Skip to content

Add Belgium's Child rules bindings and Axiom input closure (PR 4a), still inactive - #1602

Merged
MaxGhenis merged 6 commits into
mainfrom
be/pr4a-child-closure
Nov 15, 2026
Merged

MaxGhenis merged 6 commits into
mainfrom
be/pr4a-child-closure

Conversation

@MaxGhenis

@MaxGhenis MaxGhenis commented Nov 13, 2026 •

Copy link
Copy Markdown
Contributor

Ready. CI passed at c90f5c9: all 12 jobs of run 43729379512. Both Subfleet rounds are done: 5fd86ad fixes round 2's one P1, and c90f5c9 takes its P2 and most of its P3 notes; the rest are filed (see Verification).

What this adds

PR 4a of Belgium's transport port: the resources that bind and close Belgian Child content. Nothing executes them yet.

  • build/be/axiom_rules_bindings.json binds the twelve rulespec-be family-benefit modules whose programs the committed input surface compiles on the engine entity Child, pinned at rulespec-be 632e1da1. No frame entity maps to Child, so each binding sits on the person entity and declares one binding-local view (Add binding-local views to the Axiom adapter, registry, compose and closure (S2b PR 1) #1482): a bool selector column, and an outside value for each variable.
  • build/be/axiom_input_closure.json closes the modules' 277 Child root inputs. The axiom:be mapping's person_role bindings encode 35 (a child's age and enrolment facts). The other 242 take defaults, each with a reason and a citation. Every default is a proposal (new_default: true) that awaits Max's METHOD-CARD-BE ruling.
  • gates.json gains a release_input_coverage row (allow_engine_optional: false, evidence_absent_blocks: true).
  • country_package.json registers both resources, and the Belgian golden and two resource lists follow.
  • test_be_axiom_input_closure.py calls closure_partition, the closure kernel's own partition code, on the committed surface, the mapping, the bindings and the closure.
  • test_be_target_references.py: Belgium still refuses to activate, and the test now names today's gaps (the null engine pins, the unit rule, the graph) in place of the missing bindings file.
  • test_nz_spec_package.py (NZ's file; NZ OK'd both hunks): resolve_json_pointer looks a cited file up by bare name and needs exactly one match. Belgium's axiom_input_closure.json shares its name with NZ's, and compose.py fixes both names, so when several files match a cited name, bare or folder-qualified, the resolver now takes the one at that name's path in NZ's package if it is among them, and otherwise still refuses. test_a_name_other_packages_share_cites_nz_own_file pins four scratch-tree cases.

The 277 roots

module roots encoded defaulted
be-bru basic_amounts 3 2 1
be-bru eligibility 19 4 15
be-bru selected_amount 8 3 5
be-bru social_supplements 5 1 4
be-dg amounts 12 0 12
be-dg eligibility 31 2 29
be-vlg amounts 7 1 6
be-vlg eligibility 23 2 21
be-wal amounts 11 2 9
be-wal eligibility 33 4 29
be lgaf_amounts 18 2 16
be regional_routing 107 12 95
total 277 35 242

The routing bridge reads the four regional schemes' roots through its imports, so 94 of its 95 defaults repeat a regional module's value, type and citation; six of those carry a reason of their own (the four domicile copies, the Brussels nationality copy and the Flemish raised-or-studies-outside-Belgium copy), since in the bridge they reach a child only through the region code. The 242 entries cover 111 distinct names; 188 entries hold a bool, 35 an integer and 19 a decimal.

How the defaults were set

Seven direction rules, which the closure records as default_direction_rules:

  1. A status that grants a supplement, an increase or an extension of entitlement past an age limit (study, training, a jobseeker period, a derogation) defaults to not held, a lower bound on spending.
  2. A condition the law requires for any benefit (residence, nationality or a title, domicile in the scheme's region) defaults to held, for the children the scheme's selector admits.
  3. An index or revaluation factor defaults to 1, which keeps amounts unindexed.
  4. A status that removes or suspends entitlement, or keeps it open (an exclusion, work that suspends an extension, abduction, disappearance), defaults to not held. For an exclusion or a suspension that overstates entitlement for the children it should reach; for abduction or disappearance it is a lower bound.
  5. A count, rank, income or period fact that a later graph node is to derive takes the value that pays the base benefit with no supplement or extension, and its reason names the node that replaces it where one is planned. Three such facts take another value, and their reasons say why: the German-speaking months since a disappearance (read only when the disappearance flag holds), the LGAF birth order rank (no binding requests the birth allowance) and the German-speaking benefit year (the query year).
  6. The routing bridge's region code defaults to a code outside the four regions, so the bridge pays nothing until a node supplies each child's code. Until then no routing copy of a regional default changes the bridge's result.
  7. A flag whose value is fixed wherever its one reader can change a result takes that value (the German-speaking no_current_article_9_entitlement, true).

