[MaxGhenis]

Adds the NZ reporting classifiers the graph reports by: income decile (A1) and receipt group (A2), carried from the baseline into every scenario branch by stable unit id. This is package G8-p of #1183. The production graph wiring is G8-w; here only toy graphs wire complete branches.

What it adds

Shared, country-neutral code:

• microcosm/frame/reporting.py:
  • strict executable declarations (:63);
  • annual before-tax income = assigned receipt × engine amount × annualisation (:280);
  • weighted-midpoint income cells that sort by stable id, handle zero weights and publish shares. Each unit's cell is decided exactly: every float64 weight is a dyadic rational, so the cumulative sums and the midpoint-vs-boundary comparisons are done in exact integer arithmetic, with no float tolerance (d3827c0);
  • baseline joins by unit id (:370);
  • priority receipt groups with mixed counts (:391).
• build/transport/reporting_kernels.py:
  • a baseline producer (:195), typed baseline carry (:247) and a receipt producer (:290);
  • source weights resolved to units (:152);
  • artifact-schema and exact id-set checks refuse drift;
  • numeric scope platform_bitwise (:175).
• artifact_types.py:40 adds the nominal income-cells type; registry.py:384 registers the producers; entitlement.py:68 admits declarative reporting nodes.

NZ spec, build/nz/reporting_classifiers.json:

• A1 income: the calibrated weight source, tie key, S0 boundary and annualisation (:6).
• A2: priority, default and predicates (:139).
• Scenario carry (:169).
• Every executable value carries provenance from the G8 method addendum (:312).
• as_rate_bridge.json:422 (MC11) now references this resource's single superannuitant predicate.
• gates.json:235 lists the S0–S4 superannuitant structural zeros in the cell-support notes; the resource also has machine-readable rows.
• country_package.json:101 registers the resource.

Open items (fail closed, disclosed)

• Before-tax paid amounts for Jobseeker Support, Sole Parent Support and Supported Living Payment are deliberately null columns (reporting_classifiers.json:203). G8-w or the engine path must supply them before activation.
• Veteran's Pension receipt and amount are unsupported. A missing flag refuses rather than implying false. G8-w supplies them and extends the AS non-beneficiary closure.
• A1 stays provisional:
  • its gross-income basis waits on D2 (addendum A1);
  • the weighted midpoint is a disclosed provisional detail (:190);
  • rental and private-pension donor concepts lack explicit tax wording, so they are flagged for activation review.

Invariants (Hypothesis and example tests)

• Income cells are invariant to row permutation, and to any exact common scaling of the weights, for example by a power of two. Scaling by a non-dyadic factor such as 0.1 changes the stored floats themselves (0.3 × 0.1 = 0.30000000000000004), so cells may change; the method doc says so.
• Ties split by stable unit id.
• Zero, single and equal-weight units are handled.
• Clones stay distinct reporting units, so equal-income, equal-weight clones can land in different cells; ties split by stable id.
• Baseline labels survive variant changes to income, receipt and weight (test_transport_reporting_kernels.py:267).
• The native graph covers all nine scenarios and all 43 cells (:485).

Tests

Hub re-run at this head, one file per run:

File	Result
test_reporting.py (new, frame)	90 passed
test_transport_reporting_kernels.py (new)	31 passed
test_nz_spec_package.py	88 passed
test_country_spec.py	135 passed
test_spec_only_country_packages.py	8 passed
test_transport_graph.py	15 passed
test_transport_entitlement_graph.py	46 passed

ruff check, tools/ci_test_plan.py verify and git diff --check pass.

Mutants (each failing a named test, with a passing control):

• unweighted ranks → test_exact_weighted_midpoint_boundaries_and_uneven_weights;
• row-order ties → test_boundary_ties_split_by_stable_unit_id;
• beneficiary-first priority;
• net instead of before-tax payment;
• including housing payments;
• recomputing in the variant instead of carrying the baseline.

Mutant logs are in the lane report.

The NZ golden country spec is regenerated (bced1c2). Only the as_rate_bridge, gates and country_package hashes move, the new reporting_classifiers resource is added, and the fingerprint changes; each hash equals its committed file.

How it was built

• A GPT-6.1 Sol lane wrote it (a6a1deb). The 11:42Z reboot interrupted it; it resumed and finished its own tests and mutants.
• Its three required regressions (transport graph, NZ spec, entitlement graph) stalled on file reads in the lane, so the hub ran them, with the results above.
• Provenance now points into the repository: docs/nz-reporting-classifiers-method.md holds the A1/A2 method basis and the SSR 2018 reg 17/18 reading, and each provenance entry in reporting_classifiers.json anchors to it.

