[MaxGhenis]

Part of #593

FRS ages are top-coded at 80. National calibration currently gives the entire 80+ sample an 80–84 target and also asks for two 85–90 columns with no records. This change emits ons/{female,male}_80_plus and ons/{region}_age_80_plus, with open-ended compute masks and complete ONS projection sums, including 90+ and the oldest grouped ages.

The data still has no records over 80, and this change only stops calibration chasing targets the data can't represent. It does not assign older ages or establish the recalibrated fiscal or distributional effect.

2025 target values (people)

Sex	Old 80–84	Old 85–90	New 80+
female	1,064,043	766,765	2,164,978
male	853,423	524,918	1,543,695

Region	Old 80–89 target	New 80+ target
east	379,000	383,909.917
east_midlands	273,000	279,796.208
london	304,000	292,199.323
north_east	153,000	149,741.415
north_west	399,000	394,643.102
northern_ireland	93,000	95,719.000
scotland	297,000	290,656.000
south_east	559,000	569,679.705
south_west	388,000	399,727.616
wales	190,000	191,184.000
west_midlands	337,000	339,480.659
yorkshire_and_the_humber	302,000	300,471.035

National targets preserve the existing 2024-based ONS UK principal projection. They now include every age group from 80 through 110 and over, including the grouped 105–109 row. The regional top targets use directly reproducible 2022-based ONS sums: the aligned five-year-migration English subnational projection and principal country projections. This also refreshes the projection vintage and precision; lower regional bands remain unchanged. The old regional values cannot safely be interpreted as omitting 90+: e.g. East 2022 was 350,000, close to published 80+ 350,364 rather than 80–89 284,842. Their original extraction is unidentified. See tools/ons_age_targets/README.md.

Signal on the released 1.58.1 file

Measured at calibration year 2025 on released uk-data 1.58.1 enhanced_frs_2024_25.h5, SHA-256 056e80fd8e9151d78f432a25cde863340baee81304c6216a37fb575ce8563784, with locked policyengine-uk 2.122.2 / core 3.32.13. One Microsimulation was created; only demographic band compute functions and household weights were evaluated, through the same _compute_column dispatch as create_target_matrix. Person records are the sum of each household count column; nonzero households count contributing matrix rows; weighted totals are column dot household_weight for each stated year. Counts and weighted totals only were output. No dataset was built or recalibrated.

Maximum age remains 80; 7,718 records are 80 and 0 are over 80. Both sex columns and the twelve regional columns separately partition those age-80 records. These are pre-recalibration column signals, not predicted recalibrated populations or fiscal effects.

New target (ons/ prefix)	Person records	Nonzero households	People at 2024 weights	People at 2025 weights
female_80_plus	4,054	4,050	2,557,341.635	2,575,754.517
male_80_plus	3,664	3,660	1,036,411.275	1,043,873.442
east_age_80_plus	614	514	369,221.371	371,879.767
east_midlands_age_80_plus	427	353	272,817.274	274,781.560
london_age_80_plus	435	385	301,125.011	303,293.110
north_east_age_80_plus	293	253	152,583.813	153,682.416
north_west_age_80_plus	609	525	390,397.299	393,208.165
northern_ireland_age_80_plus	754	628	91,494.240	92,153.000
scotland_age_80_plus	728	616	295,627.391	297,755.915
south_east_age_80_plus	1,142	913	534,126.231	537,971.943
south_west_age_80_plus	831	712	375,899.020	378,605.497
wales_age_80_plus	673	540	186,601.655	187,945.187
west_midlands_age_80_plus	460	376	333,183.400	335,582.325
yorkshire_and_the_humber_age_80_plus	752	604	290,676.204	292,769.074

Local calibration and integration

Local age CSVs contain single years 80–89 and 90+, but local_age.py emits only half-open age/0_10 … age/70_80. Age-80 people contribute zero to every local age-specific column. Their household weights still enter other local targets and the shared national demographic loss (dense local calibration uses summed local weights for that loss; the OA path stacks national targets). The new national 80+ constraint therefore applies during local calibration. Local age targets are unchanged; there is still no local 80+ age-specific constraint.

The old literal names were checked across code, tests, docs and gap lists. Saved notebook outputs remain historical. #595 is a separate unmerged branch: when combining it with this fix, remove its two ons/*_85_90 known-gap entries and change its zero-signal gap-count assertion from 22 to 20 (release count 19 to 17).

Validation

Validation at 29ca75420b885ba7cd0ee3819817a974841a5a99, on top of reviewed head 61f8f7ef:

• UV_LINK_MODE=clone uv sync --frozen --all-extras with Python 3.13 and threads capped at four; production lock unchanged. pytest-xdist installed ephemerally for serial -n 0 runs.
• Exact reviewed-head reproduction using the public-fixture plugin: PYTHONPATH="$PWD/.job:$PWD" pytest -n 0 -p offline_ons policyengine_uk_data/tests/test_national_age_bands.py at 61f8f7ef: 2 failed, 3 passed (open-ended-band subtraction and 32-vs-34 emission count). All original file contents were used and the patch restored afterward.
• Same offline national suite at this head: 5 passed. The reader includes 80_plus, validates the terminal (80, None) band and bounded five-year bands, and compares the complete 2025 population exactly with the fixture (69,502,341).
• All five targeted files, foreground serial pytest -n 0 -p offline_ons: 46 passed — test_national_age_bands.py 5; test_ons_demographics_targets.py 16; test_age_demographics.py 13; test_matrix_builder.py 10; test_calibrate_validation_targets.py 2. The requested matching-test search found the first two files; the other three preserve round 1 coverage.
• uvx ruff format --check .: 266 files already formatted; git diff --check passes.
• Fixture CRLF -> LF normalization preserves every one of its 214 rows and all values. Regional-weight comments now describe the shared household weights and summed local weights accurately.

