it. Continuing silently meant a typo in the filename
+            # bypassed every check for that backend: exit 0, `checked: []`, no
+            # errors and no warnings.
+            issues.append(Issue(
+                "error", entry,
+                "has %s but not %s, so it is never discovered; rename it or fix the "
+                "directory name" % (", ".join(present), entry + MANIFEST_SUFFIX)))
+            continue
+        kernel_like = [f for f in contents if f.endswith((".cu", ".cuh"))]
+        if kernel_like:
+            issues.append(Issue(
+                "error", entry,
+                "has kernel sources (%s) but no %s manifest or parent backend manifest"
+                % (", ".join(sorted(kernel_like)[:3]), entry + MANIFEST_SUFFIX)))
+
+    for entry, directory, expected in found:
+        try:
+            with open(expected, "r", encoding="utf-8") as handle:
+                manifest = json.load(handle)
+        except ValueError as error:
+            issues.append(Issue("error", entry, "%s is not valid JSON: %s"
+                                % (os.path.basename(expected), error)))
+            continue
+        if not isinstance(manifest, dict):
+            issues.append(Issue("error", entry, "%s must contain a JSON object"
+                                % os.path.basename(expected)))
+            continue
+        manifest["_directory"] = directory
+        manifest["_entry"] = entry
+        manifests.append(manifest)
+    return manifests, issues
+
+
+def check_manifest(manifest, issues, repo_root=None):
+    backend = manifest.get("_entry", "?")
+    directory = manifest.get("_directory", "")
+    if repo_root is None:
+        # Best effort when called directly: four levels up from
+        # <repo>/src/backends/cuda/<name>. Callers that know the root should pass
+        # it, because the flat layout makes this guess wrong.
+        repo_root = os.path.dirname(os.path.dirname(os.path.dirname(
+            os.path.dirname(os.path.abspath(directory or ".")))))
+
+    # These were two hand-written lists that had to agree, and they drifted:
+    # `name` was reported as missing and then read anyway, so a manifest missing
+    # only `name` raised KeyError out of check_manifest and aborted the checks of
+    # every other backend, in both the text and --json paths. Deriving the guard
+    # from the same constant is what stops that recurring.
+    missing = [key for key in REQUIRED_KEYS if key not in manifest]
+    for key in missing:
+        issues.append(Issue("error", backend, "manifest is missing required key %r" % key))
+    if missing:
+        return
+
+    name = manifest["name"]
+    if not isinstance(name, str) or not name:
+        issues.append(Issue("error", backend, "name must be a non-empty string"))
+    elif name != backend:
+        issues.append(Issue("error", backend,
+                            "name %r does not match directory name %r" % (name, backend)))
+
+    check_abi_version(manifest, issues)
+
+    status = manifest["status"]
+    if status not in STATUSES:
+        issues.append(Issue("error", backend,
+                            "status must be one of %s, got %r" % (", ".join(STATUSES), status)))
+        return
+
+    sources = manifest["sources"]
+    if not isinstance(sources, list) or not all(isinstance(item, str) for item in sources):
+        issues.append(Issue("error", backend, "sources must be a list of strings"))
+    else:
+        for source in sources:
+            if os.path.isabs(source) or ".." in source.split("/"):
+                issues.append(Issue("error", backend,
+                                    "source %r must be relative to the backend directory" % source))
+                continue
+            if not os.path.isfile(os.path.join(directory, source)):
+                issues.append(Issue("error", backend, "declared source %r does not exist" % source))
+
+    build = manifest["build"]
+    if not isinstance(build, dict):
+        issues.append(Issue("error", backend, "build must be an object"))
+        return
+
+    for key in ("script", "output"):
+        if not isinstance(build.get(key), str) or not build.get(key):
+            issues.append(Issue("error", backend, "build.%s must be a non-empty string" % key))
+    architectures = build.get("architectures")
+    if not isinstance(architectures, list) or not architectures \
+            or not all(isinstance(item, int) and 50 <= item <= 200 for item in architectures):
+        issues.append(Issue("error", backend,
+                            "build.architectures must be a non-empty list of compute "
+                            "capabilities, e.g. [89, 90]"))
+        architectures = None
+    default_arch = build.get("default_arch")
+    if default_arch is not None and architectures is not None and default_arch not in architectures:
+        issues.append(Issue("error", backend,
+                            "build.default_arch %r is not listed in build.architectures %r"
+                            % (default_arch, architectures)))
+
+    # What the kernels require is currently stated only in comments and READMEs
+    # ("tensor-core kernels need sm_80 or newer"). Declaring it makes the claim
+    # checkable here rather than discoverable as a build failure on someone
+    # else's GPU.
+    min_capability = build.get("min_capability")
+    if min_capability is not None:
+        if not isinstance(min_capability, int):
+            issues.append(Issue("error", backend,
+                                "build.min_capability must be an integer compute capability, "
+                                "e.g. 80"))
+        elif architectures is not None:
+            too_low = [item for item in architectures if item < min_capability]
+            if too_low:
+                issues.append(Issue("error", backend,
+                                    "build.architectures includes %r, below build.min_capability "
+                                    "%d; the kernels would not build for that target"
+                                    % (too_low, min_capability)))
+
+    # A backend can serve more than one model. Kev and Cua-S1 share the same
+    # Qwen3.5 backbone, so listing consumers is what makes reuse visible instead
+    # of a private arrangement between two PRs.
+    models = manifest.get("models")
+    if models is not None:
+        if not isinstance(models, list) or not all(isinstance(item, str) for item in models):
+            issues.append(Issue("error", backend, "models must be a list of strings"))
+        else:
+            for model in models:
+                if not model.endswith("/") or os.path.isabs(model) or ".." in model.split("/"):
+                    issues.append(Issue("error", backend,
+                                        "models entry %r must be a repository-relative directory "
+                                        "path ending in '/'" % model))
+                elif not os.path.isdir(os.path.join(repo_root, model)):
+                    issues.append(Issue("warning", backend,
+                                        "models entry %r does not exist yet; the consumer engine "
+                                        "is not in the tree" % model))
+
+    if status == "planned":
+        return
+
+    # A non-planned backend must ship the build script it names.
+    script_path = ""
+    if isinstance(build.get("script"), str):
+        script_path = os.path.join(directory, build["script"])
+        if not os.path.isfile(script_path):
+            issues.append(Issue("error", backend,
+                                "build.script %r does not exist" % build["script"]))
+        else:
+            _check_build_script(backend, script_path, build, architectures, issues)
+            # The compile job runs the script directly, so the execute bit must
+            # be set in git, not only in a local working copy: a fresh clone is
+            # what CI checks out.
+            if os.name == "posix" and not os.access(script_path, os.X_OK):
+                issues.append(Issue("error", backend,
+                                    "build.script %r is not executable on the checked-out tree "
+                                    "(CI takes that bit from the committed one); the compile job "
+                                    "runs it as ./%s. Fix with: git update-index --chmod=+x %s"
+                                    % (build["script"], build["script"], script_path)))
+
+    numerics = manifest.get("numerics")
+    reference = manifest.get("reference")
+    if status == "validated":
+        if not isinstance(numerics, dict) or not isinstance(numerics.get("tolerance"), dict) \
+                or not numerics["tolerance"]:
+            issues.append(Issue("error", backend,
+                                "status is validated but numerics.tolerance is missing; a "
+                                "parity claim needs a tolerance declared before comparison"))
+        if not isinstance(reference, dict) or not reference.get("entrypoint"):
+            issues.append(Issue("error", backend,
+                                "status is validated but reference.entrypoint is missing"))
+    if isinstance(reference, dict) and reference.get("entrypoint"):
+        entrypoint = reference["entrypoint"]
+        # Type first. `os.path.isabs` and `split` raise TypeError on a list or a
+        # number, and that exception escaped check_manifest: the CLI printed no
+        # structured report at all, not even under --json, and every later
+        # backend went unchecked.