Review

• r1, 20261012-201438 (GPT-6.1 Sol) at bced1c2: REQUEST_CHANGES.
  • P1: the weighted-midpoint deciles broke the exact-boundary and weight-scaling invariants. Cumulative float rounding exceeded a fixed eps × n_bins × 4 snap.
  • Its counterexample: 625 units, left = tile([78, 46, 37], 104), weights = [left, 80, left]. Unit 312's exact midpoint is ½, so cell 6, but scaling every weight by 0.1 moved it to cell 5. It also found 14 boundary and 2 scaling failures in larger symmetric populations.
  • Non-blocking: the false "clones share cells" claim, and provenance that cited a source outside the checkout.
  • Its "Needs Max" (publishing reporting-method claims) was judged no: the NZ package is inactive and publishes no number.
• Fix, d3827c0 (GPT-6.1 Sol lane):
  • exact dyadic-integer midpoint ranks;
  • regressions for the reviewer's 625-unit case (exact and under 0.1 scaling) and for larger symmetric populations;
  • property generators to large n, a power-of-two scaling property, and a fractions.Fraction oracle differential;
  • a mutant restoring the old float cumsum fails 10 tests;
  • 500,000 units take 0.70 s for the decile step alone (lane's measurement). The r2 reviewer measured the full classifier at 1.65 s / 4.21 s / 11.77 s for 100k / 300k / 500k units, with sorting O(N log N) and the rest linear;
  • the clone wording and the in-repo method doc.
• The hub then recorded the lane's merge of main (c62a2c2) as merge commit d3827c0, merged current main (Settle and apply the NZ dependent-child benefit-unit rule (G8-u) #1226) and regenerated the golden (41d20a3). Relative to main, only the four G8-p resource hashes and the fingerprint differ.
• Re-run at 41d20a3: test_reporting 111, transport_reporting 31, nz_spec 89, country_spec 135, spec_only 8, entitlement_graph 67. ruff check, the CI plan and diff-check pass.
• r2, 20261013-000212 (GPT-6.1 Sol) at 41d20a3: APPROVE_WITH_NITS.
  • The ten-bin computation is exact for finite non-negative float64 weights, subnormals and mixed extremes included: 254 independent probes, including the 625-unit case at scales 1 and 0.1.
  • The Fraction oracle is independent. The old-accumulator mutant gives 10 failures. 159 new test instances each go red under mutation.
  • The resource hashes, fingerprint, 105 provenance entries and 117 citation anchors all verify.
  • Its entitlement run was interrupted at the disk floor after 46 passed. The hub's run at 41d20a3 had 67 passed.
• Its nits are applied in 70f1706:
  • the A2 method note now cites the pinned SSR 2018 PCO XML (axiom-corpus release nz-rulespec-2026-07-25, path and sha256 5f1250b9…);
  • the MC11 predicate citation in as_rate_bridge.json points to the in-repo note instead of a local path;
  • the tie-key evidence names build_benefit_units instead of stale line numbers;
  • the golden is regenerated;
  • the scaling invariant above is scoped and the benchmark explained.
• Tests at 70f1706: nz_spec 89, country_spec 135, reporting 111; diff-check ok.
• After r2:
  • 8eb7d88 adds reporting_classifiers.json to the generation-zero expected resource set (CI's engine-free job failed without it).
  • e4753f7 merges main f8502fe (Pin the NZ Chronicle feed and enforce full artifact identity (G8-f) #1249, which added chronicle_feed.json) and regenerates the golden. Every resource hash equals one merge side's; only country_package.json (edited on both sides) and the fingerprint are new.
• r3 delta, 20261013-114729 (GPT-6.1 Sol) at e4753f7: APPROVE_WITH_NITS.
  • Executable declarations and classifier/transport bytes match r2. country_package.json is the exact union of both merge parents. The golden recomputes byte-equal (4,645 bytes), with all 24 resource hashes current.
  • 388 tests pass across the five delta files. Both delta mutants fail (dropping the resource from the expected set; altering a golden hash), with controls passing before and after.
  • ruff check, the CI plan and diff-check pass.
  • Its one nit (P3): the pinned SSR XML isn't in this checkout, so it could not reproduce the hash.
• The hub closed that nit from axiom-corpus origin/main (9a17ccac):
  • .axiom/corpus-locks/nz/regulation/2026-06-16-rulespec-nz-pco.json records the cited path with sha256 5f1250b9…85d8f and size 1,272,548. The local corpus file hashes to the same value.
  • Release manifest manifests/releases/nz-rulespec-2026-07-25.json includes that regulation scope.
  • The hub read regulations 17 and 18 in that XML. Both match the method note:
    • regulation 17(1) defines beneficiary and non-beneficiary as the note says;
    • regulation 18(1)(b) and (2) apply the 25-cents-per-dollar reduction only to a non-beneficiary.
• An earlier r3 attempt on an Opus lane hit its usage limit before delivering; its notes were passed to this review as unverified claims.
• The earlier lane and restart history above is the hub's and lanes' report, not re-verified by the reviewers.
• All three reviews' "Needs Max" flags (publishing reporting-method and regs 17/18 prose) were judged no by the hub:
  • the NZ package is inactive and publishes no number;
  • the note restates the regulation text, which the hub checked against the pinned source;
  • the note itself labels the three-way split a reporting convention, not law.

🤖 Generated with Claude Code

[MaxGhenis]

Merged by the NZ hub (session local_e407a527) at e4753f7 with --squash. Gates: gh pr checks exit 0; MERGEABLE; not draft; no CHANGES_REQUESTED review. Reviewed head: e4753f7. Reviews (GPT-6.1 Sol): r1 REQUEST_CHANGES at bced1c2 (fixed), r2 APPROVE_WITH_NITS at 41d20a3, r3 delta APPROVE_WITH_NITS at e4753f7 (jobs/review-microcosm-1234/REPORT-114729.md). r3 P3 nit closed: SSR XML sha256 verified against axiom-corpus lock file at origin/main 9a17ccac. Needs Max judged no: inactive NZ package, no published number; regs 17/18 reading checked against the pinned text.