diff --git a/docs/design/alignment_machinery.md b/docs/design/alignment_machinery.md
new file mode 100644
index 00000000..078d1ce7
--- /dev/null
+++ b/docs/design/alignment_machinery.md
@@ -0,0 +1,857 @@
+# Alignment machinery
+
+**Status:** draft production-layer design for adversarial review. This document
+specifies a future implementation; it changes no code, gate, floor, threshold, or
+run artifact.
+
+The alignment layer produces a labeled, Trustees-aligned `after` projection while
+preserving the unaligned projection as the scientific and certified object. It
+reconciles discrete outcomes by selecting whole panel units, reconciles continuous
+covered wages by a disclosed scalar, and publishes the intervention needed for
+every margin and year. It never independently reweights linked people.
+
+## 1. Resolution boundary and inherited law
+
+### 1.1 The one deferred object this document resolves
+
+The M6 design names two report-only surfaces in one sentence:
+
+> “**Report-only** — the shock-window diagnostics, the `not_certified` surface
+> (mortality drift first, §4.10), the re-drawn-`T*`-seed comparison (decision 5b),
+> entrants (family B), and the alignment displacement — all published, none gated.”
+
+That is the binding phase-5 rule in
+`docs/design/m6_projection_engine.md:1838-1840`. The same document later records
+the current alignment state verbatim: the “runner records the alignment layer as
+not collected (`build_alignment_displacement(None, None)`,
+`m6_runner.py:1100`)” (`docs/design/m6_projection_engine.md:2100-2103`; the
+current call site has moved to `src/populace_dynamics/harness/m6_runner.py:1120-1128`).
+
+This document resolves only the missing producer of the aligned `after` frame and
+the target, reconciliation, and disclosure contract around that producer. It does
+**not** resolve the other null-with-reason surface. The re-drawn-`T*` comparison
+continues to report
+`reason = "successor_forward_seed_machinery_out_of_scope"`, and its pre-named
+margin continues to report
+`unresolved_reason = "pre_named_margin_absent_from_ratified_spec"`
+(`src/populace_dynamics/harness/m6_reporting.py:37-38,418-459`). Nothing here
+supplies a re-drawn state, chooses that margin, or creates the M6 decision-5
+successor gate.
+
+### 1.2 Mode separation is law
+
+M6 decision 9 is not reopened:
+
+> “**Alignment layer report-only; gate the un-aligned drift — ACCEPTED,** plus the
+> per-run **maximum alignment displacement** published as a reported (not gated)
+> magnitude (§4.8).”
+
+The quote is `docs/design/m6_projection_engine.md:2979-2981`. Family B likewise
+places external alignment, interventions, and maximum displacement outside the
+certifiable family-A surface (`docs/design/m6_projection_engine.md:2665-2680`).
+Therefore:
+
+1. A certification run always invokes the engine with alignment disabled. Supplying
+   an alignment specification to a gate-scoring path is an error.
+2. An aligned production run is a separate replay with a separate artifact. It
+   references, and never overwrites, its immutable unaligned companion.
+3. An aligned artifact always says `gate_scored = false`,
+   `alignment_applied = true`, and `certification_transfer = false`. Matching an
+   external target does not certify the transition law that missed it.
+4. A report may display aligned and unaligned series together, but only the
+   unaligned series may carry a gate verdict. Alignment cannot change an existing
+   verdict or manufacture a new one.
+
+“Unaligned forever” applies to the certification path, not to storage. A gate
+artifact may link to a separately produced aligned companion as a family-B
+disclosure. The gate runner must not construct that companion inside its scored
+projection call.
+
+### 1.3 Existing accounting remains the canonical surface
+
+`build_alignment_displacement` already implements the report-only accounting
+primitive (`src/populace_dynamics/harness/m6_reporting.py:321-415`). It is fully
+value-column-parameterized. It:
+
+- identifies `before` as unaligned and scored and `after` as aligned and
+  report-only;
+- requires unique, exactly equal keys, including `year`;
+- rejects a one-sided or key-mismatched comparison;
+- computes absolute displacement by named numeric value column;
+- publishes per-year field maxima, the run maximum, and the number of intervened
+  rows; and
+- returns an explicit `not_computed` record, never a fabricated zero, when neither
+  frame exists.
+
+This design does not replace, fork, or weaken that function. The production layer
+must construct lawful exact-key comparison views and call it. The current gate
+runner's `build_alignment_displacement(None, None)` remains valid for a run with no
+aligned companion. Its inherited result is exactly `status = "not_computed"` and
+`reason = "alignment_layer_output_not_collected"`. It is invalid for an artifact
+labeled as aligned: every applied margin in an aligned artifact must carry a
+`status = "computed"` displacement record.
+
+The function reports one frame-global `n_intervened_rows`, not a per-year weighted
+sum. It does not compute margin identifiers, target residuals, signed changes, or
+summed representation-weight displacement. Section 5 specifies the caller-owned
+group-bys around it. Claiming that the existing function itself computes gross
+displaced weight would be incorrect.
+
+## 2. External evidence and the mechanism decision
+
+The public record supports alignment, but it does not support treating an aligned
+aggregate as model validation.