+        if not isinstance(entrypoint, str):
+            issues.append(Issue("error", backend,
+                                "reference.entrypoint must be a string, got %s"
+                                % type(entrypoint).__name__))
+            entrypoint = ""
+        elif os.path.isabs(entrypoint) or ".." in entrypoint.split("/"):
+            issues.append(Issue("error", backend,
+                                "reference.entrypoint must be repository-relative"))
+        elif entrypoint:
+            manifest["_repo_relative_entrypoint"] = entrypoint
+            _check_reference_entrypoint(backend, entrypoint, issues, repo_root)
+    if isinstance(numerics, dict) and isinstance(numerics.get("tolerance"), dict):
+        for key, value in numerics["tolerance"].items():
+            if key == "note":
+                continue
+            if not isinstance(value, (int, float)):
+                issues.append(Issue("error", backend,
+                                    "numerics.tolerance.%s must be a number or a note, got %r"
+                                    % (key, value)))
+
+
+def _check_reference_entrypoint(backend, entrypoint, issues, repo_root):
+    """A declared reference that is not there cannot be run by Tier 2."""
+    if not os.path.isfile(os.path.join(repo_root, entrypoint)):
+        issues.append(Issue("warning", backend,
+                            "reference.entrypoint %r does not exist yet; the Tier-2 GPU job "
+                            "cannot run parity for this backend" % entrypoint))
+
+
+def _check_build_script(backend, script_path, build, architectures, issues):
+    try:
+        with open(script_path, "r", encoding="utf-8") as handle:
+            text = handle.read()
+    except OSError as error:
+        issues.append(Issue("error", backend, "cannot read %s: %s" % (build.get("script"), error)))
+        return
+
+    parsed = parse_build_script(text)
+
+    declared_output = build.get("output")
+    if parsed["output"] and declared_output and parsed["output"] != declared_output:
+        issues.append(Issue("error", backend,
+                            "build.output %r does not match the %r the build script writes"
+                            % (declared_output, parsed["output"])))
+
+    if architectures is None:
+        return
+
+    # What nvcc is actually told to build, from the -gencode flags. This is the
+    # authority: a variable that never reaches one of those flags does not make a
+    # target reachable. A script that assigns `arch=${2:-89}` and then writes
+    # `-gencode arch=compute_90,code=sm_90` builds sm_90 only, so a manifest
+    # declaring sm_89 is false even though an `arch=` variable says 89.
+    gencode = parsed.get("gencode_architectures") or []
+    by_variable = parsed["architectures"]
+    if gencode and by_variable:
+        unused = [item for item in by_variable if item not in gencode]
+        if unused:
+            issues.append(Issue("error", backend,
+                                "build script assigns %r but only passes %r to nvcc; the "
+                                "assigned value never reaches a -gencode flag, so targets "
+                                "read from variables alone cannot be claimed"
+                                % (by_variable, gencode)))
+    script_arch = gencode or by_variable or parsed["literal_architectures"]
+    if not script_arch:
+        issues.append(Issue("warning", backend,
+                            "build script declares no compute capability; the manifest claims "
+                            "%r but CI cannot confirm the script honours it" % (architectures,)))
+        return
+    unbuildable = [item for item in architectures if item not in script_arch]
+    if unbuildable:
+        issues.append(Issue("error", backend,
+                            "build.architectures claims %r but the build script only reaches %r "
+                            "(missing %r)" % (architectures, script_arch, unbuildable)))
+    extra = [item for item in script_arch if item not in architectures]
+    if extra:
+        issues.append(Issue("warning", backend,
+                            "build script also targets %r, which build.architectures omits"
+                            % (extra,)))
+
+
+def _declared_abi(manifest):
+    """Every ``#define *ABI_VERSION N`` in the backend's declared sources.
+
+    Read as text, like the build script: the checker must not need a compiler,
+    and the macro is a fact the source already states.
+    """
+    directory = manifest.get("_directory", "")
+    found = []
+    for source in manifest.get("sources") or []:
+        if not isinstance(source, str):
+            continue
+        path = os.path.join(directory, source)
+        if not os.path.isfile(path):
+            continue        # a missing source is already reported separately
+        try:
+            with open(path, "r", encoding="utf-8", errors="replace") as handle:
+                text = handle.read()
+        except OSError:
+            continue
+        for match in _ABI_MACRO.finditer(text):
+            found.append((os.path.basename(source), match.group(1), int(match.group(2))))
+    return found
+
+
+def check_abi_version(manifest, issues):
+    """``abi_version`` is the library's own, and its header is the authority.
+
+    This is the one manifest field whose source of truth lives in the code, and
+    the checker could not see it before: nothing read the header, so the only
+    guard against a stale manifest was the example in contract.md. #19's
+    ``ops.h`` says ``CS1_ABI_VERSION 4``; a manifest claiming 1 would have been
+    accepted.
+
+    Two libraries may legitimately differ, so there is deliberately no
+    cross-backend sameness requirement: what has to hold is that each manifest
+    agrees with its own header.
+    """
+    backend = manifest.get("_entry", "?")
+    abi = manifest.get("abi_version")
+    if not isinstance(abi, int) or abi < CONTRACT_ABI_VERSION:
+        issues.append(Issue("error", backend,
+                            "abi_version must be an integer >= %d, got %r"
+                            % (CONTRACT_ABI_VERSION, abi)))
+        return
+
+    found = _declared_abi(manifest)
+    if not found:
+        issues.append(Issue("warning", backend,
+                            "no #define *ABI_VERSION in the declared sources, so abi_version "
+                            "%d cannot be checked against the library" % abi))
+        return
+
+    versions = {version for _file, _macro, version in found}
+    if len(versions) > 1:
+        issues.append(Issue("error", backend,
+                            "the declared sources define more than one ABI version: %s"
+                            % ", ".join("%s %s=%d" % entry for entry in sorted(found))))
+        return
+
+    header_version = versions.pop()
+    if header_version != abi:
+        file_name, macro, _value = found[0]
+        issues.append(Issue("error", backend,
+                            "abi_version is %d but %s defines %s %d; the manifest and the "
+                            "library it describes have to agree"
+                            % (abi, file_name, macro, header_version)))
+
+
+def run(repo_root):
+    manifests, issues = load_manifests(repo_root)
+    for manifest in manifests:
+        check_manifest(manifest, issues, repo_root)
+    return manifests, issues
+
+
+def main(argv=None):
+    parser = argparse.ArgumentParser(description=__doc__.splitlines()[0])
+    parser.add_argument("--repo-root", default=".",
+                        help="repository root to check (default: current directory)")
+    parser.add_argument("--json", action="store_true", help="emit machine-readable results")
+    args = parser.parse_args(argv)
+
+    repo_root = os.path.abspath(args.repo_root)
+    manifests, issues = run(repo_root)
+    errors = [issue for issue in issues if issue.level == "error"]
+    warnings = [issue for issue in issues if issue.level == "warning"]
+
+    if args.json:
+        json.dump({
+            "checked": sorted(m.get("_entry", "?") for m in manifests),
+            "issues": [issue.as_dict() for issue in issues],
+            "errors": len(errors),
+            "warnings": len(warnings),
+        }, sys.stdout, indent=2)
+        sys.stdout.write("\n")
+    else:
+        for issue in issues:
+            print(issue)
+        for manifest in manifests:
+            print("checked %s: status=%s abi=%s"
+                  % (manifest.get("_entry"), manifest.get("status"), manifest.get("abi_version")))
+        print("%d backend(s), %d error(s), %d warning(s)"
+              % (len(manifests), len(errors), len(warnings)))
+        if not manifests:
+            print("no backends declared; add src/backends/cuda/<name>/<name>.backend.json")
+
+    return 1 if errors else 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())
diff --git a/src/backends/cuda/contract.md b/src/backends/cuda/contract.md
new file mode 100644
index 0000000..bfc80ba
--- /dev/null
+++ b/src/backends/cuda/contract.md
@@ -0,0 +1,249 @@
+# CUDA backend contract
+
+`src/backends/cuda/` is shared by several model engines. Each model owns its own
+operations, its own library and its own numerics; this document fixes only the
+parts that have to agree for two model libraries to be built, distributed and
+validated the same way.
+
+The rule from the [repository layout](../../../README.md) still holds: model
+orchestration, batching policy, state management and kernel selection stay with
+the model engine. A backend does not need identical internal structures to
+another backend, and no universal tensor abstraction is introduced here.
