Take-up assumptions: compare PolicyEngine and UKMOD where they are comparable

Found while researching UKMOD (CeMPA/Essex) for scorecard comparison. Verified by reading the take-up parameters out of a real policyengine-uk 2.89.2 — the certified engine — not from documentation.

The finding

PolicyEngine-UK's take-up parameters, read from the engine:

UKMOD, for the same benefits, applies non-take-up probabilities sourced from DWP (2020) "Income-related benefits: estimates of take-up" and HMRC (2019) tax credits take-up, varying by claimant circumstance (UKMOD Country Report 2023–2030, CeMPA WP 8/26, Table 3.4):

So on four legacy benefits PolicyEngine assumes everyone entitled claims, where the published DWP/HMRC evidence says between 32% and 96% do, depending on who they are.

Why it matters for the scorecard

Two distinct consequences, and I want to be careful not to overstate either:

1. A caseload or expenditure comparison for HB/IS/CTC/WTC is not comparing the same quantity. PolicyEngine is producing an entitlement count; DWP's administrative figure is a receipt count. Any lane scoring those against each other is measuring non-take-up and attributing it to the engine. This is exactly the failure GAP_REASONS and the divergence registry (#59) exist to name — but nothing currently names it.

2. Calibration may be silently absorbing it. PolicyEngine reweights to DWP/OBR caseload and expenditure targets. If entitlement is simulated at 100% take-up and then weighted to match a receipt-based total, the weights are compensating for a modelling assumption. The scorecard already distinguishes this — OBR rows carry calibration_relationship = consumed_as_target — but the take-up axis isn't represented at all.

Proposed lane

A parameter-level comparison rather than an output one, which the scorecard has none of today: record PolicyEngine's take-up parameters read live from the certified engine, alongside UKMOD's published assumptions, with the originating DWP/HMRC source cited for each.

It stages no value claims (the #86/#91 rule — the rates originate with DWP and HMRC, and UKMOD's table is a transcription of them plus mid-point choices of its own). What it records is: what each model assumes, where the assumption came from, and how stale it is.

Caveats worth stating up front

• UKMOD applies take-up at household level per benefit with a randomisation rule; PolicyEngine's are national scalars. These are not the same mechanism, so the numbers are not strictly like-for-like — the comparable object is the assumed aggregate rate, not the algorithm.
• PolicyEngine does have reported-recipient anchoring in policyengine-uk-data, so real behaviour may be closer than the bare parameter suggests. That should be checked before any conclusion is drawn about output error.
• UKMOD flags its own rates as stale: DWP/HMRC have not updated working-age take-up statistics since 2021, and CeMPA says "we have, as yet, failed to identify a suitable remedy." PolicyEngine's are older still, and carry no such flag.

Blocked on nothing. Needs no compute (#128) and no new harvest.

Start by locating the scorecard's GAP_REASONS, divergence registry, and existing lane or parameter-comparison entry points. Review how calibration_relationship and source citations are represented, then record the certified PolicyEngine take-up parameters beside UKMOD's published assumptions. Done means the comparable aggregate rates, originating DWP/HMRC sources, dates, and mechanism caveats are represented without making value claims.