+
+| model/source | documented method or margin | implication for this design |
+|---|---|---|
+| DYNASIM4 | The 2024 fact sheet names fertility, mortality, disability, immigration, and labor-force participation as aligned to SSA Office of the Chief Actuary targets ([Urban 2024, p. 2](https://www.urban.org/sites/default/files/2024-09/Urban%E2%80%99s%20Dynamic%20Simulation%20of%20Income%20Model%204.pdf)). A DYNASIM4 application says the employment selection criterion moves until sex-by-age employment meets Trustees targets and separately says wages follow Trustees real-wage growth and worker earnings targets ([Johnson et al. 2017, appendix p. 47](https://crr.bc.edu/wp-content/uploads/2017/11/wp_2017-17-1.pdf)). | DYNASIM4 supplies a selection-threshold precedent for discrete employment and a value adjustment precedent for wages. Its public demographic summary does not disclose the demographic reconciliation algorithm, so this design does not attribute one to DYNASIM4. |
+| DYNASIM3 lineage | The predecessor multiplies fertility probabilities within five maternal-age groups by target/predicted ratios and applies analogous linear mortality adjustments in 12 age-sex groups ([Favreault and Smith 2004, printed pp. 7-8](https://www.urban.org/sites/default/files/publication/71226/410961-A-Primer-on-the-Dynamic-Simulation-of-Income-Model-DYNASIM-.PDF)). | Probability ratios can align expectations, but a realized weighted panel can still miss a count and can hit probability caps. This is evidence for the rejected discrete-event alternative, not evidence that DYNASIM4 still uses it. |
+| Logit scaling | Stephensen defines a mean correction that multiplies odds by common factors, hits expected state totals, minimizes relative entropy, retains structural zeros, and keeps probabilities in range ([Stephensen 2016, §§2-3.2](https://www.microsimulation.pub/articles/00144)). | This is the strongest probability-scaling alternative: it defeats the clipping objection to linear ratios and supplies a principled probability-space distortion criterion. It must be compared with realized selection on its own merits. |
+| MINT7 | MINT assigns multiple potential donors, projects, and swaps donors when mortality or DI prevalence misses an age-sex-year target. Unequal weights can cause overshoot or undershoot and require repeated swaps ([Smith and Favreault 2013, printed pp. 7-8 and footnote 24](https://www.urban.org/sites/default/files/publication/22116/413131%20-%20A-Primer-on-Modeling-Income-in-the-Near-Term-Version-MINT-.pdf)). The same source says employment assumptions inform MINT but are not directly calibrated (printed p. 2, footnote 7). | Donor swapping is a whole-record selection precedent. Its unequal-weight warning requires an explicit resolution rule; it does not justify pretending an arbitrary real-valued target is exactly attainable. |
+| POLISIM | SSA's public description says each major submodule is aligned to historical data. It reports fixed benefit take-up totals by age, sex, and year, with an equation choosing the recipients; it also warns that “the alignment step may hide the erroneous ‘raw’ projection” and calls for retaining intermediate output before and after alignment ([McKay 2003, pp. 2, 5-6](https://web.archive.org/web/20080807142659/https://guard.canberra.edu.au/natsem/conference2003/papers/pdf/mckay_steven-1.pdf)). MINT6 also uses POLISIM target records and statistical donor matching for later cohorts ([Smith et al. 2010, chapter VII, printed pp. VII-2 and VII-4](https://www.urban.org/sites/default/files/publication/24986/412479-Modeling-Income-in-the-Near-Term-Version-.PDF)). | POLISIM supports selection of people to satisfy fixed totals and preservation of raw output. The accessible public sources do not specify its exact general event-selection algorithm; this design does not invent one. |
+
+The older comparative review makes the tradeoff explicit: DYNASIM2 scaled
+probabilities or continuous predictions; CORSIM chose among linear, nonlinear,
+and sorting algorithms. It also says large alignment factors may indicate a
+specification, estimation, or data-quality problem, while some large wage-growth
+adjustments simply impose a scenario assumption ([Toder et al. 2000, printed pp.
+52-53](https://www.urban.org/sites/default/files/publication/62701/410242-Long-Term-Model-Development-for-Social-Security-Policy-Analysis.PDF)).
+MINT6 provides a concrete example: substantial mortality and fertility alignment
+led its authors to recommend re-estimating the underlying equations rather than
+treating the adjusted match as success ([Smith et al. 2010, printed pp. I-5-I-6](https://www.urban.org/sites/default/files/publication/24986/412479-Modeling-Income-in-the-Near-Term-Version-.PDF)).
+
+Recent methodological work reaches the same separation. Li and O'Donoghue say
+alignment should not repair misspecification and evaluate both target deviation
+and before/after distribution distortion ([2014, §§2.3 and 5.3-5.8](https://www.jasss.org/17/1/15.html#2.3)).
+Cumpston records a DYNACAN mortality-coefficient error that alignment masked
+([2010, §2](https://www.microsimulation.pub/articles/00037#s2)). These sources
+support the campaign-honest rule: the aligned series is a disclosed production
+scenario; the unaligned companion remains the scientific diagnostic.
+
+### 2.1 Adjudication: selection for events, ratios for values
+
+For binary or categorical events in a discrete weighted panel, **deterministic
+stratified selection wins**. The reconciler changes whole event outcomes while
+leaving each person's representation weight fixed. It publishes who changed, how
+much weight those people represent, and the closest registered threshold-prefix
+target. Whole-unit selection preserves key and atomic-unit integrity; it does not
+by itself repair spouse, parent-child, or household state after a roster change.
+
+Annual **linear** probability-ratio adjustment loses for discrete production
+margins. It matches an expected rate before the draw, not necessarily the realized
+weighted count; clipping at one breaks the claimed ratio; and the scaling rule
+alone does not expose a unique set of displaced people. Independently changing
+person weights also loses because it can break the linked network and contaminate
+every downstream weighted outcome.
+
+Logit scaling is the stronger probability-adjustment alternative. For a binary
+event it sets
+`q_i(c) = c p_i / (1 + (c - 1) p_i)` for a common positive odds factor `c` and
+solves `sum_i w_i q_i(c) = T`. Stephensen derives the unweighted
+relative-entropy-minimizing mean correction; placing the frozen production weights
+in both the constraint and entropy objective gives the direct weighted extension.
+Within attainable support it therefore matches the expected weighted margin
+exactly, is KL-minimal, and keeps every adjusted probability in `[0, 1]` without
+clipping ([Stephensen 2016, §§2-3.2](https://www.microsimulation.pub/articles/00144)).
+Reusing each module uniform with a pinned equality rule would also give a unique,
+deterministic, auditable flip set. Clipping and displaced-set uniqueness therefore
+do not decide against this strong alternative.
+
+Selection still wins for the production target defined here. The expectation
+equation `sum_i w_i q_i(c) = T` does not imply the realized weighted-count equation
+`sum_i w_i 1[u_i < q_i(c)] = T`. The latter is a discontinuous step function,
+especially with unequal target contributions. Choosing `c` against that function,
+or resolving a jump over `T`, is itself a selection rule with an ordering, tie rule,
+and closest-attainable residual. The registered threshold-prefix mechanism controls
+the realized whole-unit outcome directly, discloses the selected prefix and exact
+displaced set, and couples every change to the original module `(p_i, u_i)` without
+a second draw. Logit scaling remains a useful probability-space diagnostic; it is
+not a complete realized-count reconciler.
+
+For a continuous value such as average covered wage, a uniform, disclosed ratio
+remains admissible. It preserves positive-earner ranks, creates no fractional
+person-state, and can match a correctly defined mean exactly apart from pinned
+floating-point tolerance. This narrow exception does not authorize event-probability
+ratios or annual person-weight repair.
+
+## 3. Margins, granularity, and vintage
+
+### 3.1 The first target bundle
+
+The first proposed production bundle is
+`ssa_trustees_2026_alternative_ii_v1`: the 2026 Trustees Report's intermediate
+assumptions, [released June 9, 2026](https://www.ssa.gov/news/en/advocates/2026-06-09.html)
+and retrieved for this design on July 15, 2026. [Alternative II is the Trustees'
+best-estimate set as of February 2026](https://www.ssa.gov/oact/TR/2026/V_A_demo.html).