+
+Status: proposed. The checker is [`check_contract.py`](check_contract.py); wiring it
+into CI is a separate change, so nothing here is enforced yet.
+
+## Why this exists
+
+Three CUDA efforts are now in flight, each with its own build path:
+
+| Effort | Kernels | Produces | Built by |
+| --- | --- | --- | --- |
+| Laya native (#14) | TileLang, exported to CUDA | `liblaya_cuda.so` | `tools/export.py` + `tools/build.py` |
+| Cua-S1 native (#19) | CUDA C++ | `libqwen3_5_cuda.so` | `build.sh` |
+| Cua-S1 multimodal (#12) | Triton / cuTile | (Python) | — |
+
+That is three build systems, three library names and three C ABIs. None of it
+conflicts today, because nothing links against anything else yet. It conflicts
+as soon as one model reuses another's kernels — which is already planned: the
+Cua-S1 native worker and a Kev engine share a Qwen3.5 base, and `#9` asks for
+CUDA paths for CLM as well.
+
+`#6` says to extract shared code once two implementations exist. Two exist now.
+This document is that extraction, kept to interfaces only.
+
+## 1. Artifact discovery
+
+Every CUDA backend declares itself in one JSON file:
+
+```text
+src/backends/cuda/<name>/<name>.backend.json
+```
+
+The checker discovers backends by that path, so adding a backend needs no CI
+change. Schema:
+
+```json
+{
+  "name": "qwen3_5",
+  "abi_version": 4,
+  "status": "validated",
+  "sources": ["common.cuh", "mma.cuh", "ops.h", "norm.cu", "elementwise.cu",
+              "attention.cu", "gdn_prefill.cu", "gemm.cu", "runtime.cu"],
+  "models": ["src/models/cua_s1/", "src/models/kev/"],
+  "build": {
+    "script": "build.sh",
+    "output": "libqwen3_5_cuda.so",
+    "default_arch": 89,
+    "architectures": [89],
+    "min_capability": 80
+  },
+  "numerics": {
+    "precision": "bfloat16",
+    "accumulation": "float32",
+    "tolerance": { "max_abs": 0.039, "note": "float32 reference, see #11" }
+  },
+  "reference": {
+    "entrypoint": "recipe/cua_s1/check_native.py",
+    "note": "pinned upstream FourBModel; kernel tolerances are declared per suite"
+  }
+}
+```
+
+Required keys are `name`, `abi_version`, `status`, `sources` and `build`.
+`status` is one of:
+
+- `planned` — directory only; the checker skips everything else.
+- `experimental` — builds, but no parity claim yet. `numerics.tolerance` may be
+  omitted.
+- `validated` — a parity claim is made. `numerics.tolerance` and
+  `reference.entrypoint` are required, and the checker fails without them.
+
+`reference.entrypoint` is a repository-relative script that runs on a GPU and
+exits non-zero when parity fails. It takes no contract-defined arguments: a
+self-hosted runner that has the weights and the GPU runs it directly.
+
+### One backend, several models
+
+`models` lists the model engines that consume this backend, as
+repository-relative directories. It is optional, and it exists because reuse is
+the point: `#19`'s kernels serve a Qwen3.5-4B backbone, and Kev from `#9` uses
+the same backbone with a different adapter and readout, so one backend directory
+serves both. Without this field that sharing is a private arrangement between
+two PRs and invisible to anyone reading either one. An entry naming a directory
+that is not in the tree is a warning, not an error, so a backend can be merged
+before its second consumer lands.
+
+### Build script interface
+
+`build.script` is invoked by the compile job as:
+
+```sh
+./<build.script> <output-dir> <compute-capability>
+```
+
+so it must be executable, accept an output directory as `$1`, and accept a
+compute capability as `$2` — defaulting to `build.default_arch` when `$2` is
+absent. `#19`'s `build.sh` already has this shape:
+
+```sh
+out=${1:?usage: build.sh <output dir> [compute capability]}
+arch=${2:-${CUDA_COMPUTE_CAP:-89}}
+```
+
+This is a deliberate narrowing, and it is the first place the two in-flight
+build paths diverge. `#19` builds from a shell script that emits one library.
+`#14` builds through `tools/export.py` and `tools/build.py`, which are Python
+programs driven by a TileLang export step and take a bundle directory, not an
+output directory plus an architecture. A backend of that shape satisfies the
+contract with a small `build.sh` wrapper that documents the real invocation
+rather than by CI growing a second code path. Making that wrapper the required
+entry point keeps CI one path and makes "how do I build this backend" answerable
+from the manifest alone.
+
+## 2. C ABI
+
+A backend library exports a flat `extern "C"` interface and is loaded at run
+time, so a Rust engine builds without a CUDA toolkit. `#19`'s `ops.h` is the
+reference for the shape. The contract fixes four points:
+
+1. **One version symbol per library, and the manifest repeats it.** A backend
+   defines `<PREFIX>_ABI_VERSION` in a header and exports
+   `uint32_t <prefix>_abi_version(void)` returning it; a loader refuses a library
+   whose value it does not know. The manifest's `abi_version` is **that number**,
+   not a version of this document, and the checker reads the macro out of the
+   declared sources and requires the two to agree. `#19`'s `ops.h` says
+   `CS1_ABI_VERSION 4` today, so a `qwen3_5` manifest declares 4.
+
+   Two backends may declare different values. They are independent libraries, and
+   the repository layout says CUDA and Metal implementations need not share
+   internal structure; forcing one number across models would invent a coupling
+   nothing needs. What must hold is that each manifest matches its own header.
+2. **Errors are `int`, not exceptions.** Every entry point returns `0` on
+   success. CUDA runtime errors are returned as `cudaError_t` values; anything
+   the library defines itself starts at `1000`. Every library exports
+   `const char* <prefix>_error_string(int)`.
+3. **Work is queued on a caller-supplied stream.** Operations take a
+   `cudaStream_t` as their last argument and must not synchronize internally, so
+   that a caller can capture them into a CUDA Graph. Allocations, copies, stream
+   creation and graph capture are exported by the library too, so a caller needs
+   no direct CUDA linkage.
+4. **GEMM algorithm choice is explicit and reproducible.** A tuned plan is
+   exported and imported rather than re-tuned at load. A plan that reduces
+   split-K in place must be refused, so a given plan always produces the same
+   result.
+
+Point 3 is what makes two model libraries composable: a model engine that
+already owns a stream and a captured graph can call into either library.
+
+### Runtime symbol sharing
+
+`#19`'s runtime block (`malloc`, `free`, `stream_create`, `stream_sync`,
+`upload`, `download`, `graph_begin`, `graph_end`, `graph_launch`,
+`graph_destroy`, `device_info`, `set_device`) is the same work every backend
+needs. Implementing it per library means two libraries loaded into one process
+each carry a copy, and a graph captured through one cannot be launched through
+the other.
+
+Extracting that set into one shared header is a follow-up, deliberately not part
+of this change: there is only one backend today, so the shape of the shared
+surface would be guesswork. It becomes worth doing when a second backend
+actually needs to be loaded alongside the first, and the interface can be
+written against two real callers instead of an imagined one.
+
+## 3. Numerics
+
+Bit-exactness is not claimed anywhere, so the contract records what varies
+instead of leaving it to be discovered in a parity failure:
+
+- **Rounding points.** Where each kernel rounds to the storage dtype is part of
+  the kernel's documentation. `#19`'s `qwen3_5/README.md` is the model: it names
+  the kernels that round where the Transformers reference rounds, and the
+  attention and Gated DeltaNet paths that keep bfloat16 intermediates as
+  FlashAttention and flash-linear-attention do.
+- **Accumulation.** Float32 unless documented otherwise.
+- **Tolerance.** Declared before comparison, per suite, in
+  `numerics.tolerance`. A single end-to-end number is not enough: an end-to-end
+  check cannot localize a failure to one kernel.
+- **Architecture.** `build.architectures` lists the compute capabilities the
+  library is built for. A kernels file that requires a newer capability than the
+  build script's default is a contract error — `#19`'s `mma.cuh` and
+  `gdn_prefill.cu` both say "sm_80 and later" in their first line, while `#14`
+  builds sm_90a only.