One agent per scheme read each rule that uses a root and proposed its default. A second agent per scheme, on a different model (Opus 5.5 checking Fable 5.2's proposals), checked every pair against the YAML: readers, type, direction, citation and reason, then traced three children by hand. Of 171 distinct (module, input) pairs the checkers agreed on 152, asked for a reason or citation fix on 9 and flagged 10. No value changed. All five hand traces matched: under the full default set a 10-year-old gets each regional scheme's base amount with no supplement. (The traces predate the region default of 99; under it the routing bridge returns 0 for every child, and the regional amounts don't change.) The checkers also read three values that round 1 of the Subfleet review then changed: they agreed with the region code at 0 and no_current_article_9_entitlement at false, and flagged months-used at 0. The current 99, true and 12 were read by the fix checkers and by round 2 of the Subfleet review.

Defaults for Max's ruling by name

  • Students aged 18 to 24, Brussels and Flanders. brussels_family_benefits_child_in_qualifying_education_training_or_stage and flanders_family_benefits_child_is_pupil_student_intern_or_school_leaver default to false, so no Brussels or Flemish child aged 18 or over is entitled, students included. The mapping binds enrolment to the Walloon and German-speaking counterparts only. A follow-up PR proposes the same binding for these two roots; the mapping sits outside this PR's limits. Every Flemish child aged 18 or over in 2026 was born before 2019, so the Flemish selector leaves out each one that decree article 210 §1 still keeps on the old child benefit (see Birth cohorts). Through the Flemish bindings the default reaches only the rest in 2026, such as a child who first came to Belgium after 2018 with no right open under the federal rules on 31 December 2018 (decree article 228), until the 2019 cohort turns 18 in 2037; the routing bridge reads it too but pays nothing while its region code stays 99. The Brussels default reaches every Brussels child aged 18 or over.
  • wallonia_family_benefits_no_government_obstacle_applies = true. Decree article 5 §3 pays for a child from 1 September of the year they turn 18 to the end of the month they turn 21, enrolled or not, unless they are in an obstacle situation the Government sets; the enrolment routes start after 21 (§4). True is that rule wherever no obstacle is recorded (age < 21 and no_government_obstacle_applies, be-wal eligibility.yaml). False would pay only enrolled 18-to-20-year-olds, through the encoded enrolment leg, which unlike §4 has no age-21 floor: less than true, though not a lower bound child by child. Children under 18 are unaffected. Every Walloon child aged 18 to 20 in 2026 was born before 2020, so their amounts come from the LGAF (decree article 120) and the Walloon selector leaves them out (see Birth cohorts). Through the Walloon bindings the default first matters in 2038, when the 2020 cohort turns 18; the routing bridge reads it too but pays nothing while its region code stays 99.
  • german_speaking_community_family_benefits_employment_suspends_training_or_extension_entitlement = false. Every enrolled 18-to-24-year-old keeps the base amount whatever they earn, which leans spending up for working students.
  • belgium_family_benefits_region = 99. No constant fits a per-child code, so the default is a code outside 0 to 3, the out-of-range code the routing module's own tests use. The bridge then returns 0 for every child under any selector, a lower bound, and a request for its binding adds nothing to the four regional amounts. A W3 region node from NIS geography replaces it.
  • german_speaking_community_family_benefits_article_10_extension_months_used = 12, with no_current_article_9_entitlement = true. Twelve months drawn shut the Article 10 extension (be-dg eligibility.yaml line 98: 12 < 12 is false) whatever the flag says, so an 18-to-24-year-old not in training gets nothing, a lower bound. True is the flag's value wherever it can change that line: only for a child of 18 to 24 who has failed the age, disability and training tests and whose work suspends nothing. 0 months would open the extension for every such child.
  • Defaults that would zero a scheme's eligibility-gated amount if set the other way (each zeroes the gated amount; the German-speaking base-entitlement flag and index factor also zero the selected amount that its amounts binding requests as a component):
    • the four domicile roots (true), which rely on each scheme's selector; the nationality and title legs are joined by or in every scheme, so setting one the other way zeroes nothing on its own;
    • german_speaking_community_family_benefits_child_has_base_entitlement (true), which leaves the eligibility-gated amount to the eligibility gate (the selected amount, which that gate doesn't reach, pays the base to every child its selector admits);
    • four exclusion flags (false): the Brussels, Flemish and Walloon study-stay exclusions and the Flemish raised-or-studies-outside-Belgium exclusion;
    • the German-speaking pre_2025_index_factor, which scales its gated and selected amounts (current_index_factor isn't read in 2026, since the 2025-2027 freeze selects pre_2025_index_factor). The LGAF index multiplier and revaluation coefficient, the other two factors read in 2026, zero only the LGAF binding's monthly total, which no gate or other binding reads.
    • The LGAF rank is not one of them: a rank of 0 zeroes the article 40 base, but article 44's age supplement still pays from age 6.
  • Income defaults (45,000 in Brussels, 70,000 in Flanders and Wallonia) withhold every income-tested supplement at the modules' unindexed ceilings until an income node exists. The Walloon article 14 supplement for a household with a disabled member reads no income; its own flag's default withholds it.