+The bundle pins the [2026 report](https://www.ssa.gov/oact/TR/2026/), the
+[single-year table index](https://www.ssa.gov/oact/TR/2026/lrIndex.html), and the
+[single-year workbook](https://www.ssa.gov/oact/TR/2026/SingleYearTRTables_TR2026.xlsx).
+It never resolves “current” or “latest” at runtime. A later report creates a new
+bundle identifier and a new content hash; it does not mutate this one.
+
+The manifest declares an applicable target-year interval. The current M6 panel
+builders require the pinned 2022 horizon
+(`src/populace_dynamics/engine/panel_builders.py:120-132`), so this document does
+not make a 2026-forward Trustees run executable. Extending the production horizon
+is a separate dependency and cannot be smuggled into the aligned-`after` producer.
+
+The direct public tables do not contain every micro-event target. In particular,
+Table V.A1's total fertility rate and standardized death rates are summary rates,
+not annual birth counts or age-sex death probabilities. A label such as “aligned
+to V.A1” is therefore insufficient for person selection. The target bundle must
+either pin an official granular OCACT extract or pin a documented, reviewable
+bridge from the public series. If neither exists, that margin remains
+`not_available` and the aligned run does not fabricate it.
+
+### 3.2 Ordered margins
+
+Demographic margins reconcile first because they determine the live roster on
+which economic margins operate. Entrant additions are selected and registered
+before `ProjectionEngine.project`; the loop validates every scheduled entrant and
+reserves IDs before its wave loop (`src/populace_dynamics/engine/loop.py:192-237`).
+Within each year, those scheduled additions materialize at the existing pre-wave
+seam; a separately registered emigration branch runs after addition and before
+mortality; mortality reconciles after raw step 1; fertility after raw step 4;
+covered employment after raw step 6; and covered wage after employment.
+
+The emigration seam precedes mortality to enforce one discrete roster-exit
+transition per person-year: a selected emigrant leaves before the mortality
+candidate ledger is formed and cannot also be counted as a death. It follows
+pre-wave additions because the year's roster must first be materialized. This
+timing does not decide whether same-year entrants are eligible for removal;
+§8 decision 2 must ratify that candidate-universe rule.
+
+| order | margin and target | production granularity | mechanism | binding and limitation |
+|---:|---|---|---|---|
+| 1 | population entry and exit | year × published V.A2 category, then age × sex × atomic unit when a granular bridge is pinned | separate whole-unit addition and removal branches | [Table V.A2, Immigration Assumptions](https://www.ssa.gov/oact/TR/2026/lr5a2.html) publishes projected inflow, outflow, adjustment-of-status, and net columns separately for the LPR and temporary-or-unlawfully-present categories, plus total net change, in thousands. Adjustment of status is an offsetting transfer between categories, not roster entry or exit. The table therefore binds annual gross addition and removal totals at its published-category granularity, but supplies neither age-sex cells nor a Social Security area-to-production-roster bridge. Its within-category semantics also require ratification because LPR outflow includes citizens and the temporary-or-unlawfully-present series is residual-method-based. The margin remains unavailable until §8 decision 2 pins those bridges and semantics. |
+| 2 | mortality | year × sex × single year of age, with a pre-registered pooling ladder for sparse cells | whole-person death selection | [Table V.A1, Fertility and Mortality Assumptions](https://www.ssa.gov/oact/TR/2026/lr5a1.html) supplies standardized summaries; [Table V.A4](https://www.ssa.gov/oact/TR/2026/lr5a4.html) and [V.A5](https://www.ssa.gov/oact/TR/2026/lr5a5.html) supply period and cohort life-expectancy checks. Selection requires a separately pinned official age-sex probability/count extract; summary life expectancy is validation-only. |
+| 3 | fertility | year × mother's single year of age, pooled only through the registered ladder; plurality reported separately | whole maternal-birth-event selection | V.A1 supplies annual total fertility rate. SSA's [2026 demographic assumptions memorandum](https://www.ssa.gov/oact/TR/2026/2026_Long-Range_Demographic_Assumptions.pdf), pp. 3 and 6, describes single-age birth-rate assumptions, but the public single-year table does not expose them. A granular extract or pinned age-pattern bridge is required. |
+| 4 | OASDI-covered work | year total; age × sex only when a definition-matched granular extract is pinned | whole-person covered-worker selection | [Table IV.B4, Covered Workers and Beneficiaries](https://www.ssa.gov/oact/TR/2026/lr4b4.html), workers paid at any time in the year for OASDI-covered employment. [Table V.B2](https://www.ssa.gov/oact/TR/2026/lr5b2.html) reports unemployment and total-employment growth; [the covered-employment methods](https://www.ssa.gov/OACT/TR/2026/V_C_prog.html) explain why its CPS average-week concepts differ from an annual covered-worker count. |
+| 5 | average annual covered-wage growth path | year, positive OASDI-covered workers | one multiplicative scalar after employment selection | [Table V.B1, Principal Economic Assumptions](https://www.ssa.gov/oact/TR/2026/lr5b1.html), nominal and real annual change in average annual wage in covered employment. It does not provide the base covered-wage level needed to construct `T_y`; that binding remains open. [Table VI.G1](https://www.ssa.gov/oact/TR/2026/lr6g1.html) provides AWI and taxable-payroll checks, not interchangeable targets. |
+
+The reconciler does not align every published quantity. Table V.A3 population
+stocks, V.A4/V.A5 life expectancy, V.B2 total employment and unemployment, and
+VI.G1 taxable payroll remain validation checks. Aligning both components and their
+derived totals would conceal accounting errors and overconstrain the panel.
+
+### 3.3 Definitions that must not drift
+
+Every target carries its universe and measurement definition:
+
+- The SSA Social Security area includes more than residents of the 50 states and
+  District of Columbia. The target bundle must publish the bridge to the
+  production roster; it cannot silently call the two populations equal
+  ([SSA 2026 long-range model documentation, Demography p. 2](https://www.ssa.gov/oact/TR/2026/2026_LR_Model_Documentation.pdf)).
+- V.A2's positive lawful-permanent-resident adjustment-of-status amount is the
+  offsetting negative amount in the temporary/unlawfully present category. It is a
+  legal-status transfer; counting either side as roster entry or exit would add or
+  remove people who do not move into or out of the SSA total.
+- IV.B4 counts anyone with covered earnings at any time during a calendar year.