+- **Kernel requirement.** `build.min_capability` states the oldest compute
+  capability the kernels actually compile for. It exists because that fact is
+  currently only in prose — `#19`'s `gdn_prefill.cu` and `mma.cuh` both say
+  "sm_80 and later" in a header comment, and its README says "Tensor-core kernels
+  need sm_80 or newer". One integer turns that into something CI checks: every
+  entry in `build.architectures` must be at least `build.min_capability`, so a
+  library cannot claim a target its own kernels reject. The field is optional;
+  a backend that omits it declares no verified floor, which is honest but not
+  checked.
+
+## 4. Validation
+
+Two tiers, because they have different requirements.
+
+**Tier 1 — contract check (no GPU).** `check_contract.py` reads the manifests and
+the build scripts and checks: the schema, that every declared source exists, that
+a build script declares the architectures the manifest claims and writes the
+library the manifest names, that an ABI version is consistent across backends, and
+that a `validated` backend declares both a tolerance and a reference entrypoint.
+It runs on a stock runner and needs no CUDA toolkit. This is the tier implemented
+in this change.
+
+It does not compile CUDA and does not prove numerics.
+
+**Tier 2 — compile (no GPU).** Building each backend with `nvcc` in a CUDA
+container, so a kernel that stops compiling cannot land. The interface it relies
+on is the one Tier 1 already checks: `./<build.script> <output-dir>
+<compute-capability>`, producing exactly the declared library. Not yet wired up;
+it needs a container image and belongs with the rest of the CI wiring rather than
+with the checker.
+
+Note what a build check does and does not prove. It shows the sources compile and
+that a file of the declared name appears. It does not prove the binary honoured
+`build.architectures`: `#19`'s `build.sh` advertises that it "also embeds PTX",
+and nothing reads the artifact to confirm which targets are inside. Inspecting
+that would need `cuobjdump`, which is a reasonable follow-up rather than
+something this change claims to do.
+
+**Tier 3 — GPU parity (self-hosted).** Running each validated backend's
+`reference.entrypoint` on a real GPU, and reporting kernel-level and end-to-end
+parity separately. Contributors with a card can attach their own runner to their
+own fork, which is how `#19` (RTX 6000 Ada, sm_89), `#12` (RTX 4090) and `#14`
+(H800, sm_90a) can each be checked on the hardware they were measured on. Not yet
+wired up.
+
+The split is the point. A CPU-only workflow that claims to validate CUDA is worse
+than no workflow, which is the gap `#14`'s own validation report already records:
+today's CI checks every Rust feature and compiles no CUDA at all. Tier 1 removes
+the part of that gap which does not need hardware; Tiers 2 and 3 need hardware and
+are therefore separate.
+
+## Reporting
+
+Kernel-level and end-to-end results stay separate, and both report the hardware,
+driver, CUDA version, compute capability, source revision and tolerance that
+applied. Load, warmup and warm latency are reported separately. A run that does
+not establish an improvement reports that, rather than an acceleration claim.
diff --git a/src/backends/cuda/tests/__init__.py b/src/backends/cuda/tests/__init__.py
new file mode 100644
index 0000000..e69de29
diff --git a/src/backends/cuda/tests/test_build_script.py b/src/backends/cuda/tests/test_build_script.py
new file mode 100644
index 0000000..1d04fd5
--- /dev/null
+++ b/src/backends/cuda/tests/test_build_script.py
@@ -0,0 +1,138 @@
+#!/usr/bin/env python3
+"""Tests for reading a build script as text.
+
+A parser validated only against invented samples proves nothing about the scripts
+it has to read, so the main case here is the real ``qwen3_5/build.sh`` added by
+#19, and the real ``build.sh`` the Laya backend wrapper exposes.
+
+Nothing is executed: these are the parsing rules CI uses to decide whether a
+manifest is honest.
+"""
+
+from __future__ import annotations
+
+import os
+import sys
+import unittest
+
+sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
+
+import build_script  # noqa: E402  (path is set above)
+
+# The argument handling of src/backends/cuda/qwen3_5/build.sh in #19: `$1` is the
+# output directory, `$2` defaults through CUDA_COMPUTE_CAP to 89.
+QWEN3_5_BUILD_SCRIPT = """#!/usr/bin/env bash
+# Build libqwen3_5_cuda.so from the kernels in this directory.
+#
+#   src/backends/cuda/qwen3_5/build.sh <output dir> [compute capability, e.g. 89]
+set -euo pipefail
+here=$(cd "$(dirname "$0")" && pwd)
+out=${1:?usage: build.sh <output dir> [compute capability]}
+arch=${2:-${CUDA_COMPUTE_CAP:-89}}
+mkdir -p "$out"
+"$nvcc" -O3 -std=c++17 -gencode "arch=compute_${arch},code=[sm_${arch},compute_${arch}]" \\
+    -shared -Xcompiler -fPIC -I"$here" "$here"/*.cu \\
+    "${link[@]}" -o "$out/libqwen3_5_cuda.so"
+echo "built $out/libqwen3_5_cuda.so for sm_${arch}"
+"""
+
+
+class StripCommentTest(unittest.TestCase):
+    def test_a_trailing_comment_is_removed(self):
+        self.assertEqual(build_script._strip_comment("arch=89 # default"), "arch=89 ")
+
+    def test_a_hash_inside_quotes_is_kept(self):
+        self.assertEqual(build_script._strip_comment('x="a # b"'), 'x="a # b"')
+
+    def test_a_comment_at_the_start_is_removed(self):
+        self.assertEqual(build_script._strip_comment("# sm_80 and later"), "")
+
+    def test_a_hash_not_preceded_by_space_is_kept(self):
+        # Not a comment in shell: `a#b` is one word.
+        self.assertEqual(build_script._strip_comment("a#b"), "a#b")
+
+
+class ParseBuildScriptTest(unittest.TestCase):
+    def test_reads_the_real_qwen3_5_script(self):
+        parsed = build_script.parse_build_script(QWEN3_5_BUILD_SCRIPT)
+        self.assertEqual(parsed["output"], "libqwen3_5_cuda.so")
+        self.assertIn(89, parsed["architectures"])
+        # `compute_${arch}` is a template, not a literal, so no architecture
+        # number is invented from it.
+        self.assertEqual(parsed["literal_architectures"], [])
+
+    def test_literal_gencode_is_read_when_no_variable_is_used(self):
+        parsed = build_script.parse_build_script(
+            "nvcc -gencode arch=compute_90,code=sm_90 -shared -o libx.so a.cu\n")
+        self.assertEqual(parsed["literal_architectures"], [90])
+        self.assertEqual(parsed["output"], "libx.so")
+
+    def test_ignores_architectures_in_comments(self):
+        parsed = build_script.parse_build_script(
+            "# works on sm_120\nout=x\nnvcc -o liby.so a.cu\n")
+        self.assertEqual(parsed["literal_architectures"], [])
+        self.assertEqual(parsed["output"], "liby.so")
+
+    def test_a_quoted_output_path_with_a_variable_prefix(self):
+        # `-o "$out/libx.so"`: the directory is unknowable, the filename is not.
+        parsed = build_script.parse_build_script(
+            'out=${1:?usage}\nnvcc -shared -o "$out/libqwen3_5_cuda.so" ./*.cu\n')
+        self.assertEqual(parsed["output"], "libqwen3_5_cuda.so")
+
+    def test_nested_defaults_are_expanded(self):
+        parsed = build_script.parse_build_script(
+            "arch=${2:-${CUDA_COMPUTE_CAP:-89}}\nnvcc -gencode \"x=sm_${arch}\" -o l.so a.cu\n")
+        self.assertEqual(parsed["architectures"], [89])
+
+    def test_every_declaration_is_collected_not_only_the_last(self):
+        # A script that builds two targets assigns arch twice; both count, since
+        # the claim is about what the script can reach.
