[MaxGhenis]

report-logger filed every UK INVESTIGATE data issue in PolicyEngine/policyengine-uk-data. policyengine.py 6.16.0 loads Microcosm's microcosm_uk_2024_25 whenever a UK caller names no dataset, and policyengine-api runs only that default. So those issues went to a repo that didn't build the run's data. report-logger now files each data issue in the repo that built the run's dataset:

Run	Data issue goes to
US	PolicyEngine/microcosm (unchanged)
UK on the certified default (every API run)	PolicyEngine/microcosm
UK on a named Enhanced FRS or FRS file (enhanced_frs_*, frs_*)	PolicyEngine/policyengine-uk-data
Any other country	PolicyEngine/policyengine-{country}-data (unchanged)

Structural issues still go to PolicyEngine/policyengine-{country}.

What changes

• report-logger.md picks the repo from the run's country and, for UK runs, a new dataset_at_run frontmatter field, in both its routing prose and its issue-filing bash. Microcosm now receives data issues from both countries, so the issue body names the dataset and its data version.
• microsim-runner.md adds UK guidance for the API path (region uk, baseline policy 1, no dataset parameter) and for the local path, where pe.uk.ensure_datasets(datasets=[dataset], years=[year]) loads a named dataset. It accepts an optional dataset for local runs and returns dataset and data_version with each result. /analyze-policy gains --dataset, which needs --mode local, and records dataset_at_run in the archive.
• calibration-diagnostics.md points a UK hypothesis at Microcosm's UK target files, or at policyengine-uk-data when the run used one of its files. It also tells the agent that the calibration API lists only US releases.
• The policyengine skill said the pre-2026 datasets "are gone", and the policyengine-data skill said the per-area files "no longer exist" and "were removed". The policyengine and policyengine-uk skills now say that policyengine.py serves the Enhanced FRS and FRS files only when a caller names one, from the gated policyengine/policyengine-uk-data-private repo. The policyengine and policyengine-data skills now say that the bundle serves no per-area file.
• test_country_routing.py checks the rule with example, property and differential tests. The differential tests run report-logger's bash and resolve_issue_repo, a Python restatement of the rule, on the same inputs.
• The workflow's log_to description and target-repo comment, reform-pipeline.md, docs/structured-output-and-chaining.md and tests/pipeline/README.md repeat the rule, and analyses/README.md documents dataset_at_run. The stale-reference guard's message for per-area paths now says the bundle serves no such file.

Evidence

Claim	Source
policyengine.py 6.16.0's bundle manifest names microcosm_uk_2024_25 as the UK default	manifest.json:106
The manifest lists the Enhanced FRS and FRS files under dataset_overlays, in policyengine/policyengine-uk-data-private, and _apply_dataset_overlays raises an error for any overlay that would shadow a certified dataset	manifest.json:261-320, provenance/manifest.py:756-788
resolve_managed_dataset_reference returns the certified default when the caller names no dataset, and the named file otherwise	provenance/manifest.py:1357
pe.uk.ensure_datasets loads named datasets, and pe.uk exports economic_impact_analysis but no calculate_budgetary_impact	uk/datasets.py:378-424, uk/__init__.py:46-61
policyengine-api builds every economy request with dataset="default"	routes/economy_routes.py:128, :290
Its economy query models reject undeclared fields. GET /uk/economy/1/over/1?region=uk&time_period=2026&dataset=enhanced_frs_2024_25 returned HTTP 400, "dataset: Extra inputs are not permitted", on 2026-11-08, and the US route answered the same way	query_parameters.py:114-117, :152
Each economy result carries a policyengine_bundle with the model version, data version and dataset	services/economy_impact_execution.py:402
/uk/metadata lists one dataset under economy_options.datasets: microcosm_uk_2024_25	live API, 2026-11-08
Microcosm keeps its UK targets in build/uk/ and its UK build stages in build/uk_runtime/	microcosm 3db6515
The calibration API lists 30 releases, all US	GET https://calibration-diagnostics.vercel.app/calibration/dashboard/api/microcosm/releases, 2026-11-08
Hugging Face's policyengine/policyengine-us-data still holds 51 state and 436 district H5 files, and the bundle manifest references no file from that repo	Hugging Face tree API and manifest.json, 2026-11-08

Invariants

test_country_routing.py checks these for any dataset string Hypothesis generates:

• routing returns None or one PolicyEngine/<repo> name;
• only INVESTIGATE and structural verdicts open an issue;
• structural routing and US INVESTIGATE routing ignore the dataset;
• UK INVESTIGATE routing lands in microcosm or policyengine-uk-data and reads only the dataset's name, which it takes as policyengine.py's dataset_logical_name does (drop any @revision, keep the last path segment), so a dataset name and its hf:// URI route alike;
• report-logger's bash and resolve_issue_repo agree for us and uk.

A strict xfail records the one case where the two disagree: the bash sends ca to policyengine-ca-data, the prose sends it to policyengine-canada-data, and neither repo exists.

Not in this PR

• microsim-runner's US API path still sends a dataset query parameter, which policyengine-api answers with HTTP 400. Its table of advertised US datasets also lists names that /us/metadata no longer serves. model-corroborator.md:105 sends the same parameter, and tests/pipeline/test_dataset_naming.py treats the old US names as advertised and the live default as a bad name.
• Neither PolicyEngine/microcosm nor PolicyEngine/policyengine-uk-data has the calibration or auto-filed label that report-logger's gh issue create --label passes, and gh stops before it creates an issue whose label the repo lacks (params.go:106-109).
• The ca route, which sends data issues to repos that don't exist.

Tests

uv run pytest --ignore=tests/test_skill_examples.py passes 256 tests and xfails the ca case. scripts/build_claude_wrapper.py copies the eight edited agents, commands and skills into the wrapper byte for byte.

Before merge

• Rebase onto main after Stop the data skill steering agents to microimpute and uk-data #108 merges and rerun the suite.
• Get an independent review that agrees at head.
• Squash-merge on green checks and confirm that the Sync Claude Wrapper run carries the change to policyengine-claude.

🤖 Generated with Claude Code