+  [V.B2's employment and unemployment concepts use average-week CPS measures](https://www.ssa.gov/OACT/TR/2026/V_C_prog.html).
+  They cannot share a column or target.
+- V.B1's average covered wage, the national AWI in VI.G1, and taxable payroll are
+  distinct. The alignment scalar targets only the first; the others diagnose it.
+- Estimated, preliminary, or revised historical values stay exactly as published
+  in the pinned vintage. A revised SSA series requires a new bundle and a
+  before/after vintage comparison.
+
+## 4. Production mechanism
+
+### 4.1 Two immutable projections
+
+One production request creates two logically separate results from the same
+initial slice, model commit, draw index, and module RNG registry:
+
+1. `before`: the ordinary `ProjectionEngine.project` result, with no target bundle
+   visible to the engine; and
+2. `after`: a replay through an alignment-aware production wrapper that inserts
+   reconciliation hooks at the named seams.
+
+“Same” means identical immutable inputs and stream addresses, not shared mutable
+objects. Each replay receives a fresh `assemble_period_modules` closure, metadata
+mapping, synthetic-ID allocator, and output collectors. The assembly closes over
+mutable marital, fertility, disability, and household caches
+(`src/populace_dynamics/engine/assembly.py:195-211,230-250`), while the loop places a
+mutable allocator in metadata (`src/populace_dynamics/engine/loop.py:225-237`). A
+shallow copy would allow one path to contaminate the other. The term “alignment
+layer” in `src/populace_dynamics/engine/assembly.py:1-8` names the existing
+native-panel adapter; the external target reconciler designed here is a distinct
+production wrapper around freshly assembled modules.
+
+The word “wrapper” names that production entry point; it does not choose who owns
+the seams. Section 8 decision 9 must select either (a) optional reconciliation and
+frozen-identity-registry hook parameters on the existing eight-step loop, reachable
+only through a production facade, or (b) a parallel production-only loop that
+orchestrates the same module callables in the same order. The first concentrates
+orchestration but changes the certified loop's extension surface; the second leaves
+that loop byte-identical but must prevent the two orchestrations from drifting.
+
+Gate isolation is structural in either architecture. The gate runner calls an
+unaligned-only facade with a closed request schema and allow-listed metadata. That
+request cannot express a target bundle, alignment specification, reconciliation
+hook registry, or frozen replay registry, and its return type cannot represent an
+aligned result. The aligned producer implementation remains outside the gate
+runner's import closure. CI enforces the boundary with dependency and signature
+checks, rejection of unknown alignment-bearing metadata, and a fixed-fixture check
+that canonical gate artifacts are byte-identical when the aligned producer is
+importable but uncalled and when its import is deliberately blocked. A runtime
+`alignment_enabled = false` flag alone is not structural isolation.
+
+The ordinary engine remains the eight-step loop in
+`src/populace_dynamics/engine/loop.py:268-321`. The wrapper does not add a ninth
+certified module or change that order. It captures a step's pre-state, raw
+post-state, candidate probabilities or values, and stable draw keys; it then
+reconciles the raw result before the next existing step reads it. This timing is
+load-bearing: an aligned death must leave before aging, marriage, disability, and
+earnings; an aligned birth must enter before later waves; aligned employment must
+exist before the wage scalar.
+
+The wrapper writes a new `ProjectionResult`-shaped object plus an alignment ledger.
+It never mutates `before`, shares a mutable frame with it, or returns `after` to a
+gate scorer.
+
+### 4.2 Discrete selection algorithm
+
+Each binary margin-year-stratum supplies a candidate ledger with:
+
+`alignment_unit_id`, `person_id` or linked-unit members, `year`, `stratum_id`,
+positive `unit_weight`, positive scalar `target_contribution`, raw outcome `e_raw`,
+predicted probability `p`, and the module uniform `u`. `unit_weight` is the atomic
+unit's pre-reconciliation production year-`y` displacement weight, not the M6
+start-wave gate weight. `target_contribution` is the amount that selecting the unit
+adds to this scalar target cell; for a one-person count it is the frozen year-`y`
+person weight. The two fields are not assumed equal. The registered selection score
+is `p - u`. The raw Monte Carlo outcome is the zero-threshold selection
+(`p - u > 0`), so moving one common threshold preserves the realized latent
+draw-distance ordering. It does not claim to preserve probability-only risk
+ordering.
+
+**Flip-ordering decision -- additive draw distance.** A common-odds realization
+would instead rank candidates by `logit(p) - logit(u)`; for a `0→1` addition its
+crossing factor is `[u / (1 - u)] / [p / (1 - p)]`, with the smallest factor
+crossing first. The additive and odds-distance orders coincide when `p` is constant,
+including within a single age-sex mortality cell under the registered mortality
+model. They can select different people when fertility probabilities vary by parity
+or cohort and in a year-total covered-work pool with heterogeneous probabilities.
+This design chooses additive `p - u` because it measures distance from the engine's
+actual probability-scale uniform comparison; the target bundle does not identify a
+common logit-intercept shock, and additive distance needs no logit-link or
+boundary-infinity convention. The algorithm revision records this choice, and the
+validation panel reports its compositional effects. This is a methodological
+choice, not a claim that the two orderings are interchangeable.
+
+The reconciler:
+
+1. sorts candidates by descending `p - u`, then by a dedicated alignment-stream
+   tie key, then by canonical `alignment_unit_id`;
+2. computes cumulative `target_contribution` for every prefix, including the empty
+   prefix;
+3. chooses the prefix whose cumulative target contribution is closest to the
+   registered target;
+4. breaks equal absolute residuals by fewer flips from `e_raw`, then no overshoot,
+   then the smaller prefix; and
+5. records every `0→1` and `1→0` flip before applying the selected whole-unit
+   outcomes.
+
+This is a deterministic stratified-selection rule, not an iterative search whose
+stopping point can vary. With unequal positive target contributions it may not hit
+an arbitrary real target exactly. A non-prefix subset may be closer, but packing
+units by contribution would abandon the registered threshold rule and
+model-relative ordering. The selected prefix and its residual are unique,
+auditable, and byte-reproducible.
+Section 6 freezes each registered threshold-prefix resolution floor at its seam,
+before that margin's selection.
+
+For a linked unit, `unit_weight` is reported once and the outcome applies atomically
+to every required member. The bundle must separately register each member's person
+weight and the unit's target contribution. Migration-addition candidates use a
+pinned donor pool and donor-distance tier in place of a model event probability;
+within a tier, the dedicated alignment key supplies the order. The bundle must name
+whether the unit is a person, maternal event, couple, family, or household. It may
+not fall back to person selection when that would orphan a relationship.