+        parsed = build_script.parse_build_script(
+            "arch=90\nnvcc -o a.so x.cu\narch=89\nnvcc -o a.so x.cu\n")
+        self.assertEqual(sorted(parsed["architectures"]), [89, 90])
+
+    def test_a_script_declaring_nothing_reports_nothing(self):
+        parsed = build_script.parse_build_script("#!/usr/bin/env bash\nnvcc -o l.so ./*.cu\n")
+        self.assertEqual(parsed["architectures"], [])
+        self.assertEqual(parsed["literal_architectures"], [])
+        self.assertEqual(parsed["output"], "l.so")
+
+    def test_the_last_output_wins(self):
+        parsed = build_script.parse_build_script(
+            "nvcc -o first.so a.cu\nnvcc -o second.so b.cu\n")
+        self.assertEqual(parsed["output"], "second.so")
+
+    def test_a_trailing_quote_does_not_end_up_in_the_name(self):
+        parsed = build_script.parse_build_script('nvcc -o "libx.so" a.cu\n')
+        self.assertEqual(parsed["output"], "libx.so")
+
+    def test_gencode_targets_are_read_from_the_flags_not_the_variables(self):
+        parsed = build_script.parse_build_script(
+            "out=${1:?}\narch=${2:-89}\n"
+            'nvcc -gencode "arch=compute_90,code=sm_90" -shared '
+            '-o "$out/libx.so" ./*.cu\n')
+        self.assertEqual(parsed["architectures"], [89])
+        self.assertEqual(parsed["gencode_architectures"], [90])
+
+    def test_each_gencode_flag_resolves_against_the_value_in_force(self):
+        # A loop that reassigns `arch` builds both targets; resolving every line
+        # against the final value would report 90 twice and lose 89.
+        parsed = build_script.parse_build_script(
+            "out=${1:?}\n"
+            'arch=89\nnvcc -gencode "arch=compute_${arch},code=sm_${arch}" -o "$out/a.so" x.cu\n'
+            'arch=90\nnvcc -gencode "arch=compute_${arch},code=sm_${arch}" -o "$out/a.so" x.cu\n')
+        self.assertEqual(parsed["gencode_architectures"], [89, 90])
+
+    def test_a_script_with_no_gencode_reports_no_gencode_targets(self):
+        parsed = build_script.parse_build_script("nvcc -shared -o l.so ./*.cu\n")
+        self.assertEqual(parsed["gencode_architectures"], [])
+
+    def test_the_parser_never_executes_the_script(self):
+        # A script whose first command would be destructive must still be read.
+        parsed = build_script.parse_build_script(
+            "rm -rf /\nout=${1}\nnvcc -o \"$out/libz.so\" a.cu\n")
+        self.assertEqual(parsed["output"], "libz.so")
+
+
+if __name__ == "__main__":
+    unittest.main(verbosity=2)
diff --git a/src/backends/cuda/tests/test_check_contract.py b/src/backends/cuda/tests/test_check_contract.py
new file mode 100644
index 0000000..553bd33
--- /dev/null
+++ b/src/backends/cuda/tests/test_check_contract.py
@@ -0,0 +1,636 @@
+#!/usr/bin/env python3
+"""Tests for the Tier-1 CUDA backend contract check.
+
+No GPU, no nvcc, no CUDA toolkit. Run from the repository root:
+
+    python3 -m unittest discover -s src/backends/cuda/tests -v
+    python3 src/backends/cuda/tests/test_check_contract.py
+
+The build-script case uses the real text of the ``qwen3_5/build.sh`` added by
+PR #19, because a parser validated only against invented samples proves nothing
+about the script it has to read.
+"""
+
+from __future__ import annotations
+
+import contextlib
+import io
+import json
+import os
+import shutil
+import sys
+import tempfile
+import unittest
+
+sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
+
+import check_contract  # noqa: E402  (path is set above)
+
+# The argument handling of src/backends/cuda/qwen3_5/build.sh in PR #19: `$1` is
+# the output directory and `$2` defaults through CUDA_COMPUTE_CAP to 89.
+QWEN3_5_BUILD_SCRIPT = """#!/usr/bin/env bash
+# Build libqwen3_5_cuda.so from the kernels in this directory.
+set -euo pipefail
+here=$(cd "$(dirname "$0")" && pwd)
+out=${1:?usage: build.sh <output dir> [compute capability]}
+arch=${2:-${CUDA_COMPUTE_CAP:-89}}
+mkdir -p "$out"
+"$nvcc" -O3 -std=c++17 -gencode "arch=compute_${arch},code=[sm_${arch},compute_${arch}]" \\
+    -shared -Xcompiler -fPIC -I"$here" "$here"/*.cu \\
+    "${link[@]}" -o "$out/libqwen3_5_cuda.so"
+echo "built $out/libqwen3_5_cuda.so for sm_${arch}"
+"""
+
+KERNELS_CU = """
+#include "ops.h"
+extern "C" int fake_kernel(const void* x, int n);
+"""
+
+
+def manifest(**overrides):
+    """A minimal schema-complete manifest, matching PR #19's qwen3_5 backend.
+
+    ``build.architectures`` is ``[89]`` because that is the whole truth about
+    this build script: one invocation produces machine code for one compute
+    capability, and PTX for newer parts to compile at load time.
+    """
+    value = {
+        "name": "qwen3_5",
+        "abi_version": 1,
+        "status": "validated",
+        "sources": ["ops.h", "kernels.cu"],
+        "build": {
+            "script": "build.sh",
+            "output": "libqwen3_5_cuda.so",
+            "default_arch": 89,
+            "architectures": [89],
+        },
+        "numerics": {
+            "precision": "bfloat16",
+            "accumulation": "float32",
+            "tolerance": {"max_abs": 0.039},
+        },
+        "reference": {"entrypoint": "recipe/cua_s1/check_native.py"},
+    }
+    value.update(overrides)
+    return value
+
+
+class CheckManifestTest(unittest.TestCase):
+    def setUp(self):
+        self.root = tempfile.mkdtemp(prefix="omni-contract-")
+        self.directory = os.path.join(self.root, "src", "backends", "cuda", "qwen3_5")
+        os.makedirs(self.directory)
+        with open(os.path.join(self.directory, "kernels.cu"), "w", encoding="utf-8") as handle:
+            handle.write(KERNELS_CU)
+        with open(os.path.join(self.directory, "ops.h"), "w", encoding="utf-8") as handle:
+            handle.write("#pragma once\n#define TEST_ABI_VERSION 1\n")
+        with open(os.path.join(self.directory, "build.sh"), "w", encoding="utf-8") as handle:
+            handle.write(QWEN3_5_BUILD_SCRIPT)
+        os.chmod(os.path.join(self.directory, "build.sh"), 0o755)
+        # The validated manifest declares this reference, so the fixture ships it.
+        recipe = os.path.join(self.root, "recipe", "cua_s1")
+        os.makedirs(recipe)
+        with open(os.path.join(recipe, "check_native.py"), "w", encoding="utf-8") as handle:
+            handle.write("import sys\nsys.exit(0)\n")
+
+    def tearDown(self):
+        shutil.rmtree(self.root, ignore_errors=True)
+
+    def write_manifest(self, value):
+        path = os.path.join(self.directory, "qwen3_5.backend.json")
+        with open(path, "w", encoding="utf-8") as handle:
+            json.dump(value, handle)
+
+    def check(self):
+        manifests, issues = check_contract.run(self.root)
+        return manifests, [issue for issue in issues if issue.level == "error"], \
+            [issue for issue in issues if issue.level == "warning"]
+
+    def messages(self, issues):
+        return " | ".join(issue.message for issue in issues)
+
+    def test_a_complete_manifest_passes(self):
+        self.write_manifest(manifest())
+        manifests, errors, warnings = self.check()
+        self.assertEqual(len(manifests), 1)
+        self.assertEqual(errors, [], self.messages(errors))
+        self.assertEqual(warnings, [], self.messages(warnings))
+
+    def test_kernels_without_a_manifest_are_an_error(self):
+        manifests, errors, _ = self.check()
+        self.assertEqual(manifests, [])
+        self.assertEqual(len(errors), 1)
+        self.assertIn("no qwen3_5.backend.json manifest", errors[0].message)
+
+    def test_every_required_key_is_reported_and_then_guarded(self):
+        """Removing any required key must produce a report, never an exception.
