Reuse a function from unfinished research

Share an independent prediction while unrelated experiments stay unfinished. Then add a defect to a required helper, observe refusal and correct it legitimately.

Expected result and prerequisites

serving.predict(50.0) returns REST 0.4 and MCP {"result":0.4}. The full audit retains the unrelated ambiguity in old_notebook.py; only serving.predict is public.

Use Apizr 0.4.5 core, Python 3.11–3.14 on Linux/macOS and a generated REST/MCP runtime. Download research-serving.zip, extract outside the checkout and open research-serving/. Existing GPL-3.0-or-later files need no scientific library or dataset: features.py normalizes a number, serving.py uses it, experiments.py has a decorated generator and old_notebook.py imports that unrelated ambiguous experiment.

Authorize and build

python - <<'PY'
import json
from pathlib import Path
Path('operator.json').write_text(json.dumps({'schema':'apizr.operator-policy/v1',
 'grants':[{'adapter':'repository','operation':'analyze',
 'target':{'kind':'local','root':str(Path.cwd().resolve())},'permissions':['source.analyze']}]}))
Path('readiness.json').write_text('{"execution":{"modes":["direct"]}}')
PY
export RESEARCH_BUNDLES="$(cd .. && pwd -P)/research-bundles"
mkdir -p "$RESEARCH_BUNDLES"
apizr expose plan . --operator-policy operator.json --readiness-policy readiness.json \
  --interface rest --interface mcp --execution-mode direct --select python:serving:predict --plan > plan.json
apizr expose build rest . --operator-policy operator.json --readiness-policy readiness.json \
  --interface rest --execution-mode direct --select python:serving:predict --output-dir "$RESEARCH_BUNDLES/rest"
apizr expose build mcp . --operator-policy operator.json --readiness-policy readiness.json \
  --interface mcp --execution-mode direct --select python:serving:predict --output-dir "$RESEARCH_BUNDLES/mcp"

Contracts contain only serving.predict. All inspected source remains evidence; unrelated bundled files are not imported simply because they are present. See selection scope.

Diagnose and correct a required-helper failure

cp features.py features.original
printf '\nraise RuntimeError("unsafe required helper initialization")\n' >> features.py
if apizr expose plan . --operator-policy operator.json --readiness-policy readiness.json \
  --interface rest --execution-mode direct --select python:serving:predict > refused-plan.json 2> refusal.txt; then
  echo 'Expected required-helper refusal' >&2
  exit 1
fi
cat refusal.txt
mv features.original features.py
apizr expose plan . --operator-policy operator.json --readiness-policy readiness.json \
  --interface rest --interface mcp --execution-mode direct --select python:serving:predict --plan > corrected-plan.json

The diagnostic identifies the selected function's required features dependency. Restoring the original legitimate helper permits the same selection. Conditional opt-in cannot waive required helper defects.

Verify real calls

Install both emitted requirements in a separate runtime:

python -m venv ../research-runtime
../research-runtime/bin/python -m pip install -r ../research-bundles/rest/requirements.txt -r ../research-bundles/mcp/requirements.txt

Starting these direct services executes trusted source. Run REST in one terminal:

../research-runtime/bin/python -m uvicorn app:app --app-dir ../research-bundles/rest --host 127.0.0.1 --port 8000

From another shell in the extracted directory:

curl --fail http://127.0.0.1:8000/capabilities/serving.predict -H 'Content-Type: application/json' -d '{"value":50.0}'

Check 0.4. Call MCP with the official SDK, without an AI account:

../research-runtime/bin/python - <<'PY'
from pathlib import Path
import anyio
from mcp import Client, StdioServerParameters
async def main():
    async with Client(StdioServerParameters(
        command=str(Path('../research-runtime/bin/python').absolute()),
        args=[str(Path('../research-bundles/mcp/server.py').resolve())])) as client:
        assert [t.name for t in (await client.list_tools()).tools] == ['serving.predict']
        result = await client.call_tool('serving.predict', {'value':50.0})
        assert not result.is_error and result.structured_content == {'result':0.4}
        print(result.structured_content)
anyio.run(main)
PY

Installed qualification executes these exact blocks, the helper refusal and the restored selection. The MCP context owns and stops its server.

Optional analysis MCP exploration

Install the mcp profile and configure the analysis server for this exact root/policies. Request analysis/readiness summary, then detail for python:serving:predict; exposure planning remains static. These analysis tools are distinct from the generated serving.predict tool that executes the calculation.

Limits and cleanup

Repository-wide integrity/resource failures still block generation. Independent selection is not runtime safety approval. Stop owned REST processes; MCP clients close their servers. Remove only the extracted project, generated bundles and runtime after retaining diagnostics needed.