+
+The scalar prefix rule cannot align an atomic family independently in several
+age-sex cells: one selected family contributes a vector of member weights to those
+cells. Multi-cell linked-unit migration therefore requires a separately registered
+joint vector objective, optimizer, and deterministic tie rule. §8 decision 2
+leaves that mechanism open, so the producer cannot activate family/household
+migration by running independent prefixes that would split or double-count the
+unit.
+
+The positive-prefix schema applies to one-sided events. Migration uses separately
+targeted addition and removal prefixes. V.A2 supplies the gross inflow and outflow
+columns needed to bind those branches at year × published category; its offsetting
+adjustments of status are category transfers, not roster-addition or roster-removal
+events. The producer may not discard those source-bound gross targets and
+reconstruct branches from total net change alone. Activation remains blocked on
+§8 decision 2 because the age-sex and roster bridges, candidate-unit and donor
+rules, within-category flow semantics, and linked-unit vector objective are not
+ratified.
+
+The implementation must expose the complete eligible candidate ledger, not only
+raw positive events. Otherwise it could remove excess events but could not add a
+shortfall without an unregistered redraw. A target outside the attainable range
+is `structurally_infeasible`; the runner publishes the reason and does not issue an
+aligned panel.
+
+That complete ledger is an implementation precondition, not a claim about current
+outputs. `simulate_maternal_births` currently evolves parity internally and returns
+only positive events (`src/populace_dynamics/engine/marital.py:288-352`). A future
+seam must expose each live-mother candidate's probability, uniform, parity, and
+prospective birth order without a second draw; fertility remains `not_available`
+until it does. The current earnings interface returns a continuous draw, not a
+covered-work participation probability and uniform
+(`src/populace_dynamics/engine/steps.py:169-287`); covered-work selection likewise
+remains unavailable until §8 decision 3 pins a lawful score.
+
+### 4.3 Continuous covered-wage ratio
+
+After covered-worker selection, let `R_y` be the panel's definition-matched
+weighted mean of positive annual covered wages and `T_y` the V.B1-derived target
+level for year `y`. Because V.B1 publishes growth, the bundle must construct `T_y`
+from a separately pinned base covered-wage level and the named nominal or real
+growth column. Without that base, the wage margin is `not_available`. The scalar is
+
+`a_y = T_y / R_y`, and `earnings_after_i = a_y × earnings_before_i`
+
+for positive covered earnings only. The ledger publishes `R_y`, `T_y`, `a_y`, the
+number and representation weight of changed workers, and weighted and unweighted
+absolute dollar displacement. If `R_y <= 0 < T_y`, the margin is infeasible; the
+runner never substitutes a zero, cap, or donor value. The bundle pins rounding,
+numeric dtype, and whether self-employment is included before execution.
+
+The wage scalar does not change representation weights, zero earnings, employment
+status, or ranks among positive earners. It does change the chained state read by
+later waves. On an even projection year, the producer applies the same scalar or
+employment override to both current `earnings` and the just-written
+`gen_earn_w2`; it leaves `gen_earn_w4` as the already aligned older lag. On an odd
+year it changes current `earnings` only, matching the no-shift behavior in
+`apply_earnings` (`src/populace_dynamics/engine/steps.py:210-287`). The next even
+draw reads all three fields
+(`src/populace_dynamics/engine/forward_earnings.py:1046-1070`). Every changed chain
+field enters the incremental and canonical audit views. Taxable payroll, AIME,
+PIA, taxes, and benefits are recomputed downstream from aligned state; they are not
+separately forced to SSA totals.
+
+### 4.4 RNG and byte reproducibility
+
+Alignment consumes no existing module stream. A separate registry derives each
+tie stream from canonical integer words encoding the root draw index,
+`alignment_spec_sha256`, calendar year, margin identifier, stratum identifier, and
+algorithm revision. It uses a cryptographic digest, never Python's process-random
+`hash()`.
+
+Within that stream, each candidate's tie value uses a counter-based derivation keyed
+by the canonical typed `alignment_unit_id`; it is never consumed by candidate-frame
+position. Adding or suppressing another candidate therefore cannot change an
+unchanged unit's tie value.
+
+Candidate frames sort by canonical typed keys before scoring. The implementation
+pins float64 arithmetic, stable sorting, target units, missing-value behavior, and
+tie rules. It records the stream address and hashes of `before`, `after`, the target
+bundle, and the displacement sidecar. A bit-exact replay check must reproduce all
+four hashes before the first production release.
+
+Adding or removing a production person must not shift any ordinary person's later
+draws. The current loop pre-registers scheduled-entry IDs before the wave loop but
+assigns newborn ordinals sequentially after each year
+(`src/populace_dynamics/engine/loop.py:192-237,323-331`); person streams use those
+ordinals, not person-ID values (`src/populace_dynamics/engine/loop.py:93-107`;
+`src/populace_dynamics/engine/rng.py:158-169`). Merely choosing stable IDs is
+therefore insufficient.
+
+The producer completes `before` first and freezes a replay identity registry. Every
+ordinary person, including each raw-positive synthetic child, retains the exact
+`person_id`, person ordinal, and auxiliary values used by `before`. Suppressed raw
+events leave their identities and ordinals reserved but inactive. The complete
+candidate schedule assigns every counterfactual entrant or child a deterministic
+ID and ordinal in disjoint ranges above the corresponding maxima in `before`,
+ordered by stable event key; unused reservations remain holes. Common raw-positive
+births reuse the ordinary path's child sex and other auxiliary draws, while a
+counterfactual `0→1` birth uses per-event alignment streams.
+
+The aligned loop must accept that frozen ordinal registry instead of appending
+`len(person_ordinals)`, and child materialization must accept the reserved identity
+and auxiliary record instead of the current length-dependent vector draw and
+sequential allocator (`src/populace_dynamics/engine/steps.py:381-437`). Until this
+production seam passes a zero-flip identity test and an add/suppress isolation
+test, every roster-changing alignment margin is unavailable.
+
+### 4.5 Roster and scoring universes remain separate
+
+Alignment operates on the lawful production roster; it does not redefine a gate
+universe.
+
+- The earnings-domain law marks people who are closed-panel members for flow
+  modules but open additions for the 2014-anchored earnings law
+  (`docs/design/m6_projection_engine.md:1525-1604`). Economic alignment may select
+  only people for whom the production earnings adapter exposes a lawful potential
+  earnings state. It cannot backfill a gated earnings anchor or turn a marked
+  person into gated support.