+
+        The guard and the report were two hand-written lists, and `name` was in
+        one but not the other: a manifest missing only `name` raised KeyError out
+        of check_manifest, which also aborted the checks of every other backend
+        and broke --json. Iterating over the keys rather than testing one of them
+        is what covers the whole class.
+        """
+        for key in check_contract.REQUIRED_KEYS:
+            value = manifest()
+            del value[key]
+            self.write_manifest(value)
+            try:
+                _, errors, _ = self.check()
+            except KeyError as error:
+                self.fail("removing %r raised %s instead of reporting it" % (key, error))
+            self.assertIn("missing required key %r" % key, self.messages(errors))
+
+    def test_a_manifest_missing_name_does_not_stop_other_backends(self):
+        # The failure mode the maintainer reproduced: one bad manifest must not
+        # hide the results for the rest.
+        value = manifest()
+        del value["name"]
+        self.write_manifest(value)
+        other = os.path.join(self.root, "src", "backends", "cuda", "other")
+        os.makedirs(other)
+        with open(os.path.join(other, "k.cu"), "w", encoding="utf-8") as handle:
+            handle.write("// kernel\n")
+        with open(os.path.join(other, "other.backend.json"), "w", encoding="utf-8") as handle:
+            json.dump({"name": "other", "abi_version": 1, "status": "planned",
+                       "sources": ["k.cu"],
+                       "build": {"script": "build.sh", "output": "libother.so",
+                                 "default_arch": 89, "architectures": [89]}}, handle)
+        manifests, issues = check_contract.run(self.root)
+        # A manifest missing `name` is still discovered and still listed; it
+        # is the entry key that identifies it, which is why the checker
+        # reports against the directory name rather than the missing field.
+        self.assertEqual(sorted(m["_entry"] for m in manifests), ["other", "qwen3_5"])
+        self.assertTrue(any("missing required key 'name'" in i.message for i in issues))
+
+    def test_missing_required_key_is_an_error(self):
+        value = manifest()
+        del value["sources"]
+        self.write_manifest(value)
+        _, errors, _ = self.check()
+        self.assertIn("missing required key 'sources'", self.messages(errors))
+
+    def test_declared_source_that_does_not_exist_is_an_error(self):
+        self.write_manifest(manifest(sources=["ops.h", "gdn_prefill.cu"]))
+        _, errors, _ = self.check()
+        self.assertIn("declared source 'gdn_prefill.cu' does not exist", self.messages(errors))
+
+    def test_source_escaping_the_backend_directory_is_an_error(self):
+        self.write_manifest(manifest(sources=["ops.h", "../../../etc/passwd"]))
+        _, errors, _ = self.check()
+        self.assertIn("must be relative to the backend directory", self.messages(errors))
+
+    def test_output_name_must_match_the_build_script(self):
+        value = manifest()
+        value["build"]["output"] = "libwrong.so"
+        self.write_manifest(value)
+        _, errors, _ = self.check()
+        self.assertIn("does not match the 'libqwen3_5_cuda.so' the build script writes",
+                      self.messages(errors))
+
+    def test_architecture_the_script_cannot_reach_is_an_error(self):
+        # The build script defaults one architecture and offers no way to reach
+        # another, so a manifest claiming [89, 90] is over-claiming.
+        value = manifest()
+        value["build"]["architectures"] = [89, 90]
+        self.write_manifest(value)
+        _, errors, _ = self.check()
+        self.assertIn("build.architectures claims [89, 90] but the build script only reaches [89]",
+                      self.messages(errors))
+
+    def test_two_architectures_in_one_script_are_accepted(self):
+        # The architectures have to be visible in the script. A `for arch in ...`
+        # loop alone is not read as a declaration, so the script names them.
+        script = """#!/usr/bin/env bash
+set -euo pipefail
+out=${1:?usage: build.sh <output dir>}
+arch=89
+"$nvcc" -gencode "arch=compute_${arch},code=sm_${arch}" -shared \\
+    -o "$out/libmulti_cuda_${arch}.so" ./*.cu
+arch=90
+"$nvcc" -gencode "arch=compute_${arch},code=sm_${arch}" -shared \\
+    -o "$out/libmulti_cuda_${arch}.so" ./*.cu
+"""
+        with open(os.path.join(self.directory, "build.sh"), "w", encoding="utf-8") as handle:
+            handle.write(script)
+        value = manifest()
+        value["build"]["architectures"] = [89, 90]
+        # The script's last assignment is arch=90, so that is the library a run
+        # produces; the manifest has to name the same file.
+        value["build"]["output"] = "libmulti_cuda_90.so"
+        self.write_manifest(value)
+        _, errors, _ = self.check()
+        self.assertEqual(errors, [], self.messages(errors))
+
+    def test_warns_when_a_script_declares_no_compute_capability(self):
+        with open(os.path.join(self.directory, "build.sh"), "w", encoding="utf-8") as handle:
+            handle.write('#!/usr/bin/env bash\nnvcc -shared -o libqwen3_5_cuda.so ./*.cu\n')
+        self.write_manifest(manifest())
+        _, errors, warnings = self.check()
+        self.assertEqual(errors, [], self.messages(errors))
+        self.assertIn("declares no compute capability", self.messages(warnings))
+
+    def test_min_capability_covered_by_the_architectures_passes(self):
+        # The real #19 case: kernels documented as sm_80-and-later, built sm_89.
+        value = manifest()
+        value["build"]["min_capability"] = 80
+        self.write_manifest(value)
+        _, errors, _ = self.check()
+        self.assertEqual(errors, [], self.messages(errors))
+
+    def test_an_architecture_below_the_kernel_requirement_is_an_error(self):
+        value = manifest()
+        value["build"]["min_capability"] = 90
+        self.write_manifest(value)
+        _, errors, _ = self.check()
+        self.assertIn("below build.min_capability 90", self.messages(errors))
+
+    def test_a_non_integer_min_capability_is_an_error(self):
+        value = manifest()
+        value["build"]["min_capability"] = "80"
+        self.write_manifest(value)
+        _, errors, _ = self.check()
+        self.assertIn("must be an integer compute capability", self.messages(errors))
+
+    def test_listing_a_consumer_engine_that_exists_passes(self):
+        # Kev and Cua-S1 share the Qwen3.5 backbone, so one backend can serve
+        # both without either model owning a private copy of the kernels.
+        os.makedirs(os.path.join(self.root, "src", "models", "kev"))
+        value = manifest(models=["src/models/cua_s1/", "src/models/kev/"])
+        self.write_manifest(value)
+        _, errors, warnings = self.check()
+        self.assertEqual(errors, [], self.messages(errors))
+        self.assertIn("src/models/cua_s1/", self.messages(warnings))
+
+    def test_a_consumer_path_that_is_not_a_directory_is_an_error(self):
+        value = manifest(models=["src/models/kev"])
+        self.write_manifest(value)
+        _, errors, _ = self.check()
+        self.assertIn("must be a repository-relative directory path", self.messages(errors))
+
+    def test_a_non_list_models_field_is_an_error(self):
+        value = manifest(models="src/models/kev/")
+        self.write_manifest(value)
+        _, errors, _ = self.check()
+        self.assertIn("models must be a list of strings", self.messages(errors))
+
+    def test_min_capability_is_optional(self):
+        # Omitting it must not fail: not every backend states a requirement yet.
+        self.write_manifest(manifest())
+        _, errors, _ = self.check()
+        self.assertEqual(errors, [], self.messages(errors))
+
+    def test_default_arch_outside_the_list_is_an_error(self):
+        value = manifest()
+        value["build"]["default_arch"] = 75
+        self.write_manifest(value)
+        _, errors, _ = self.check()
+        self.assertIn("is not listed in build.architectures", self.messages(errors))
+
+    def test_validated_without_a_tolerance_is_an_error(self):
+        value = manifest()
+        del value["numerics"]["tolerance"]
+        self.write_manifest(value)
+        _, errors, _ = self.check()
+        self.assertIn("status is validated but numerics.tolerance is missing",
+                      self.messages(errors))
+
+    def test_validated_without_a_reference_entrypoint_is_an_error(self):
+        value = manifest()
+        del value["reference"]
+        self.write_manifest(value)
+        _, errors, _ = self.check()
+        self.assertIn("status is validated but reference.entrypoint is missing",
+                      self.messages(errors))
+
+    def test_experimental_needs_neither_tolerance_nor_entrypoint(self):
+        self.write_manifest(manifest(status="experimental", numerics={}, reference={}))
+        _, errors, warnings = self.check()
+        self.assertEqual(errors, [], self.messages(errors))
+        self.assertEqual(warnings, [], self.messages(warnings))
+
+    def test_planned_backend_is_not_required_to_build(self):
+        # `planned` means the directory is declared but nothing is built yet, so
+        # a build script that does not exist is fine. The schema still applies.