Judgment calls

  • Period. One label, 2026-01, with explicit bounds and kind calendar_month, the name rulespec-be's own tests give the monthly period. Every requested variable runs by Month, and no rule version in the twelve modules starts after 2026-01-01.
  • Variables. Each eligibility binding requests the scheme's gated monthly amount and its 0/1 entitlement (the German-speaking one its statutory base entitlement). Each other regional binding requests one monthly amount before the eligibility gate, marked as a component: the scheme's selected total, or for Brussels the article 7 base, the selected amount and the article 9 social supplement. The LGAF binding requests its indexed monthly amount and the routing bridge its gated monthly amount. No binding requests a Year rule or a birth or adoption premium. The test pins each binding's variables in order.
  • Selectors. Six bool person columns, named in the bindings file with their meaning. No node produces them yet; PR 4b adds the producers and sets each cut-off. The Flemish and Walloon selectors leave out the older cohort (see Birth cohorts).
  • Outside values. 0.0 for a Money output and 0 for a Count output. The surface's engine (04315d94) gives a Count a decimal dtype, which takes either; an integer 0 stays valid if a Count becomes an integer output, for which the adapter refuses a float.
  • The LGAF module is bound so its 18 roots are closed, behind a selector that admits no 2026 child. No bound module imports it and the routing bridge returns none of its outputs; its one importer at the pin, guaranteed_family_benefits.yaml, describes a pre-2019 layer that the surface marks dense_unsupported.
  • Entity names hold person and household, the two the mapping maps. The registry also needs the unit entity that a benefit unit rule declares; Belgium has none yet, and entity_names_note says so.
  • Engine pins stay null. Belgium's surface records engine 04315d94; a follow-up refreshes it at the commit NZ pins.

Birth cohorts

None of the regional modules bound here reads a birth date, and every amount they return is their reform scheme's. Flanders and Wallonia keep an older cohort on the old LGAF-based amounts:

  • Flanders. Decree article 210 §1 keeps a child born before 1 January 2019 who met the old rules' conditions on 31 December 2018 on the old child benefit. Article 8's conditions still decide that child's entitlement, article 210 §2 sets the amount from the old rules, and article 13's base amount, which both requested Flemish monthly amounts include, excludes that child. Every Flemish child aged 8 or over in 2026 was born before 2019.
  • Wallonia. Decree article 120 keeps LGAF articles 40 to 50septies, 52 to 55 and 56ter to 76bis applying to a child born before the reform date, 1 January 2020, from which the bound Walloon rules apply. Article 121 lets the decree's article 5 prevail for a child who reaches 18 in the reform year. Every Walloon child aged 7 or over in 2026 was born before 2020.

So the Flemish and Walloon selector meanings now leave that cohort out: a graph that admitted it would pay it the reform base amount, which the law withholds from it. The routing bridge's selector meaning says the same must hold once the bridge routes children. rulespec-be encodes only a one-child slice of the cohort's amounts (first-rank 2025 amounts, chosen by age rather than birth date) in be/statutes/family_benefits/child_benefit_base_2025.yaml, which these bindings don't cover. Brussels and the German-speaking Community adjust amounts instead, which no selector cut-off can express: Brussels ordinance article 35 takes 10 euros off each article 7 base for a child born before 1 December 2019, Brussels article 39 keeps the LGAF payment rules where the December 2019 LGAF rate gives more, and German-speaking decree article 111 computes a transition amount from December 2018. METHOD-CARD-BE's row BE-MC0 now gives the cohort as a second reason Flemish and Walloon child benefits are not modelled at level, beside indexation.

What the test can't check yet

  • frame_missing and relation_missing, which need a producer's frame slice (frame_columns stays None).
  • The donor half of release_input_coverage.
  • The engine-commit match, which needs the engine pins.
  • The new gates.json row, which has no evidence before PR 4b.
  • Types and directions of the defaults: nothing executes these modules, so the readings are the check: the per-scheme checkers above (which read three values before round 1 changed them), the fix checkers, and the two Subfleet rounds, the second of which read all 242 defaults against their rules.