+- The marital entry law keeps anchor-present minors in the roster and seeds them
+  only when they enter the certified marital risk set
+  (`docs/design/m6_projection_engine.md:1009-1152`). Alignment cannot remove them
+  merely because they were outside a module's year-0 domain.
+- Amendment 3h, merged into `master` by
+  [PR #216](https://github.com/PolicyEngine/populace-dynamics/pull/216), separates
+  the frame-independent fertility schedule used for scoring from the live
+  post-mortality roster used for child materialization. Its law appears in
+  [§2.8.2h at merge commit `0e27be2`](https://github.com/PolicyEngine/populace-dynamics/blob/0e27be2d857719b30e33b556580b4a360808b5e0/docs/design/m6_projection_engine.md#L1183-L1264): only roster-present mothers materialize children; a scheduled birth to a removed mother enters report-only reconciliation. That lawful `roster_absent_births` drop stays in its upstream roster-reconciliation ledger. It is not an external-alignment flip. Fertility candidates are constructed only after the live-mother filter, and alignment must not resurrect or replace a mother to satisfy a target.
+
+Every margin record therefore names `candidate_universe`, `materialization_universe`,
+`measurement_universe`, its corresponding module scoring domain,
+`gate_consumption = none`, and `aligned_values_consumed_by_gate = false`. Counts
+from one universe may not be presented as counts from another.
+
+Whole-unit selection does not automatically reconcile dependent relational state.
+After mortality or emigration, the producer must validate spouse, parent-child,
+household, cached marital, and cached composition references before the next
+consumer. It may apply only a separately ratified dependent-state rule; it cannot
+clear or invent relationships opportunistically. A margin whose roster change
+would leave unresolved dependent state is unavailable pending §8 decision 8.
+
+## 5. Displacement accounting and publication
+
+### 5.1 Two ledgers, one canonical before/after meaning
+
+The existing function's labels make `before` the immutable unaligned scored path
+and `after` the final aligned report-only path. An immediate raw snapshot at a
+later hook is already conditional on earlier alignment, so passing it as `before`
+would make the function's `scored_path = "unaligned"` false. The producer keeps two
+separate ledgers:
+
+1. a **canonical path-comparison ledger**, derived from the immutable unaligned
+   companion and final aligned result and passed to the existing function; and
+2. an **incremental intervention ledger**, comparing raw and reconciled outcomes
+   inside the aligned replay and owned entirely by the production caller.
+
+Raw projection panels cannot serve as the canonical pair for roster-changing
+events: deaths remove rows, while births and entrants add them. The producer builds
+a common audit keyspace over the union of stable
+`(comparison_unit_id, year, margin_id, stratum_id)` keys. Initial people retain
+their canonical IDs; entrants use donor-event IDs; synthetic children use the
+event-derived identities required by §4.4. Each path receives every union key and
+explicit numeric `eligible` and `present` indicators. This is a typed audit view,
+not a padded live panel.
+
+Each registered measure has one unit and a defined absence contribution:
+
+| measure | unit and absence rule |
+|---|---|
+| `eligible` or `present` | binary; zero explicitly means the keyed unit is in the union but not eligible/present on that path |
+| `represented_presence` | production-year represented people; `present * reference_weight`, accompanied by `present` |
+| `event_selected` | binary event outcome; accompanied by path-specific `eligible` |
+| `target_contribution` | the event's signed or unsigned contribution in the declared target unit; zero for an unselected event or explicit ineligibility, distinguished by `event_selected` and `eligible` |
+| `covered_wage_contribution` | current covered-wage contribution; zero for explicit absence/noncoverage, with those indicators retained |
+
+Undefined underlying state never becomes zero. The named contribution fields have
+zero as part of their definition, and their presence/eligibility indicators make
+the reason observable. `reference_weight` comes from the stable unit registry and
+is identical in both union views: an existing person uses the frozen year-`y`
+production weight, a prospective child uses the live mother's frozen year-`y`
+weight that it will inherit, and an entrant uses the registered donor-member
+weight. The producer rejects a cross-path mismatch, null keys, null or non-finite
+numeric values, non-positive registered weights, duplicate keys, or an unmappable
+live row before calling the existing function. The function's own exact-key check
+then remains a second line of defense.
+
+The producer invokes
+[`build_alignment_displacement`](../../src/populace_dynamics/harness/m6_reporting.py)
+separately for each margin and **homogeneous measure**, across all requested years.
+The implementation at
+`src/populace_dynamics/harness/m6_reporting.py:321-415` remains authoritative. A
+binary event, represented-person contribution, materialized-person count, and
+dollar value never share one call: its rowwise cross-field maximum would mix
+units. Each single-unit call retains native `per_year_maximum`, run
+`maximum_alignment_displacement`, and frame-global `n_intervened_rows`. The caller
+computes per-year counts and sums. It publishes no scalar maximum across
+incommensurate margins.
+
+### 5.2 Incremental per-margin, per-year attribution
+
+Every aligned run publishes one caller-owned record for every requested
+`(margin_id, year, stratum_id)`, including records with no direct intervention.
+Here `raw` means the raw module outcome conditional on all earlier aligned hooks;
+it is never labeled as the gate-scored path. Each record contains:
+
+- target, raw, and aligned weighted totals in one declared unit;
+- `implied_alignment_factor = target_weighted_total / raw_weighted_total` for every
+  `computed` cell with a nonzero raw total; when the raw total is zero the field is
+  JSON `null`, never infinity or a fabricated zero or one, and
+  `implied_alignment_factor_reason` distinguishes `zero_raw_zero_target` from
+  `zero_raw_nonzero_target`;
+- raw residual, aligned residual, and the precomputed threshold-prefix resolution
+  floor;
+- counts of `0→1` and `1→0` unit flips;
+- distinct affected decision people, added/removed materialized people, and
+  displaced atomic units, with a margin-specific role code;
+- `displaced_person_count`, `gross_displaced_person_weight`,
+  `displaced_unit_count`, and `gross_displaced_unit_weight`, plus separate `0→1`
+  and `1→0` fields;
+- absolute target-contribution displacement and its maximum over a single unit;
+- for wages, the scalar, changed-worker count and representation weight, and
+  weighted and unweighted absolute dollar displacement;
+- references to the canonical path-comparison results and hashes of the
+  incremental pre/post hook views; and
+- status, reason, candidate/materialization/measurement universes, target source,
+  stream address, and algorithm revision.
+
+The role code defines whose state changed. Mortality and covered work name the
+person; a fertility event names the mother as decision person and the child as a
+materialized addition; an entrant family names every entering member and one
+atomic donor unit. `displaced_person_count` is the distinct union of these affected
+people, never a rounded weight. `gross_displaced_person_weight` sums each affected
+person's frozen pre-reconciliation production-year weight once. For a not-yet-live
+addition, it uses the same pre-registered child or donor-member `reference_weight`
+defined for the canonical views.