The current continuation used only the committed public fixture and synthetic test data: no source/dataset downloads, dataset build or production calibration. Earlier regional extraction replay, released-file signals and main-versus-head checks remain documented in the round 1 hand-off and independent review; they were not repeated here. Independent review and merging remain with the hub and Max's d833 batch go. The rebuild-queue row was already installed by the hub; no build GO was requested or used.

Reproducer: policyengine_uk_data/tests/test_national_age_bands.py::test_national_gender_age_targets_are_emitted_once_per_band

axiom: n/a: calibration demographic targets only

🤖 Generated with Claude Code

[MaxGhenis]

Code hand-off at 61f8f7ef00297a1344345b87e0f97e9cf5efe398 (Part of #593).

• 41 targeted tests passed, serial foreground -n 0; all 14 new cases fail on
  main. uvx ruff format --check . passed (266 files). Public regional source
  extract reproduced byte for byte in the locked venv.
• One released 1.58.1 demographic-column measurement, aggregates only: all 14
  new 80+ columns have signal and each sex/region partition covers 7,718 records.
  New 2025 targets are female 2,164,978 and male 1,543,695. The PR body contains
  every regional old/new value and each column's counts and weighted totals.
• The data still has no records over 80. No dataset was built or recalibrated;
  local age targets remain unchanged. Regional values also refresh projection
  vintage/precision; legacy top values cannot be assumed to exclude 90+.
• Hub: queue row installed (rebuild-queue row 18); review independently, and remove the two
  obsolete ons/*_85_90 gaps when integrating Fail the calibration build on targets it drops or cannot fit #595 (coverage count 22 to 20).
  Recalibrated effect and merge wait for Max's d833 batch go. No self review or
  merge performed.

The build job's push failed on GitHub TLS timeouts. The UK hub published the branch from the job's bundle (same commit) and opened this PR. The hub changed only the Reproducer: line, to a test node for the Fix evidence check. The job's hand-off, impact report and bundle are kept in the hub's job folder (jobs/fix-593-age-topcode/job-copy/).

[MaxGhenis]

Addressed the independent review of 61f8f7ef in 29ca75420b885ba7cd0ee3819817a974841a5a99 (round 3; saved round 2 patch applied).

1. P2 national-band regression: accept ..._80_plus names, require the terminal (80, None) band, validate bounded inclusive five-year bands without subtracting None, and compare the complete emitted population with the fixture exactly: 69,502,341, including both 80+ targets. At exact 61f8f7ef contents, the fixture-backed national suite reproduces 2 failed, 3 passed (None subtraction and 32 instead of 34 targets); with this patch, 5 passed.
2. Inherited regional-weight comment: corrected the source comment and test-module explanation. Regional ten-year controls constrain the shared household weights, including summed local weights in dense local calibration; national five-year sex bands provide finer age resolution.
3. Fixture hygiene: normalized CRLF to LF. Verified byte equality after newline normalization and exact equality of all 214 rows; population remains 69,502,341.

Commands (worktree locked venv on PATH; numerical threads <=4):

UV_LINK_MODE=clone uv sync --frozen --all-extras
# Copied review offline_ons.py into .job/ so its fixture path resolves here.
PYTHONPATH="$PWD/.job:$PWD" pytest -n 0 -p offline_ons policyengine_uk_data/tests/test_national_age_bands.py
PYTHONPATH="$PWD/.job:$PWD" pytest -n 0 -p offline_ons \
  policyengine_uk_data/tests/test_national_age_bands.py \
  policyengine_uk_data/tests/test_ons_demographics_targets.py \
  policyengine_uk_data/tests/test_age_demographics.py \
  policyengine_uk_data/tests/test_matrix_builder.py \
  policyengine_uk_data/tests/test_calibrate_validation_targets.py
uvx ruff format --check .
git diff --check

Five-file result: 46 passed (national 5; ONS targets 16; local-age 13; matrix builder 10; local-validation 2). Ruff: 266 files already formatted. Original-head reproduction and current-head logs/XML, progress and hand-off are under .job/.

Code hand-off: 29ca75420b885ba7cd0ee3819817a974841a5a99 pushed to fix/593-age-topcode, preserving the existing author/history; PR Validation and test-node Reproducer updated. Impact is regression coverage, accurate comments and fixture newline hygiene; this continuation changes no target values. No source/dataset downloads, dataset build, production calibration, self-review or merge. Hub independent review and Max's d833 batch go remain outstanding.

#595 integration: whichever of #595 (eb247c64) and #604 lands second must drop ons/female_85_90 and ons/male_85_90 from targets/known_gaps.yaml; update tests/test_target_coverage.py zero-signal count 22 -> 20 and released count 19 -> 17. These files are absent from #604's base/head and were not edited here.