Invariants the test pins

  • Every Child root of every bound module is encoded or defaulted, exactly once: zero gap, zero unexpected, zero unmapped_entity, and partition_counts sums to each module's roots.
  • A root is encoded exactly when the mapping binds it.
  • A name keeps one default value and type in every module that reads it.
  • The routing bridge's region default is an integer outside 0 to 3, so the bridge routes no child.
  • Dropping any one of the 242 defaults leaves a gap that names it (all 242 are tried).
  • Each binding id keeps its module, its selector column, and its variables in order. Each outside value is a non-bool zero of its output's type.
  • No selector names a structural column, a person_<group>_id membership column of any unit, a binding variable or a root input.
  • The selectors block describes exactly the columns in use. It and entity_names_note hold sentences, not stubs.
  • Child is live only through each binding's view: without the views all 277 roots read unmapped_entity.

Verification

Independent review (three Opus 5.5 reviewers, one lens each; the author ran on Fable 5.2). First round at d7b617e:

  • Contracts (registry, adapter, closure kernel, gate): request changes. One blocker: entity_names lacks the unit entity the registry needs, and the file didn't say so. Fixed with entity_names_note. It ran the registry's view checks, InputClosure.check and the partition on the committed files, and confirmed all 16 outside values cast.
  • Test: request changes. One blocker: the activation test this PR broke. Two should-fix: no test pinned a binding's module and selector or the outside values, and the structural-column set was wrong. All fixed; five notes taken too. It ran 30 mutants in memory.
  • Content: approve with fixes. The "no rule at the pin" claim about the LGAF module was false, and this body's zero-scheme list was wrong. Both fixed. It recomputed all 107 routing entries from their owners and traced the three children per scheme.
  • 0ffd634 holds the fixes.

Second round at 0ffd634 (a fresh Opus 5.5 reviewer, all five areas of the first round's brief): approve, with eight notes and no blocker or should-fix. b81ac72 takes five of them:

  • the test pins each binding's variables and the type of each outside value;
  • a selector may not match any person_<group>_id name;
  • the structural-set comment says which set is the adapter's and which the registry's;
  • entity_names_note and the selector texts must be sentences;
  • the closure cites METHOD-CARD-BE by row (BE-MC0), not by line, and says the card sits outside this repository.

Left as is:

  • An assertion in the activation test can't fail first today, but it would catch a plausible change to validate_transport_activation.
  • Two notes concern NZ's files: the resolver tie-break had no in-file test (e068356 adds one, whose bytes NZ OK'd), and a holdout-separation docstring still says every spec refuses on missing bindings, which went to NZ.

b81ac72 had its own check: nine Opus 5.5 agents.

  • Mutation lens: 70 mutants; every JSON-only mutant from the brief fails a test.
  • Prose lens: only the description and eight reasons changed, with every value, citation and the entry order as before, and BE-MC0 holds the cited caveat.
  • Independent review of the test change: two notes, both confirmed by a skeptic and both taken. The table-independent zero and distinct-variable checks stay beside the pins, and the Count comment states the pinned engine's decimal dtype.
  • A completeness critic found nothing further in the code.

Independent review, round 1 (Subfleet hard tier, job 20261113-130631-be-pr4a-review-sf, at b81ac72): approve, with four P2 and seven P3 findings and nothing at P0 or P1. e068356 takes all of them:

  • P2-1: the LGAF texts now speak of the one requested variable, the monthly total, and say age picks its article 44 supplement.
  • P2-2: the region code moves from 0 to 99, so the bridge routes no child. The bridge's selector text, five routing copies that described a regional selector (c90f5c9 fixes a sixth, after round 2) and a new direction rule say so, and a test pins the code outside 0 to 3.
  • P2-3: the German-speaking extension is shut through article_10_extension_months_used = 12, with no_current_article_9_entitlement = true.
  • P2-4: NZ's file gains the tie-break test, whose bytes NZ OK'd.
  • P3: six reasons are reworded (Walloon diploma training, the German-speaking benefit year and base entitlement, file names on the German-speaking line references, Flemish shared housing as the upper choice, the Brussels single-child flag hiding article 7(b)), and this body's zero-scheme list is corrected.

Six Opus 5.5 agents then checked the fixes, one per finding group plus a completeness critic. They raised two P2 points: a generator bug that wrote one new Brussels sentence twice (the build now deduplicates its ruling keys) and this body's stale by-name lines. Their P3 points are taken too: the Brussels nationality routing copy, a closer LGAF note, a narrower reason for no_current_article_9_entitlement, the disappeared-child reason (false drops a disappeared child of 18 to 24), the shared-housing factor's value, the abduction and disappearance direction rule, a seventh direction rule, the region test and this body's other stale lines. Four mutants of the new NZ test survive; the extra cases that kill them are offered to NZ as a follow-up.