+        self.write_manifest(manifest(status="planned", sources=["ops.h"], build={
+            "script": "build.sh", "output": "libqwen3_5_cuda.so",
+            "default_arch": 89, "architectures": [89],
+        }))
+        _, errors, _ = self.check()
+        self.assertEqual(errors, [], self.messages(errors))
+
+    def test_a_non_executable_build_script_is_an_error(self):
+        # The compile job runs `./build.sh`, so the execute bit is part of the
+        # contract, not a local detail.
+        os.chmod(os.path.join(self.directory, "build.sh"), 0o644)
+        self.write_manifest(manifest())
+        _, errors, _ = self.check()
+        self.assertIn("is not executable", self.messages(errors))
+
+    def test_a_non_string_reference_entrypoint_is_reported_not_raised(self):
+        # A list here reached os.path.isabs, whose TypeError escaped
+        # check_manifest: no report at all, not even --json, and every later
+        # backend went unchecked.
+        self.write_manifest(manifest(reference={"entrypoint": ["reference.py"]}))
+        _, errors, _ = self.check()
+        self.assertIn("reference.entrypoint must be a string, got list",
+                      self.messages(errors))
+
+    def test_an_empty_reference_entrypoint_is_reported(self):
+        self.write_manifest(manifest(reference={"entrypoint": ""}))
+        _, errors, _ = self.check()
+        self.assertIn("status is validated but reference.entrypoint is missing",
+                      self.messages(errors))
+
+    def test_an_assigned_architecture_never_passed_to_nvcc_is_an_error(self):
+        # The variable says 89; the compiler is told 90. Reading the variable
+        # alone accepted a manifest declaring sm_89 that the script cannot build.
+        script = ("#!/usr/bin/env bash\n"
+                  "out=${1:?usage}\n"
+                  "arch=${2:-89}\n"
+                  'nvcc -gencode "arch=compute_90,code=sm_90" '
+                  '-shared -o "$out/libqwen3_5_cuda.so" ./*.cu\n')
+        with open(os.path.join(self.directory, "build.sh"), "w", encoding="utf-8") as handle:
+            handle.write(script)
+        os.chmod(os.path.join(self.directory, "build.sh"), 0o755)
+        self.write_manifest(manifest())
+        _, errors, _ = self.check()
+        self.assertIn("never reaches a -gencode flag", self.messages(errors))
+
+    def test_a_misnamed_manifest_is_an_error_not_an_exemption(self):
+        # Discovery looks for <dir>/<dir>.backend.json. A typo there used to be
+        # exempted silently: exit 0, checked: [], no errors.
+        self.write_manifest(manifest())
+        os.rename(os.path.join(self.directory, "qwen3_5.backend.json"),
+                  os.path.join(self.directory, "typo.backend.json"))
+        manifests, issues = check_contract.run(self.root)
+        self.assertEqual(manifests, [])
+        messages = " | ".join(issue.message for issue in issues)
+        self.assertIn("has typo.backend.json but not qwen3_5.backend.json", messages)
+        self.assertIn("never discovered", messages)
+
+    def test_unknown_status_is_an_error(self):
+        self.write_manifest(manifest(status="done"))
+        _, errors, _ = self.check()
+        self.assertIn("status must be one of", self.messages(errors))
+
+    def test_name_must_match_the_directory(self):
+        self.write_manifest(manifest(name="laya"))
+        _, errors, _ = self.check()
+        self.assertIn("does not match directory name 'qwen3_5'", self.messages(errors))
+
+    def test_non_numeric_tolerance_is_an_error(self):
+        value = manifest()
+        value["numerics"]["tolerance"] = {"max_abs": "small"}
+        self.write_manifest(value)
+        _, errors, _ = self.check()
+        self.assertIn("numerics.tolerance.max_abs must be a number", self.messages(errors))
+
+    def test_reference_entrypoint_that_is_not_there_warns(self):
+        # The fixture ships the reference; a declared-but-missing one must warn,
+        # because Tier 2 would have nothing to run.
+        os.remove(os.path.join(self.root, "recipe", "cua_s1", "check_native.py"))
+        self.write_manifest(manifest())
+        _, errors, warnings = self.check()
+        self.assertEqual(errors, [], self.messages(errors))
+        self.assertIn("reference.entrypoint", self.messages(warnings))
+
+    def test_absolute_reference_entrypoint_is_an_error(self):
+        self.write_manifest(manifest(reference={"entrypoint": "/tmp/check.py"}))
+        _, errors, _ = self.check()
+        self.assertIn("must be repository-relative", self.messages(errors))
+
+    def test_invalid_json_is_reported_not_raised(self):
+        with open(os.path.join(self.directory, "qwen3_5.backend.json"), "w",
+                  encoding="utf-8") as handle:
+            handle.write("{not json")
+        manifests, errors, _ = self.check()
+        self.assertEqual(manifests, [])
+        self.assertIn("is not valid JSON", self.messages(errors))
+
+
+class AbiVersionTest(unittest.TestCase):
+    """`abi_version` is the library's own, taken from its header.
+
+    Read as text, so the checker needs no compiler. What has to hold is that a
+    manifest agrees with the header of the library it describes. Two libraries
+    may differ from each other, so there is deliberately no cross-backend
+    sameness rule: the old code required one, which would have forced unrelated
+    model engines onto a shared interface version.
+    """
+
+    def setUp(self):
+        self.root = tempfile.mkdtemp(prefix="omni-abi-")
+        self.directory = os.path.join(self.root, "src", "backends", "cuda")
+        os.makedirs(self.directory)
+
+    def tearDown(self):
+        shutil.rmtree(self.root, ignore_errors=True)
+
+    def add_backend(self, name, abi_version, macro=None, extra_source=None):
+        directory = os.path.join(self.directory, name)
+        os.makedirs(directory)
+        header = "#pragma once\n"
+        if macro is not None:
+            header += "#define %s_ABI_VERSION %d\n" % (name.upper(), macro)
+        with open(os.path.join(directory, "ops.h"), "w", encoding="utf-8") as handle:
+            handle.write(header)
+        sources = ["ops.h"]
+        if extra_source is not None:
+            with open(os.path.join(directory, "extra.h"), "w", encoding="utf-8") as handle:
+                handle.write("#define OTHER_ABI_VERSION %d\n" % extra_source)
+            sources.append("extra.h")
+        with open(os.path.join(directory, "build.sh"), "w", encoding="utf-8") as handle:
+            handle.write("#!/usr/bin/env bash\nout=${1:?}\narch=${2:-89}\n"
+                         'nvcc -o "$out/lib%s.so" ./*.cu\n' % name)
+        os.chmod(os.path.join(directory, "build.sh"), 0o755)
+        value = {
+            "name": name,
+            "abi_version": abi_version,
+            "status": "planned",
+            "sources": sources,
+            "build": {"script": "build.sh", "output": "lib%s.so" % name,
+                      "default_arch": 89, "architectures": [89]},
+        }
+        with open(os.path.join(directory, name + ".backend.json"), "w",
+                  encoding="utf-8") as handle:
+            json.dump(value, handle)
+
+    def check(self, root=None):
+        _, issues = check_contract.run(root or self.root)
+        errors = [i.message for i in issues if i.level == "error"]
+        warnings = [i.message for i in issues if i.level == "warning"]
+        return errors, warnings
+
+    def test_a_manifest_agreeing_with_its_header_passes(self):
+        self.add_backend("qwen3_5", 4, macro=4)
+        errors, warnings = self.check()
+        self.assertEqual(errors, [])
+        self.assertEqual(warnings, [])
+
+    def test_a_manifest_disagreeing_with_its_header_is_an_error(self):
+        # The case nothing caught before: #19's ops.h says 4, and a manifest
+        # claiming 1 was accepted because no check read the header.