+`gross_displaced_unit_weight` separately sums each changed atomic unit's registered
+unit weight once. Adds and removals are never netted. The ledger also reports
+decision-person and materialized-person subtotals so a maternal event does not
+silently count a mother and child as interchangeable observations.
+
+These sums supplement rather than reinterpret the canonical function. Its
+`n_intervened_rows` is not relabeled as a per-year person count, a maximum change
+in `event_selected` is not relabeled as aggregate displacement, and the canonical
+path difference is not claimed to be direct causal attribution to one hook. The
+run summary keeps both the path-comparison and incremental records for each margin
+and year.
+
+### 5.3 Artifact labels and completeness
+
+An aligned artifact is publishable only if at least one requested margin is
+`computed` and every requested margin has either:
+
+- `computed`, with a complete displacement ledger and fidelity result; or
+- `not_available`, with a pre-run reason such as a missing definition-matched
+  granular target.
+
+`not_available` does not mean that the producer silently skipped an advertised
+margin. The artifact title and machine label list the margins actually applied.
+If no margin can be applied, the producer publishes only an unavailable-attempt
+sidecar; the inherited displacement surface remains
+`status = "not_computed"`, `reason = "alignment_layer_output_not_collected"`, and no
+panel is labeled `aligned_production`.
+`structurally_infeasible`, audit-view key mismatch, materialization mismatch, or
+hash mismatch withholds the aligned panel and publishes a failed production
+attempt sidecar. It never converts the unaligned companion into a failure of the
+M6 gate.
+
+Every aligned publication carries this statement, without abbreviation:
+
+> This production scenario was reconciled to pinned external targets. The M6 gate
+> scored the separate unaligned projection. Alignment displacement is disclosed
+> for every applied margin and year, and this aligned output does not transfer or
+> strengthen certification.
+
+The machine-readable companion fields are
+`projection_mode = "aligned_production"`, `alignment_applied = true`,
+`gate_scored = false`, `certification_transfer = false`, and
+`unaligned_companion_sha256`. Removing the label or publishing the aligned panel
+without its displacement sidecar is an artifact-schema failure.
+
+## 6. Evaluation: floors first
+
+### 6.1 Pre-selection threshold-prefix resolution floors
+
+The target bundle, candidate-construction and weight rules, strata, pooling ladder,
+minimum-candidate rule, score, tie rule, and numeric rules are frozen before an
+aligned replay. The realized candidates for a later margin can depend on earlier
+aligned hooks, so the producer materializes and hashes that margin's ledger only at
+its named seam. Before selecting any prefix, it enumerates the empty and every
+ordered prefix and records the minimum prefix residual in target units. That is the
+**selection-resolution floor** for the discrete margin-year-stratum. It is not the
+minimum over every subset. The producer separately records whether the target lies
+outside the empty-to-total candidate range.
+
+The floor still comes first at each seam: the producer writes and hashes the
+realized ledger and floor before selecting the aligned prefix, and the evaluator
+reads rather than rewrites them. It may not declare exact agreement when the
+registered threshold rule cannot achieve it. A continuous covered-wage margin has
+a zero arithmetic floor only when its denominator is positive and all registered
+finite-value conditions hold; otherwise it is infeasible.
+
+Sparse cells follow only the target bundle's pre-registered pooling ladder and
+pre-registered minimum-candidate rule. At the seam, the producer decides pooling
+from the realized candidate count before inspecting
+prefix residuals, publishes both cell definitions, and may not pool merely because
+a non-prefix subset would fit better. A target outside the candidate range remains
+`structurally_infeasible`; pooling cannot manufacture eligible people.
+
+### 6.2 Alignment-fidelity check
+
+For every applied cell, the evaluator independently recomputes the aligned margin
+from the materialized panel using the declared measurement universe. The check
+passes only when:
+
+`abs(aligned_total - target) <= selection_resolution_floor + numeric_tolerance`.
+
+The bundle pins `numeric_tolerance` in target units; it cannot be a percentage
+chosen after execution. The evaluator also requires raw and aligned ledger totals
+to reproduce from their hashed audit views, selected units to reproduce from the
+registered ordering, and materialized events to reconcile to selected events.
+
+This is a production-integrity check, not a scientific gate. Failure withholds the
+aligned output and leaves the unaligned gate artifact and verdict unchanged.
+Passing says only that the disclosed reconciler produced the pinned scenario.
+
+### 6.3 Displacement-magnitude corridors
+
+Fidelity alone rewards a machinery that forces any target, however violently.
+Every applied cell therefore receives a report-only displacement corridor for:
+
+- displaced-person share of the eligible candidate universe;
+- gross displaced representation-weight share;
+- absolute target-contribution displacement relative to the target; and
+- for wages, the absolute log scalar and weighted absolute dollar displacement.
+
+Corridors are authored before release from pre-boundary historical backcasts,
+definition-matched target revisions, and registered synthetic target shocks. Their
+ceremony pins source vintages, years, horizons, strata, seeds, quantile rule,
+pooling ladder, and artifact hash. The labels are `within_reference`,
+`above_reference`, and `not_comparable`; they are not pass/fail labels. An
+`above_reference` result triggers model review and must appear beside the aligned
+series, but it neither invalidates nor rescues the unaligned M6 gate.
+
+No numerical corridor is ratified by this design document. Until a reviewed
+corridor artifact exists, every aligned run publishes the raw magnitude and
+`corridor_status = "not_available"`; it must not substitute zero, an informal
+percentage, or the selection-resolution floor. This preserves the floors-first order while
+keeping a missing research benchmark visibly missing.
+
+### 6.4 Validation panel
+
+The release comparison contains the unaligned, target, and aligned series on the
+same axes. It also compares untargeted distributions before and after alignment:
+age, sex, family structure, earnings quantiles, covered-work transitions, life
+expectancy, population stock, AWI, and taxable payroll where definition-matched.
+The evaluator reports changes in these diagnostics even when every target margin
+passes fidelity. This implements the target-fit and distribution-distortion split
+recommended by [Li and O'Donoghue (2014), §§5.3-5.8](https://www.jasss.org/17/1/15.html#5.3).
+
+A pre-boundary validation exercise turns alignment off for the module under test
+and compares its raw projection with observed/reference data. That follows the
+dynamic-microsimulation validation guidance to retain current, baseline, and
+reference outputs and switch alignment off for the target module
+([Harding et al. 2010, pp. 53-54](https://www.microsimulation.pub/articles/00038)).