CI at e068356: one failure on both Pythons, test_spec_source_citations.py::test_no_resource_cites_a_line_it_does_not_pin[be] (10,617 passed). Naming the file behind each German-speaking line reference made each a line citation, and the guard refuses a line citation that binds to no commit. 5fd86ad cites each of the 15 at 632e1da1, the closure's rulespec pin, as the UK package does; no value or other text changes. Run locally, every citation guard in that file finds nothing in Belgium's resources, and all 15 rule-file line citations bind to 632e1da1.

CI at 5fd86ad: every job passes: lint, select-countries, wheels, installed-emitter, integration-uk, and engine-free, engine-uk and engine-us on 3.13 and 3.14.

Independent review, round 2 (Subfleet hard tier, job 20261114-195133-be-pr4a-review-r2, the whole change at e068356): request changes, on one P1: the 15 unpinned line citations CI caught, which 5fd86ad fixes and the reviewer checked. It read all 242 defaults against their rules, found every round-1 finding answered, and found that the registry, the adapter and the closure kernel accept these files once Belgium has engine pins and a unit entity. It raised one P2 and twelve P3 notes. c90f5c9 takes the P2 and most P3s:

  • P2-1, birth cohorts (see Birth cohorts): the Flemish, Walloon and bridge selector meanings, the views note, the closure description, METHOD-CARD-BE's row BE-MC0 and the two by-name bullets now say who the older cohort is.
  • P3s taken: this body's Walloon obstacle bullet, zero-scheme list, income line, routing-copy count, checker caveat, resolver wording and engine wording; the test comment's engine wording; direction rules 1, 4 and 5; the sixth routing copy; and reasons for the German-speaking residence alternative and months-used limit, the Flemish income thresholds and domicile replacement, the Brussels single-child replacement and nationality leg, and the Walloon derogations and income line, plus the German-speaking amounts note, the period note's Year rules, the judgment policy and the gate note's path. No value changes: 22 entries' reasons change, and ten texts (the three direction rules count as one).
  • Five Opus 5.5 agents then checked every new claim against the decrees in the Axiom corpus and the rules at the pin (one per region group, one on the reasons, and a completeness critic). They found no value problem. Their wording points are all taken: the old-cohort amounts in child_benefit_base_2025.yaml are a one-child slice; the Flemish cohort test is a right open under the federal rules anywhere in Belgium; German-speaking article 7's insured-person route; the requested amounts, not every module output, include the base; the bridge's selector; the Walloon false option; and this body's sixth routing copy, rule 5's three exceptions, the zero-scheme index item and the income line.
  • Filed as follow-ups: three LGAF reasons that describe an else-branch too narrowly, a guard that no graph requests the LGAF binding, the engine-citation regex, the Belgian test's weaker assertions (81 in-memory mutants, 66 caught), and the NZ tie-break test's surviving mutants. Its out-of-scope notes are filed too; the one that matters next is that Belgium's gate rows fail the transport gate registry's parameter validation, which PR 4b must fix.

Tests. Local, Python 3.14.7, one file per process, four workers:

  • At d7b617e, twelve files:

    • Ten passed: the closure test (18 passed), country bundles (22), source citations (138), country spec (144), transport views (59), composed graph (472), NZ's closure test (19 passed, 2 skipped), gate battery (30), holdout separation (51) and the CI test plan (253).
    • test_be_target_references.py failed the activation test, which 0ffd634 fixes.
    • test_spec_engine_reemission.py failed one of eight tests on "duplicate installed distribution identity: microcosm-emitter-interface". The local run had two dependency directories on its path, each holding that package, so this is a local artifact; CI's engine-free job runs the file in a clean environment.
    • Run separately, NZ's test_nz_spec_package.py failed four citation tests, which 0ffd634 fixes.
  • The mutant from the brief, at 0ffd634: deleting the Walloon amounts module's household_annual_income default fails five of the closure test's 18 tests, and the closure check names the root.

  • At 0ffd634, the six files the fixes touch: the closure test (18 passed), test_be_target_references.py (156 passed, 32 skipped), country spec (144), country bundles (22), source citations (138) and test_nz_spec_package.py (1204 passed).