+        self.add_backend("qwen3_5", 1, macro=4)
+        errors, _ = self.check()
+        self.assertEqual(len(errors), 1)
+        self.assertIn("abi_version is 1 but ops.h defines QWEN3_5_ABI_VERSION 4", errors[0])
+
+    def test_sources_without_an_abi_macro_warn(self):
+        self.add_backend("qwen3_5", 1, macro=None)
+        errors, warnings = self.check()
+        self.assertEqual(errors, [])
+        self.assertIn("cannot be checked against the library", warnings[0])
+
+    def test_two_versions_in_one_backends_sources_are_an_error(self):
+        self.add_backend("qwen3_5", 4, macro=4, extra_source=7)
+        errors, _ = self.check()
+        self.assertEqual(len(errors), 1)
+        self.assertIn("more than one ABI version", errors[0])
+
+    def test_abi_version_below_the_contract_is_an_error(self):
+        self.add_backend("qwen3_5", 0, macro=0)
+        errors, _ = self.check()
+        self.assertIn("abi_version must be an integer >= 1", " | ".join(errors))
+
+    def test_two_backends_may_declare_different_versions(self):
+        # Deliberately allowed now. Requiring them to match would force unrelated
+        # model engines onto one interface version, which the repository layout
+        # explicitly does not ask for.
+        self.add_backend("qwen3_5", 4, macro=4)
+        self.add_backend("laya", 2, macro=2)
+        errors, warnings = self.check()
+        self.assertEqual(errors, [])
+        self.assertEqual(warnings, [])
+
+    def test_a_planned_backend_is_still_checked(self):
+        self.add_backend("qwen3_5", 9, macro=4)
+        errors, _ = self.check()
+        self.assertIn("abi_version is 9", " | ".join(errors))
+
+
+class RepositoryStateTest(unittest.TestCase):
+    """The checker must pass on a repository that satisfies the contract.
+
+    The fixture mirrors the qwen3_5 backend as PR #19 defines it, including a
+    build script with the same argument handling, so this exercises discovery,
+    script parsing, schema checks and the ABI rule together.
+    """
+
+    def setUp(self):
+        self.root = tempfile.mkdtemp(prefix="omni-repo-")
+        self.directory = os.path.join(self.root, "src", "backends", "cuda", "qwen3_5")
+        os.makedirs(self.directory)
+        with open(os.path.join(self.directory, "kernels.cu"), "w", encoding="utf-8") as handle:
+            handle.write(KERNELS_CU)
+        with open(os.path.join(self.directory, "ops.h"), "w", encoding="utf-8") as handle:
+            handle.write("#pragma once\n#define TEST_ABI_VERSION 1\n")
+        with open(os.path.join(self.directory, "build.sh"), "w", encoding="utf-8") as handle:
+            handle.write(QWEN3_5_BUILD_SCRIPT)
+        os.chmod(os.path.join(self.directory, "build.sh"), 0o755)
+        recipe = os.path.join(self.root, "recipe", "cua_s1")
+        os.makedirs(recipe)
+        with open(os.path.join(recipe, "check_native.py"), "w", encoding="utf-8") as handle:
+            handle.write("import sys\nsys.exit(0)\n")
+        with open(os.path.join(self.directory, "qwen3_5.backend.json"), "w",
+                  encoding="utf-8") as handle:
+            json.dump(manifest(), handle)
+
+    def tearDown(self):
+        shutil.rmtree(self.root, ignore_errors=True)
+
+    def test_a_contract_satisfying_repository_has_no_findings(self):
+        manifests, issues = check_contract.run(self.root)
+        self.assertEqual([issue.message for issue in issues], [])
+        self.assertEqual([m["name"] for m in manifests], ["qwen3_5"])
+
+    def test_the_repository_under_test_is_discovered_by_the_script(self):
+        """Running the module as CI does must succeed on this fixture."""
+        import contextlib
+        import io
+
+        buffer = io.StringIO()
+        with contextlib.redirect_stdout(buffer):
+            exit_code = check_contract.main(["--repo-root", self.root, "--json"])
+        self.assertEqual(exit_code, 0)
+        self.assertIn('"errors": 0', buffer.getvalue())
+
+
+class FlatLayoutTest(unittest.TestCase):
+    """A backend whose files sit directly under src/backends/cuda/.
+
+    Laya's kernels and tools are `src/backends/cuda/kernels/` and
+    `src/backends/cuda/tools/` rather than a `<name>/` subdirectory, so the
+    manifest sits beside them as `laya.backend.json`. The layout is the model
+    author's call; only the manifest's contents are fixed.
+    """
+
+    def setUp(self):
+        self.root = tempfile.mkdtemp(prefix="omni-flat-")
+        self.cuda = os.path.join(self.root, "src", "backends", "cuda")
+        os.makedirs(os.path.join(self.cuda, "kernels"))
+        os.makedirs(os.path.join(self.cuda, "tools"))
+        for path in ("kernels/runtime.cu", "kernels/model_ops.cu"):
+            with open(os.path.join(self.cuda, path), "w", encoding="utf-8") as handle:
+                handle.write("// kernel\n")
+        with open(os.path.join(self.cuda, "ops.h"), "w", encoding="utf-8") as handle:
+            handle.write("#pragma once\n#define LAYA_ABI_VERSION 1\n")
+        with open(os.path.join(self.cuda, "build.sh"), "w", encoding="utf-8") as handle:
+            handle.write("#!/usr/bin/env bash\n"
+                         "out=${1:?usage}\n"
+                         "arch=${2:-90}\n"
+                         'nvcc -gencode "arch=compute_${arch},code=sm_${arch}" '
+                         '-shared -o "$out/liblaya_cuda.so" ./*.cu\n')
+        os.chmod(os.path.join(self.cuda, "build.sh"), 0o755)
+        os.makedirs(os.path.join(self.root, "recipe", "laya"))
+        with open(os.path.join(self.root, "recipe", "laya", "check.py"), "w",
+                  encoding="utf-8") as handle:
+            handle.write("import sys\nsys.exit(0)\n")
+        manifest = {
+            "name": "laya",
+            "abi_version": 1,
+            "status": "experimental",
+            "sources": ["ops.h", "kernels/runtime.cu", "kernels/model_ops.cu"],
+            "build": {
+                "script": "build.sh",
+                "output": "liblaya_cuda.so",
+                "default_arch": 90,
+                "architectures": [90],
+                "min_capability": 90,
+            },
+        }
+        with open(os.path.join(self.cuda, "laya.backend.json"), "w", encoding="utf-8") as handle:
+            json.dump(manifest, handle)
+
+    def tearDown(self):
+        shutil.rmtree(self.root, ignore_errors=True)
+
+    def test_a_flat_backend_is_discovered(self):
+        manifests, issues = check_contract.run(self.root)
+        self.assertEqual([m["name"] for m in manifests], ["laya"])
+        self.assertEqual([i.message for i in issues], [])
+
+    def test_its_sources_are_resolved_against_the_cuda_directory(self):
+        manifests, _ = check_contract.run(self.root)
+        self.assertEqual(manifests[0]["_directory"],
+                         os.path.join(self.root, "src", "backends", "cuda"))
+
+    def test_a_validated_flat_backend_finds_its_reference(self):
+        # The repo root must not be inferred by walking up from the backend
+        # directory: for the flat layout that lands one level too high.
+        path = os.path.join(self.cuda, "laya.backend.json")
+        with open(path, encoding="utf-8") as handle:
+            manifest = json.load(handle)
+        manifest["status"] = "validated"
+        manifest["numerics"] = {"tolerance": {"max_abs": 0.002}}
+        manifest["reference"] = {"entrypoint": "recipe/laya/check.py"}
+        with open(path, "w", encoding="utf-8") as handle:
+            json.dump(manifest, handle)
+        _, issues = check_contract.run(self.root)
+        self.assertEqual([i.message for i in issues], [],
+                         "an existing reference must not be reported missing")
+
+
+if __name__ == "__main__":
+    unittest.main(verbosity=2)