+It is evidence about model behavior; the production fidelity check is evidence
+about reconciliation behavior. The report never collapses the two.
+
+## 7. External bindings and artifact identity
+
+### 7.1 Target manifest
+
+The target bundle is a content-addressed manifest. Each source series records:
+
+- bundle identifier and schema/algorithm revision;
+- publisher, report title, report date, Trustees alternative, table, row, column,
+  year range, and source URL;
+- retrieved-at timestamp, raw-file SHA-256, extraction-code commit, extracted-data
+  SHA-256, and any signed official-extract identifier;
+- unit, nominal/real basis, frequency, event timing, population universe, geography,
+  status/coverage definition, age/sex categories, and revision status;
+- transformation from source to annual target level, including the base level for
+  a growth series, rounding, and reconciliation identities;
+- bridge or pooling-ladder identifier and hash where the public table is not the
+  micro-event target; and
+- `available`, `not_available`, or `validation_only`, with a reason.
+
+The manifest pins the 2026 Alternative II vintage described in §3.1. Network
+access is forbidden during a production run; the runner accepts only the reviewed
+local content hashes. Redirects, revised spreadsheets at the same URL, or a new
+Trustees Report fail identity rather than updating a target in place.
+
+### 7.2 Production manifest
+
+The aligned sidecar binds the target manifest to the unaligned scientific object:
+repository commit, environment lock hash, initial-slice hash, draw index, ordinary
+module RNG registry hash, alignment RNG registry hash, algorithm revision,
+candidate-ledger hashes, selection-resolution-floor hash, corridor hash or explicit null
+reason, unaligned result hash, aligned result hash, and displacement-ledger hash.
+It also records the exact ordered margins and the seam after which each ran.
+
+Publishing a different target vintage, bridge, pooling ladder, candidate universe,
+or algorithm creates a different production identity. Comparison tooling may show
+two identities side by side, but it cannot call one a rerun of the other.
+
+## 8. Open decisions
+
+Implementation remains blocked on the following bounded choices. None authorizes
+a gate-path change.
+
+1. **Granular demographic target authority.** Obtain a reviewable official OCACT
+   age-sex mortality series and single-age maternal fertility series, or ratify a
+   public age-pattern bridge and sparse-cell pooling ladder. V.A1 summary rates
+   alone are insufficient.
+2. **Social Security area and migration units.** Ratify the bridge between the
+   production roster and the Social Security area and the age-sex/unit allocation
+   from V.A2's published LPR and temporary-or-unlawfully-present inflow and outflow
+   columns. Treat adjustment of status as a category transfer; specify how LPR
+   outflow's included citizens and the residual-method temporary-or-unlawfully-
+   present series map to roster additions and removals. Pin donor microdata and
+   vintage, the atomic entrant/removal unit, and whether same-year entrants enter
+   the removal-candidate universe. For multi-person units, pin the joint vector
+   objective and tie rule across age-sex cells.
+3. **Covered-work candidate state.** Define a lawful potential covered-earnings
+   value, participation probability/uniform or alternate registered selection
+   score for a raw nonworker selected `0→1`, and coherent chain-state updates.
+   Settle wage-and-salary versus self-employment coverage and bind the result to
+   IV.B4's “paid at any time in year” concept without altering the certified
+   earnings domain.
+4. **Fertility measurement.** Choose maternal birth events, live births, or
+   children as the target unit; expose the complete probability/uniform schedule;
+   specify parity evolution, birth order, plurality, and infant-death timing; and
+   bind the measure to post-mortality live-roster materialization without changing
+   the frame-independent scoring schedule.
+5. **Displacement-corridor ceremony.** Pre-register the historical window,
+   horizon bands, strata, target-revision vintages, synthetic shocks, seeds,
+   quantiles, and response to `above_reference`. This design intentionally
+   ratifies no numerical corridor.
+6. **Policy-counterfactual semantics.** Decide whether an aligned reform run uses
+   the baseline Trustees levels, aligns only common exogenous margins, or carries
+   reform-specific external targets; specify common-random coupling and labels so
+   alignment cannot erase a modeled policy effect.
+7. **Release scope, horizon, and wage base.** Choose the first target-year interval
+   and complete the separately reviewed horizon extension beyond the current 2022
+   M6 panel builders. Pin the covered-wage base level/year and nominal-or-real V.B1
+   growth column. Decide whether demographic margins ship before economics or the
+   release waits for the full ordered bundle.
+8. **Dependent relational state.** Ratify how mortality and emigration update
+   spouse, parent-child, household, marital-core, and household-composition state,
+   including amendment 3h's flagged removed-spouse case, without rewriting a
+   certified scoring schedule.
+9. **Seam ownership and structural gate isolation.** Choose hooks in the existing
+   loop behind the unaligned-only facade or a parallel production-wrapper loop.
+   Name the owner of every reconciliation seam and frozen-identity-registry handoff,
+   and pin how orchestration drift is prevented. Ratify the closed gate request and
+   metadata schemas, production-module import boundary, dependency/signature
+   checks, and byte-identical present-versus-import-blocked gate-artifact check
+   required by §4.1.
+
+## 9. What this does not change
+
+This design does not:
+
+- choose or implement either seam-ownership architecture in §8 decision 9; this
+  docs-only change does not edit the eight-step engine loop, assembly law, current
+  gate runner, gate configuration, threshold, holdout floor, one-shot rule, or
+  certification claim;
+- change the M6 decision that the gate scores the unaligned projection;
+- replace or alter `build_alignment_displacement`; it supplies the missing lawful
+  `after` producer and exact-key inputs around the existing function;
+- ratify a re-drawn-`T*` seed margin, build successor forward-seed machinery, or
+  convert any other null-with-reason surface to a numeric result;
+- redefine the roster, resurrect dead people, create a gated state for family-B
+  additions, or collapse the scoring/materialization universes in amendment 3h;
+- recalibrate panel weights, use target matching to certify a module, or align
+  validation-only derived totals; or
+- claim that citing an SSA aggregate supplies a missing granular target, bridge,
+  donor pool, corridor, or production implementation.
+
+The next implementation proposal must resolve the relevant open decisions, add
+code and tests in a separate change, and demonstrate the bit-exact replay,
+alignment-fidelity, materialization-reconciliation, artifact-label, and
+unaligned-gate-isolation checks. Until then, target margins may remain
+`not_available`, while the inherited displacement surface remains
+`status = "not_computed"` with
+`reason = "alignment_layer_output_not_collected"`; the certified projection
+remains unaligned.