  • resolve_json_pointer probe. No existing test builds two files of one name in a scratch tree, so a probe patches the checkout onto scratch trees and calls the function. Two foreign copies refuse. An NZ copy and a foreign copy resolve to NZ's, with a bare name or a folder-qualified one. Two foreign folder-qualified copies refuse, as does an NZ copy sitting only in a subfolder. One file resolves as before. In the real checkout, NZ's citation of axiom_input_closure.json reads NZ's file. 7 of 7 cases pass.

  • At b81ac72 (a clean tree at start and end): the closure test (18 passed), country spec (144), country bundles (22), source citations (138), test_be_target_references.py (156 passed, 32 skipped), holdout separation (51) and test_nz_spec_package.py (1204 passed).

CI at 0ffd634: every job passes: lint, select-countries, wheels, installed-emitter, integration-uk, and engine-free, engine-uk and engine-us on 3.13 and 3.14. Engine-free runs test_spec_engine_reemission.py and NZ's citation tests.

  • On the first build of these fixes, before the claim check's wording edits: the closure test (19 passed), source citations (138) and country spec (144, with the regenerated golden). Host load stayed above 100 after that, so CI is the test run at c90f5c9. A rebuild from the rulings at c90f5c9 changes no value, class, citation or entry order; 22 reasons change.

CI at c90f5c9: run 43729379512 passed all 12 jobs: lint, select-countries, wheels, installed-emitter, integration-uk, and engine-free, engine-uk and engine-us on 3.13 and 3.14.

🤖 Generated with Claude Code

MaxGhenis and others added 2 commits November 13, 2026 07:21
axiom_rules_bindings.json binds the twelve rulespec-be family-benefit
modules whose programs run on the engine entity Child. Each binding sits
on the person entity and declares one binding-local view, with a selector
column and an outside value for each variable.

axiom_input_closure.json closes the modules' 277 Child root inputs: the
axiom:be mapping's person_role bindings encode 35, and 242 take defaults
with reasons and citations. Every default is a proposal that awaits Max's
METHOD-CARD-BE ruling.

gates.json gains a release_input_coverage row, and country_package.json
registers both resources. test_be_axiom_input_closure.py runs the closure
kernel's partition on the committed input surface.

Nothing executes these modules yet: the engine pins stay null, no node
produces a selector column, and Belgium has no transport graph.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Bindings and closure:
- entity_names_note says the registry also needs the unit entity that a
  benefit unit rule declares, and engine.status no longer names the pins
  as the only gap.
- The LGAF selector no longer claims that no rule at the pin sends a child
  to the schedule: guaranteed_family_benefits.yaml imports it.
- views_note names the Brussels article 35 cohort beside the Flemish and
  Walloon ones.
- The undetermined-judgment reason says nothing reads the policy yet.
- Two reasons say a selector will admit, since no node produces one.

Tests:
- test_be_axiom_input_closure.py pins each binding id to its module and
  selector column, the selectors block, and zero outside values. It takes
  the structural columns from BE_SCHEMA, checks engine line citations both
  ways, and asks for sentence-length reasons.
- test_belgium_still_refuses_to_activate names today's gaps: the bindings
  file exists, so its null engine pins block activation.
- resolve_json_pointer in test_nz_spec_package.py picks NZ's own file when
  another country's package holds a resource of the same name. Belgium's
  axiom_input_closure.json made four NZ citation tests find two files.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
From the second review of PR 4a:
- test_be_axiom_input_closure.py pins each binding's requested variables
  in order and their outside values with their types: 0.0 for Money and
  0 for a Count. The pinned engine gives a Count a decimal dtype, which
  takes either, and an integer 0 stays valid if a Count becomes an
  integer output, for which _outside_values refuses a float. The checks
  that hold without the table (non-empty, distinct variables; non-bool
  zero outside values) stay beside the pins.
- A selector must not match any person_<group>_id name, since the
  registry's transport schema adds each unit's membership column to the
  adapter's structural set. The comment now says which set is which.
- entity_names_note and the selector descriptions must be sentences, not
  stubs.
- Eight reasons and the closure description cite METHOD-CARD-BE's row
  BE-MC0, not a line number, and the description says the method card
  sits outside this repository. No value or citation changes.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
MaxGhenis and others added 3 commits November 14, 2026 19:49
The Subfleet hard-tier review (job 20261113-130631-be-pr4a-review-sf)
approved b81ac72 with four P2 and seven P3 findings. This takes all of
them, plus the points six checking agents raised on the fixes.

- The routing bridge's region code defaults to 99, outside 0 to 3, so
  the bridge routes no child until a node supplies each child's code.
  The bridge selector text, five routing copies' reasons and a new
  direction rule say so, and a test pins the code outside 0 to 3.
- The German-speaking Article 10 extension is shut through
  article_10_extension_months_used = 12, with
  no_current_article_9_entitlement = true, the flag's value wherever it
  can change the rule.
- The LGAF texts speak of the one requested variable, the monthly total.
- NZ's test file gains the resolver tie-break test, whose bytes NZ OK'd.
- Eight more names' reasons and two direction rules are reworded to
  claim no more than the rules at rulespec-be 632e1da1 support.
- The golden follows the closure and bindings.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
e068356 named the file behind each German-speaking line reference,
which made each one a line citation, and test_spec_source_citations
refuses a line citation that binds to no commit. Each of the 15 reasons
now cites its rule file's lines at 632e1da1, the closure's rulespec pin,
as the UK package cites source lines. No value or other text changes.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Subfleet review 20261114-195133-be-pr4a-review-r2 asked for changes on
one P1, the unpinned line citations that 5fd86ad fixed. This commit
takes its P2 and most of its P3 notes. No default changes value.

P2-1, birth cohorts. Flemish decree article 210 section 1 keeps a child
born before 1 January 2019 who met the old rules' conditions on
31 December 2018 on the old child benefit, and article 13's base amount
excludes that child. Walloon decree article 120 keeps most of LGAF
articles 40 to 76bis for a child born before 1 January 2020. Both monthly amounts
the Flemish and Walloon bindings request include the reform base, so
their selector meanings, and the routing bridge's once it routes
children, now leave that cohort out. The views note says why no
selector expresses Brussels articles 35 and 39 or German-speaking
article 111, and the closure description gives the cohort as BE-MC0's
second reason Flemish and Walloon child benefits are not modelled at
level.

P3s: direction rules 1, 4 and 5 now cover extension statuses,
suspension and the two directions, and rule 5's three exceptions. The last
routing copy that described a regional selector now describes the
bridge. Nine shared reason sentences (among them German-speaking article
7's insured-person route), the German-speaking amounts note, the period
note, the judgment policy and the gate note's path are corrected, and
the test comment names the surface's engine.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@MaxGhenis
MaxGhenis marked this pull request as ready for review November 15, 2026 19:18
@MaxGhenis
MaxGhenis merged commit e555b4d into main Nov 15, 2026
12 checks passed
MaxGhenis added a commit that referenced this pull request Nov 15, 2026
Main since 085672f: #1615 (the engine adapters read one year), #1618
and #1624 (reference documents trimmed to the keys their kernels read),
#1096 (solve's exact softmax cap projection) and #1602 (Belgium Child
rules bindings). None changes a file this branch changes except four
pin files, where #1096 re-pinned the digests solve.py moves:

- docs/evidence/spec-engine/us-f0-coverage.json (5 hunks)
- spec_engine/inventory_coverage.py (seed_map, seed_protocol,
  source_manifest)
- tests/engine_free/shared/test_spec_engine_loader.py (golden)
- tests/engine_workflow/us/test_us_multispine_pool_tool.py (spec_sha256)

Each hunk takes main's comment and values, with this branch's #776
comment lines after main's; then the tools write the values on the
merged tree:
- tools/repin_us_spec_engine_contracts.py: seed_map 2935a411...,
  seed_protocol e13541c1..., source_manifest ee208024... (this branch's
  value; main never moved it), loader golden 627e2563..., pool-tool
  spec_sha256 6fb922d2... (= the generator's printed digest).
- tools/spec_engine_coverage.py: the same digests in
  us-f0-coverage.json; 42457/42457 fields, 41/41 inventory checks.
- tools/generate_us_bundle_from_constants.py and
  tools/build_us_release_input_coverage_manifest.py rewrote nothing.

ruff format on the three repinned .py files (quote style). Every
--check mode (repinner, legacy repinner, coverage, generator, pool
engine contracts), ci_test_plan verify, the manifest comparison and
the transfer test's three contract digests agree with this tree.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
MaxGhenis added a commit that referenced this pull request Nov 15, 2026
Main gained 39 commits since 085672f, this branch's last merge:
#1615 (US and UK build simulations stop at the year they read), #1618
and #1624 (reference documents trimmed to the keys their kernels read),
#1096 (solve's opt-in softmax parametrization projects exactly onto the
capped simplex) and #1602 (Belgium's Child rules bindings and Axiom
input closure, still inactive).

Four files conflicted, all of them pin files. Each conflicted hunk takes
main's text, keeps this branch's comment sentences after main's, and
leaves the value to the tools.

- spec_engine/inventory_coverage.py, one hunk at seed_map and
  seed_protocol: main's three #1096 comment lines byte for byte, then
  this branch's three #1444 and #1453 lines byte for byte.
- tests/engine_free/shared/test_spec_engine_loader.py, one hunk: main's
  #1096 comment line, then this branch's two #1453 lines, then the
  golden value.
- tests/engine_workflow/us/test_us_multispine_pool_tool.py, one hunk:
  main's "attested bytes." line and its #1096 line, then this branch's
  #1453 sentence on two new lines (same words), then the spec_sha256.
- docs/evidence/spec-engine/us-f0-coverage.json, five hunks, digests
  only, each to main's side.

Git merged three other files that both sides changed:
tools/build_us_fiscal_refresh_release.py,
tests/engine_free/us/test_us_fiscal_refresh_builder.py and
tests/engine_free/us/test_us_post_export_scoring.py. In each, every
definition outside the pins was changed by one side, and the merged AST
equals that side's. In the tool, main changed seven helpers (among them
_dataset_from_frame and _materialize_base_simulation_columns) and this
branch changed _main, _with_local_jurisdiction_flags and
LOCAL_JURISDICTION_FLAG_SUPPORTS. The tool's imports hold main's
us_dataset_through and this branch's local-jurisdiction import. Main
touched no us_runtime module and no binding-identity list.

Pins on the merged tree, each written by its tool:

- tools/repin_us_spec_engine_contracts.py: seed_map 5230ab2f and
  seed_protocol 4ac422f9 in inventory_coverage.py; the golden 23b97b69
  in shared/test_spec_engine_loader.py; the US spec_sha256 97f2b938 in
  engine_workflow test_us_multispine_pool_tool.py. Its covered_item_count
  in engine_contract test_spec_engine_coverage_tool.py ends at main's 41.
- tools/spec_engine_coverage.py: us-f0-coverage.json (the seed map and
  protocol, the seed-stream implementation digest and spec_sha256, each
  twice); 42395/42395 configuration fields and 41/41 inventory checks.
- Unmoved by the merge, as the coverage tool observes them again:
  full_checkpoint 47d4474d and the stacked pool's code digest 3c5e14be,
  this branch's values on lines main did not touch.

tools/generate_us_bundle_from_constants.py --check reports the same US
spec_sha256, 97f2b938. tools/repin_us_pool_engine_contracts.py
--emit-json gives this branch's pinned manifest, 6e70bc74 over 1,059
rows. The release input coverage manifest built on the merged tree
equals the committed one. The calibrate parity fixture (direct.csv,
graph.json, inputs.csv, pins.json) has main's blob ids; this branch
changes nothing under microcosm.calibrate.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
MaxGhenis added a commit that referenced this pull request Nov 15, 2026
…-2) (#1611)

Plan step PR 4a-2 of the Belgian transport port. Brussels article 25 §2(b) and Flemish article 8 §2 item 3 pay family benefits to a child in education up to 25; PR 4a defaulted both flags to false. This binds each flag, like Wallonia's and the German-speaking Community's, to a person_role predicate over fact:person.education_enrollment's six enrolled bands, in its scheme's eligibility module and in the regional routing bridge. The closure goes from 35 encoded and 242 defaulted roots to 39 and 238; a new test pins all four regions' post-18 flags to one enrolment predicate; the coverage and country-spec goldens are regenerated; the transport ADR now says encode runs Belgium's person_role bindings and defers only its 11 other group bindings.

A Subfleet hard-tier review (20261113-190022-be-pr4a2-review-sf, at 8890611) approved it with three P3 notes, two taken and one needing no change. The commit was rebuilt on #1602's later rounds (same eight names in 15 entries) and moved onto main with the same patch-id. CI on main: run 43751784209; engine-free (3.13) first failed on a wall-clock assert in a test this PR doesn't touch (2.13 s against 1 s; its exact-count assert passed; NZ removes the timer in #1629), and passed on rerun (attempt 2).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@MaxGhenis
MaxGhenis deleted the be/pr4a-child-closure branch November 15, 2026 23:03
MaxGhenis added a commit that referenced this pull request Nov 16, 2026
The BE column of the "Two postures, one executor" table in
docs/gate-battery-contract.md still counted nine entries, eight blocking
and one diagnostic. #1602 added release_input_coverage, so be/gates.json
declares ten: nine release-blocking and one diagnostic. #1626 made two of
them (calibration_per_family_fit and commune_fiscal_income_fit) not
applicable, which the entries row now says.

The review of the count fix noted that the incumbent-posture cell read as
if external oracles already replace self-parity. be/gates.json's policy
text leaves their implementation (EUROMOD-BE baselines and Federal
Planning Bureau reform scores) to microcosm#264, so the cell now says they
are planned there.

Review: Subfleet 20261116-091127-be-gate-contract-count-r1 approved
08b8548 with no in-scope finding.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant