diff --git a/docs/design/mortality_ses_sm2_freeze.md b/docs/design/mortality_ses_sm2_freeze.md
new file mode 100644
index 00000000..96db6385
--- /dev/null
+++ b/docs/design/mortality_ses_sm2_freeze.md
@@ -0,0 +1,1983 @@
+# SM2 freeze: sources, mappings, joint route, paths and conformance
+
+*Design record, 1 November 2026, branch `mgf-sm2-freeze-20261101`, written
+from `origin/master` `1808285925a069143047ae7e7bb2412012db0ad7` by the
+mortality, groups and financing lane's SM2 freeze job (Subfleet build,
+Claude Code, Opus 5.5).*
+
+*Revised 3 November 2026 by the same lane's fix round 1 (Subfleet build,
+Claude Code, Opus 5.5), which answers independent review r1 and changed
+only this file; §13.6 records what that round read and fetched.*
+
+## 1. Status and authority
+
+**This record freezes SM2 under section 5.1 of the lane design**
+(`docs/design/mortality_groups_financing.md`): "Source hashes, module
+versions, seeds, populations, estimands, tolerances and alternatives are
+frozen before any real-data outcome is computed." **It is not a
+registration on #42, and it claims no blind status.** The job that wrote
+it read every source it freezes, so it and this file count as exposed to
+those sources; nothing here feeds a blind role.
+
+- **Verification class.** External-target conformance, labeled
+  **"calibrated, not validated"** (design section 3.2): SM2 admits that
+  the fitted relative risks reproduce the targets the fit used,
+  never that the model predicts. SM2 claims **no prospective prediction
+  origin and no unused holdout** (S0.1 section 2.4). The held-out test of
+  the gradient is SM3.
+- **Authority.** Max approved the lane design, with its methodology, on
+  23 October 2026 (decision d1922). S0.1's audit
+  (`docs/design/s0_1_source_audit.md`, #588, merged 1 November 2026) is
+  this record's input; its section 2 is SM2's record. Its header says a
+  review and the gate's registration "confirm or amend" each
+  **[proposed]** item before anyone computes an outcome, and "may tighten or
+  rescope it but never loosen it". This record tightens or rescopes in
+  the places section 2's table names, and loosens nothing.
+- **Rulings that bind this record, unamended.**
+  - **Q4:** within each single age × sex × year cell, SM1's normalizer
+    scales relative risks to the baseline's expected deaths
+    (`src/populace_dynamics/mortality_ses/normalizer.py:328–382`).
+  - **Q5:** lifetime earnings is the primary dimension; education and race
+    and Hispanic origin are further dimensions of the same model.
+  - **Q6:** the gradient holds at its last observed level as the primary
+    path; a widening path is a registered alternative.
+  - **Q11:** SM owns death alignment.
+- **What it computes.** Nothing. This record states no fitted
+  coefficient, no contrast between categories, no life expectancy and no
+  comparison between sources' gradients, and it prints no source table
+  value. It names tables, pages, age bands, years and categories only.
+  Every number in it is a year, an age, a page or line, a count of table
+  cells, targets, files, PDF objects or drawing operators, a file's byte
+  size or hash, a code from a codebook, a tolerance the design or S0.1
+  sets, a rule, constant or seed this record sets, or a constant it
+  derives from those (the Bonferroni critical values and the widening
+  they imply).
+- **Data.** Neither the freeze job nor its fix round opened a PSID, HRS,
+  SIPP, NHIS or ACS microdata file, and neither computed anything on real
+  data. PSID use follows the project's data-access record. This record
+  uses and compares no DYNASIM, Urban Institute or PWBM material, which
+  reaches these gates only through a validation lane (S0.1 scope). CBO
+  publication 49659 stays closed; its methods count as open until a
+  validation lane supplies a bounded definitions-only extract.
+- **Markers.** **[frozen here]** marks a decision this record makes.
+  **[frozen, S0.1 §x]** keeps a decision S0.1 already froze, with its
+  pointer. **[open]** marks an item this record leaves unsettled and names
+  what settles it.
+
+## 2. Item table
+
+S0.1's section 2 runs from line 324 to line 707 of the audit at
+`18082859`. This command recounts its bracketed markers that contain
+"proposed" or "open":
+
+```
+sed -n '324,707p' docs/design/s0_1_source_audit.md \
+  | grep -oE '\[[^]]*(proposed|open)[^]]*\]' | wc -l
+```
+
+It returns **20**. The rows S2-01 to S2-20 below take them in the audit's
+order. A word search over the same lines
+(`awk 'NR>=324 && NR<=707 && /(^|[^a-z])(open|proposed|pending|unsettled)([^a-z]|$)/'`)
+finds five more lines that call something open or proposed without a
+bracketed marker; rows S2-21 to S2-25 cover them, so the table has
+**25 rows**. Where a row says "tightened" or "rescoped", the change is
+narrower than S0.1's proposal or adds a requirement; no row loosens one.
+
+| Row | Audit line | S0.1 item | Resolution | Where | State |
+|---|---|---|---|---|---|
+| S2-01 | 424 | **[proposed]** mapping assumption: OACT rate ratios used directly as annual-q multipliers, reported with the dynamic-AIME versus frozen-rank distinction and the beneficiary population difference | Confirmed and tightened: adds Study 129's exposure convention (deaths get a full year of exposure, which makes each published rate an initial-exposure estimate of an annual probability, p. 5 and Appendix B pp. 13–14), the age-basis mapping, and three named mapping assumptions (M-E2 to M-E4) that every report carries | §4.1 | frozen here |
+| S2-02 | 440 | **[open]** availability of earlier NCHS education editions | Settled by the NVSR listing and twelve earlier *Deaths: Final Data* reports: 2010–2017 cover reporting states only (31 rising to 49 states plus DC); 2018–2020 are national in the internet tables I–7 and I–19; 2021–2022 are national in printed Table 14 and Table VIII; the listing of 1 November 2026 holds no edition for 2023. The fit uses 2018–2022; 2010–2017 form a labeled sensitivity. A fitted year whose layout fails the loader's check counts as unobserved under §7.5, labeled (fix round 1) | §3.2 | frozen here |
+| S2-03 | 474 | **[open]** harmonization of earlier NHIS years | Rescoped to NHIS 2009–2018 sample adults. The 2009, 2016 and 2018 layouts carry `EDUC1`, `HISPAN_I`, `RACERPI2`, `AGE_P`, `SEX`, `SRVY_YR`, `INTV_QRT` and `HHX`/`FMX`/`FPX` with identical code lists; the design variables are `STRAT_P`/`PSU_P` in 2009 and `PSTRAT`/`PPSU` in 2016 and 2018. The fit job checks each intervening year's layout against these names and code lists before use; a year that fails leaves the pool, labeled "NHIS year dropped: layout", and if no year passes, `SOURCE_UNAVAILABLE` (ii) applies (fix round 1) | §3.4, §9.3 | frozen here (window and variables); intervening layouts checked by the fit job, with the failure outcome frozen |
+| S2-04 | 481 | **[open]** a three-dimension joint gradient conditional on lifetime earnings | Resolved by the joint route: joint marginal calibration on a public reference population, with the earnings–education–race dependence taken from a stated proxy. The dependence itself stays unidentified in public data, and §6.4 says so. Where the fit applies the proxy's dependence outside its estimation sample, this record names the copy assumption M-J1, with sensitivity JR-C-IND (§6.5, §6.6) | §6 | frozen here (route); the dependence stays unidentified and named |
+| S2-05 | 511 | **[open mapping]** Cristia's odds ratios | Frozen: predicted probabilities on a named reference population, then ratios to `q3`'s probability. Cristia enters only as a sensitivity and a source reported beside the fit; no log odds ratio enters as a log relative risk | §4.4 | frozen here |
+| S2-06 | 520 | **[proposed]** CBO 41681 as context only | Confirmed. It supplies the reason the widening alternative leaves race and Hispanic origin flat (pp. 1–2), with p. 1's qualifier that the black–white gap held for men and narrowed for women, and no coefficient or target | §3.1, §7.4 | frozen here |
+| S2-07 | 538 | **[proposed schema]** dimensions, columns, categories, references and unclassified treatment | Confirmed against `mortality_ses/specification.py`; tightened to one unclassified string per dimension, a fixed category order, and a crosswalk per source. `R_ref`'s persons under 25 hold the reference categories by rule, labeled `reference_below_25` (§6.2, §8.1). The ACS crosswalk maps every `SCHL`, `HISP` and `RAC1P` code in the 2010–2019 and 2021–2024 dictionaries; 2020's codes stay open until the fit job's first step (§8.2). NHIS `HISPAN_I` 00–11 map to `hispanic`, following S0.1's frozen reading, with 10 and 11 unclassified in X-R1 | §8 | frozen here (2020's ACS codes: checked by the fit job first) |
+| S2-08 | 564 | **[proposed loader]** OACT contrasts `log(r_q) − log(r_q3)`, band expansion, a source-year matrix that says what happens before a band first appears | Confirmed; tightened with the band-backcast rule, the figure-year rule and two independent table transcriptions | §3.1, §4.1, §5 | frozen here |
+| S2-09 | 570 | **[open]** earnings ages below 62 and 90–120 | Frozen: the 62–64 coefficient holds down to age 25 and the age-25 value below; the 85–89 coefficient holds up to 120; four registered sensitivities. `R_ref` holds everyone under 25 at the reference categories, so the values below 25 reach no one in the fit (§6.2) | §5, §6.2 | frozen here |
+| S2-10 | 579 | **[open]** terminal 100+ rule and any unsupported other-group coefficient | Frozen: race contrasts at 100–120 hold the age-99 contrast, and the fit reads no contrast from the 100+ interval; the e65 targets read that interval's `Tx` and `lx` for NCHS's closing only (§9.2); `other_non_hispanic` comes from NHIS–LMF, with NCHS's AIAN and Asian tables reported beside as a partial group | §4.3, §5, §9.2 | frozen here |
+| S2-11 | 587 | **[open]** extraction and digitization of the 2022 gradient, its uncertainty, complete coverage | Frozen: vector extraction of Study 129 Figures 7–12, validated against Tables 7–12 at every table year, with a digitization half-width the validation sets; complete tables at every knot | §7.1–7.3 | frozen here |
+| S2-12 | 591 | **[open]** widening `CoefficientPath`: slope, origin, ages, caps; knot years and synchronization | Frozen: origin 2024, slopes from a frozen estimator on OACT's pre-pandemic annual figures, caps of 30 years and a 1.5 multiplier, race held flat; annual knots 2010–2024 with one synchronization rule | §7.4–7.5 | frozen here |
+| S2-13 | 602 | **[proposed]** `unclassified_rule="reference"` for SM3's overall denominator | Confirmed for SM3; `refuse` for conformance runs and for any projection until the frame carries the attributes. On `R_ref`, `refuse` applies after the under-25 rule and the ACS crosswalk, which map every code the dictionaries read list (§6.2, §8.2) | §8.3 | frozen here |
+| S2-14 | 606 | **[open: identification]** marginal contrasts do not determine mutually adjusted coefficients; freeze a joint route and its reference distribution | Five candidate routes and one unavailable route; primary JR-C; JR-A (plain addition, labeled), JR-B, JR-D and JR-E as sensitivities. NHIS targets compare as age-standardized hazard ratios on `R_ref(t)`'s single-age population, under named assumption M-H1 (§4, §4.5). The ordered logit fits once per knot year and sex on that year's ACS, and again inside every replicate, with a frozen optimizer, convergence rule and tie rule (§6.2). The ACS income item passes a frozen acceptance rule, and JR-C-W covers the proxy's weakness for women (§6.2, §6.5) | §4, §6 | frozen here |
+| S2-15 | 629 | **[open]** exact release dates of any earlier NHIS cohort used | Left open. The linked-mortality `.dat` files carry the listing date 26 April 2022; each NHIS year's release date settles from that year's survey description, which the fit job records. Nothing in SM2 depends on it, because SM2 claims no prospective origin | §3.4 | left open; settled by the fit job's per-year release ledger |
+| S2-16 | 641 | **[frozen tolerance; proposed operationalization]** e65 target from source ratios on one baseline, with a fixed reference population, marginalization and tail rule | Frozen; the 0.1-year limit stays. Corrected in fix round 1 to keep NCHS's age and closing conventions, which S0.1 requires (audit lines 647–648) and the first draft dropped: the race and Hispanic-origin targets close NCHS's "100 and older" interval as NCHS does, and the model side keeps the repository's closing, so a mismatch in the tail reaches the test. The earnings targets keep the fixed tail rule S0.1 asks of OACT (audit lines 643–646) | §9.2 | frozen here (operationalization); tolerance frozen, S0.1 §2.5 |
+| S2-17 | 649 | **[open]** whether the different source populations and baselines can satisfy the targets together | This record freezes the procedure: one joint calibration per knot on one reference population, and a named failure when it cannot reproduce every target. Whether it can is an outcome, so this record does not settle it | §6.3, §9 | left open; settled by the fit under §9's rule |
+| S2-18 | 659 | **[open until the loader plan is frozen]** the critical family's integer size and which joint-survey targets it holds | Frozen: **6,884** targets, itemized | §9.4 | frozen here |
+| S2-19 | 677 | **[proposed named procedure]** `SM2_SOURCE_RECONSTRUCTION_AND_CLUSTER_BOOTSTRAP` | Filled in step by step | §9 | frozen here |
+| S2-20 | 688 | **[proposed suppression]** at least 100 unweighted persons per published microdata subgroup | Confirmed and tightened: also at least 20 deaths in any published death-based cell, and the rule applies to ACS as well as NHIS aggregates | §11.5 | frozen here |
+| S2-21 | 353 | Unmarked: CBO 49659's "safely bounded definitions remain open" | Left open as the brief requires; SM2 uses no part of 49659 | §3.1 | left open; settled by a validation lane's definitions-only extract |
+| S2-22 | 523 | Unmarked: "CBO 49659's mixed-source methods remain open pending a safely bounded definitions-only extract" | Same item as S2-21 | §3.1 | left open; as S2-21 |
+| S2-23 | 588 | Unmarked: "A 2020 table-only path is a proposed registered sensitivity" | Confirmed as sensitivity P-TABLE; it cannot become primary without amending Q6 | §7.3 | frozen here |
+| S2-24 | 669 | Unmarked: a source that does not publish the uncertainty a target needs leaves that item "open or inconclusive" | Applied: OACT publishes no deaths or exposures, so OACT's random-variation component stays open, and every OACT e65 member stays inconclusive (or fails, if its point residual lies outside ±0.1 year) until the component closes. That keeps lifetime earnings, and with it SM2 as a whole, from admission. The fit reports a Poisson approximation from `R_ref`'s recipients beside each member, labeled as likely understating the variation; it changes no verdict. Applied too to the public LMF's perturbation, which no public file lets the fit assess: that sensitivity stays open, outside every bound as U4 always was, and every NHIS result carries a label (fix round 1, low finding 5) | §9.3, §9.5, §10 | left open; settled by OACT exposure or death counts by sex × band × quintile × year; the perturbation, only by restricted-use linked files |
+| S2-25 | 698 | Unmarked: "`SOURCE_UNAVAILABLE` / open identification applies to the complete modern three-dimension joint source and missing source-age mappings" | Each trigger and fallback stated | §10 | frozen here |
+
+**Counts.** Of the 25 rows, this record freezes 20 and leaves 5 open, each
+with what settles it (S2-15, S2-17, S2-21, S2-22, S2-24); none goes to Max.
+S2-24's open item keeps lifetime earnings inconclusive, and with it SM2's
+admission, until OACT's counts close it (§9.3); the same row leaves the
+public LMF's perturbation sensitivity open. S2-02, S2-03 and S2-07 carry
+checks the fit job performs (the NCHS education layouts, the intervening
+NHIS layouts and 2020's ACS codes), each with a failure outcome this
+record freezes. Section 12 holds two rules for cases outside this table.
+
+## 3. Source-year matrix
+
+This job read every source below at the SHA-256 that section 13 lists.
+"Fit" means the source supplies a target the joint calibration
+reproduces; "beside" means the record reports the source next to the fit
+and fits nothing to it; "context" means it informs a reason and supplies no
+number;
+"reference" means it supplies the reference population, not a target.
+S0.1's design line requires at least one source not used in fitting,
+reported beside the fit; this record names one or more for each dimension
+(§3.6). **[frozen here]** for every role and year rule in this section.
+
+### 3.1 Lifetime earnings
+
+| Source | Table or figure, PDF page | Categories | Sex | Age bands | Years published | Role |
+|---|---|---|---|---|---|---|
+| OACT Actuarial Study No. 129 (December 2024, Figure 6 corrected June 2025) | Tables 7–12, relative mortality ratios (RMRs), pp. 27–32 (PDF page equals printed page) | Lowest, 2nd, 3rd, 4th and Highest AIME quintile, Total, and a share row ("Percentages represent the share of workers in that sex and age group") | Male, female | 62–64, 65–69, 70–74, 75–79, 80–84, 85–89 | 1995, 2000, 2005, 2010, 2015, 2020 | Fit at 2010, 2015 and 2020 |
+| Study 129 | Figures 7–12, RMRs by band, pp. 33–38, one figure per band with a men's and a women's panel and five series | Five quintiles | Male, female | One band per figure | Each calendar year 1995–2022 | Fit at 2011–2014, 2016–2019, 2021–2022 (fully supported bands only) |
+| Study 129 | Tables 1–6, death rates, pp. 15–20 | As Tables 7–12 | Male, female | As Tables 7–12 | As Tables 7–12 | Cross-check only (§9.1): ratios of printed rates must agree with printed RMRs within both rounding intervals |
+| Study 129 | Figures 1–6, death rates, pp. 21–26 | Five quintiles | Male, female | One band per figure | 1995–2022 | Not used |
+| OACT Actuarial Study No. 124 (April 2018) | Tables 4–8, RMRs by AIME quintile, pp. 11–15; Tables 9–13, hypothetical-worker intervals, pp. 21–25 | Five quintiles and Total; five intervals | Male, female | 62–64 to 80–84 and a Total column | 1995, 2000, 2005, 2010, 2015 | Beside (vintage revision of the same quantity on the June 2017 MBR, birth cohorts 1930 and later, p. 6) |
+| Cristia (2007), CBO Working Paper 2007-11 | Tables 3–4, age-standardized mortality ratios, PDF pp. 38–39; Tables 5–6, logistic regressions, PDF pp. 40–41 | Lifetime-earnings quintiles (own and household); race and ethnicity; education; marital status | Men (Tables 3, 5), women (Tables 4, 6) | 35–75, 35–49, 50–64, 65–75 | Person-years 1983–2003 | Beside, and sensitivities JR-D (§6) and T-E2 (§5) |
+| CBO publication 41681 (17 April 2008) | pp. 1–3, Figures 1–2 | Synthesis | — | — | — | Context (§7.4 reason) |
+| CBO publication 49659 (7 November 2014) | — | — | — | — | — | Closed; methods open until a validation lane supplies a bounded definitions-only extract |
+| OACT presentation, Tiffany Bosley, 14 January 2025 | Slides 1, 3, 23 | — | — | — | — | Identity of the January 2025 update only (S0.1 §2.1) |
+
+**Bands and years in Study 129.** Each table prints "-" for the bands it
+does not report. The study takes only workers reaching 62 in 1993 or later
+(p. 6; notes under Tables 1–12), so a band `[a1, a2]` holds only such
+workers at every one of its ages from year `1993 + (a2 − 62)`, which this
+record calls the band's first fully supported year: 62–64 from 1995,
+65–69 from 2000, 70–74 from 2005, 75–79 from 2010, 80–84 from 2015 and
+85–89 from 2020. The printed "-" cells match that schedule exactly. Study
+124's Tables 4–8 print the same pattern through 80–84 in 2015, although
+that study takes birth cohorts from 1930 (p. 6).
+
+**Printing.** Tables 7–12 print RMRs to two decimals and the share row to
+one decimal; Tables 1–6 print rates to four decimals (counted from the
+text layer without reading values: 252 two-decimal and 252 four-decimal
+quintile-or-total cells, which equals 21 band-years × 2 sexes × 6 rows).
+
+**Figures as vector content.** Page 34's content stream holds 1,281
+line-to and 846 move-to operators and no raster plot image; pages 34–38
+list no image object at all, and page 33 lists only 14 small 16 × 16
+stencils and one 192 × 2493 ccitt image (`pdfimages -list`). A digitizer can
+therefore recover the annual series from path coordinates (§7.2).
+
+**Appendix lettering.** The body text places Tables 7–12 in Appendix E
+and Figures 7–12 in Appendix F (pp. 7–9), but the printed headings read
+"D. Age Group Death Rate Figures" (p. 21) and "E. Age Group Relative
+Mortality Ratios by AIME Quintile Levels" above Figure 7 (p. 33), and
+Tables 7–12 carry no letter (p. 27). This record cites tables and figures
+by number and page only.
+
+**Before a band first appears.** For a knot year earlier than a band's
+first fully supported year, the band's targets take their values from that
+first year (`BAND_BACKCAST_HOLD_FIRST`): 80–84 for 2010–2014 from 2015,
+and 85–89 for 2010–2019 from 2020. These copies are assumptions, not
+published values; they enter no critical-family count as new targets
+(§9.4), and §5 registers the alternative, T-B1. The fit records, and never fits, a
+figure point plotted in a year before its band's first fully supported
+year.
+
+### 3.2 Education
+
+| Data year | Report | Education table, PDF page | Population and SE table, PDF page | Coverage | Role |
+|---|---|---|---|---|---|
+| 2010 | NVSR 61(4) | Internet tables | Internet tables | 31 reporting states (titles also read "31 reporting states and the District of Columbia") | Sensitivity E-S1 only |
+| 2011 | NVSR 63(3) | Internet tables | Internet tables | 33 reporting states | E-S1 |
+| 2012 | NVSR 63(9) | Internet tables | Internet tables | 38 reporting states | E-S1 |
+| 2013 | NVSR 64(2) | Internet tables | Internet tables | 39 reporting states and DC | E-S1 |
+| 2014 | NVSR 65(4) | Internet tables | Internet tables | 43 reporting states (and a DC variant) | E-S1 |
+| 2015 | NVSR 66(6) | Internet tables | Internet tables | 46 reporting states | E-S1 |
+| 2016 | NVSR 67(5) | Internet tables | Internet tables | 49 reporting states | E-S1 |
+| 2017 | NVSR 68(9) | Internet tables | Internet tables | 49 reporting states and DC | E-S1 |
+| 2018 | NVSR 69(13) | Internet Table I–7, p. 15 | Internet Table I–19, p. 72 | United States | Fit |
+| 2019 | NVSR 70(8) | Internet Table I–7, p. 15 | Internet Table I–19, p. 71 | United States | Fit |
+| 2020 | NVSR 72(10) | Internet Table I–7, p. 20 | Internet Table I–19, p. 88 | United States | Fit |
+| 2021 | NVSR 73(8) | Table 14, p. 87 | Table VIII, p. 133 | United States | Fit |
+| 2022 | NVSR 74(4) | Table 14, p. 86 | Table VIII, p. 133 | United States | Fit |
+
+- **Layout, 2018–2022.** Number of deaths, death rate and age-adjusted
+  death rate for ages 25–34, 35–44, 45–54, 55–64 and 25–64, by sex (both,
+  male, female) and three education levels: less than high school diploma
+  or GED; high school diploma or GED; some college or collegiate degree;
+  plus "not stated" in the counts. Rates are per 100,000 population; the
+  population tables print each estimate with its standard error. This job
+  read the 2018, 2020 and 2022 layouts with values masked and the 2019 and
+  2021 titles only; the fit job's loader checks every year's layout. A
+  year whose layout does not match this one (bands, sexes, the three
+  education levels and the columns) counts as unobserved under §7.5,
+  labeled, as a missing file does under `SOURCE_UNAVAILABLE` (vii) (§10);
+  the same rule covers every life-table year in §3.3.
+- **Limits stated by the source.** No data below 25 (schooling may be
+  incomplete) and none at 65 and over ("reporting quality is poorer at
+  older ages"), with completeness the technical notes state (NVSR 74(4)
+  p. 127). The
+  denominators come from 1-year ACS PUMS adjusted to July 1 resident
+  control totals (p. 132), and the rates' standard errors combine Poisson
+  variation in deaths with the ACS standard error (Formulas 9 and 10,
+  p. 132). The 2020 denominators use the 2020 1-year ACS PUMS (NVSR 72(10)
+  internet tables, p. 2).
+- **No later edition.** The NVSR listing fetched on 1 November 2026 holds
+  no *Deaths: Final Data for 2023*; its latest final-deaths report is NVSR
+  74(4) for 2022. The last observed education year is therefore 2022.
+- **Why the fit leaves out 2010–2017.** Their populations are the
+  reporting states, not the United States, so their denominators and
+  composition differ from the reference population (§6). This job did not
+  check their category sets and age bands; sensitivity E-S1's loader checks
+  them before use, and a year whose categories or bands differ from the
+  2018–2022 layout leaves E-S1, labeled.
+- **Ages 65 and over and the `some_college`/`college` split.** NCHS
+  supplies neither (§4.2). NHIS–LMF supplies both (§3.4).
+
+### 3.3 Race and Hispanic origin
+
+| Data years | Reports | Groups with life tables (each total, male, female) | Table numbers verified in this job |
+|---|---|---|---|
+| 2010–2017 | NVSR 63(7), 64(11), 65(8), 66(3), 66(4), 67(7), 68(4), 68(7) | Total; white; black; Hispanic; non-Hispanic white; non-Hispanic black (bridged race) | 2010 and 2017: Hispanic Tables 10–12, NH white 13–15, NH black 16–18 |
+| 2018 | NVSR 69(12) | Total; Hispanic; NH white; NH black (single race) | Hispanic 4–6, NH white 7–9, NH black 10–12 |
+| 2019–2024 | NVSR 70(19), 71(1), 72(12), 74(2), 74(6), 75(5) | Total; Hispanic; NH AIAN; NH Asian; NH black; NH white (single race) | 2019, 2022, 2023, 2024: Hispanic 4–6, NH AIAN 7–9, NH Asian 10–12, NH black 13–15, NH white 16–18 |
+
+- **Layout.** Complete period life tables with single ages 0–99 and an
+  open "100 and older" interval; columns `qx`, `lx`, `dx`, `Lx`, `Tx`,
+  `ex`; `qx` printed to six decimals and `ex` to one (NVSR 75(5),
+  pp. 2, 16–49). In NVSR 74(2) and 75(5) the male and female tables start
+  on pp. 16 and 18 (total), 22 and 24 (Hispanic), 28 and 30 (AIAN), 34
+  and 36 (Asian), 40 and 42 (black) and 46 and 48 (white).
+- **Methods the source states.** Hispanic, NH white and NH black tables
+  begin with 2006 data; AIAN and Asian NH tables begin with 2019 data
+  (NVSR 75(5) p. 53). Single-race categories under the 1997 standards
+  begin with 2018, replacing bridged race (pp. 1–2, 54). Rates carry a
+  misclassification adjustment; Medicare data supplement ages 66–99 for
+  the total, NH black and NH white tables only, and NCHS models old-age
+  mortality for the Hispanic, AIAN and Asian NH tables (pp. 2, 54–55); the
+  2008-onward method smooths death rates above about 85, at ages that
+  depend on the group (p. 53). A Kannisto logistic model smooths the
+  Medicare-blended rates at 85–99 and predicts them at 100–120 (p. 57); a
+  Brass relational logit, with the white non-Hispanic age pattern as its
+  standard, predicts the older ages of the Hispanic, AIAN and Asian NH
+  tables (pp. 59–62). Every table then closes its "100 and older" row with
+  `∞q100 = 1.0`, `∞d100 = l100` and `∞L100 = T100`, the sum of the
+  extrapolated `Lx` for 100–120 (pp. 61–62).
+- **Fit.** Hispanic, NH white and NH black for every knot year 2010–2024
+  (bridged race through 2017, single race from 2018, both labeled). The
+  2006–2009 tables lie before the first knot, and the fit does not use
+  them.
+- **The `other_non_hispanic` group.** No NCHS table covers it. AIAN and
+  Asian NH tables leave out Native Hawaiian and Pacific Islander,
+  multiracial and other people, and exist only from 2019, so a weighted
+  average of them is not the full group (S0.1 §2.3). The fit takes this
+  group's contrast from NHIS–LMF (§3.4); the record reports the
+  AIAN-and-Asian composite for 2019–2024 beside it, labeled "partial
+  group".
+
+### 3.4 Public NHIS–LMF
+
+- **Release.** The 2019 public-use linkage: NHIS 1986–2018 interviews,
+  mortality follow-up through 31 December 2019 (description p. 1). The
+  NCHS FTP listing fetched on 1 November 2026 holds
+  `NHIS_<year>_MORT_2019_PUBLIC.dat` for 1986–2018, all dated
+  26 April 2022, and no later public linkage. The Linked Mortality Files
+  page (revised 7 July 2026) announces follow-up through 31 December 2022
+  for NHIS 1986–2021, but only as restricted-use files; they stay
+  excluded.
+- **Window [frozen here]: sample adults interviewed 2009–2018.** Reasons:
+  a single sample-adult universe with eligibility-adjusted weights
+  `SA_WGT_NEW` (codebook p. 4); identical `EDUC1`, `HISPAN_I` and
+  `RACERPI2` code lists in the 2009, 2016 and 2018 layouts; deaths
+  2009–2019, which overlap the fitted years of every other source. From
+  2015 only sample adults were linkage-eligible (description p. 2).
+- **Variables.** `EDUC1` (person file; 2009 layout p. 265, 2016 p. 296,
+  2018 p. 294); `HISPAN_I`, `RACERPI2`, `AGE_P`, `SEX` (sample adult file;
+  2009 pp. 7, 8, 10; 2016 pp. 8, 9, 11; 2018 pp. 8, 9, 11); design
+  variables `STRAT_P`/`PSU_P` (2009, p. 6) and `PSTRAT`/`PPSU` (2016 and
+  2018, p. 6); `MORTSTAT`, `ELIGSTAT` (codebook p. 1), `DODQTR`, `DODYEAR`
+  (p. 3), `SA_WGT_NEW` (p. 4); merge by `PUBLICID` (description p. 4).
+- **What the public files do not hold.** Ages above 84 at interview
+  (`AGE_P` is 85 for every older person), month and year of birth, and the
+  flags that show whether NCHS imputed race or ethnicity (2018 survey
+  description pp. 25–26); any lifetime-earnings history; deaths after
+  2019. Some records carry synthetic follow-up time or cause (description
+  p. 1).
+- **Roles.** Fit: education contrasts at attained ages 65–74, 75–84 and
+  85+, the `college`/`some_college` split at 25–64, and
+  `other_non_hispanic` contrasts. Beside: education contrasts at 25–64
+  and `hispanic` and `black_non_hispanic` contrasts, next to NCHS's.
+- **Release dates of the 2009–2018 NHIS files** stay **[open]** (row
+  S2-15); the fit job records each from that year's survey description,
+  or "not stated" where a description gives none.
+
+### 3.5 Reference population and baseline
+
+- **Reference population (reference, not a target) [frozen here].** The
+  Census Bureau's ACS 1-year PUMS person files for each knot year
+  2010–2024, with their person weights and replicate weights. NCHS built
+  its 2020 education denominators from the 2020 1-year PUMS (NVSR 72(10)
+  internet tables, p. 2), and the fit uses that file too. The Census
+  Bureau "did not release its standard 2020 ACS 1-year" PUMS because of
+  the pandemic and released the file "with experimental weights and
+  replicate weights" (its 2020 PUMS documentation page, §13.6); the fit
+  labels the 2020 knot's reference population accordingly and records any
+  further note. §6 defines the population built from these files. The
+  first freeze job opened no ACS file or documentation. Fix round 1 read
+  the ACS data dictionaries for 2010–2019 and 2021–2024, the subject
+  definitions for 2010–2024 and the 2024 top-coded-values document
+  (§13.6), and opened no ACS microdata; the data files' URLs stay
+  unchecked (§11.6).
+- **Baseline [frozen, S0.1 §1.1].** OACT's TR 2026 single-age, sex and
+  year `q` through the pinned accessor `data/tr2026.death_probability`,
+  historical through 2023 and Alternative II from 2024, ages 0–120 with
+  age 120 repeating age 119 (S0.1 §1.1 citing
+  `baselines/tr2026_intermediate.py:154`). `BaselineYearAwareMortality`
+  applies the year-`t` period `q` at the start-of-year age
+  `t − 1 − birth_year` (`baselines/base.py:19–28`).
+
+### 3.6 Roles by dimension
+
+| Dimension | Fitted sources | Reported beside, not fitted | Context |
+|---|---|---|---|
+| Lifetime earnings | Study 129 Tables 7–12 and Figures 7–12 | Study 124 Tables 4–8; Cristia Tables 3–6 | CBO 41681 |
+| Education | NCHS 2018–2022 (25–64); NHIS–LMF (65+, split) | NHIS–LMF at 25–64; Cristia Tables 3–6; NCHS 2010–2017 reporting-state tables | CBO 41681 |
+| Race and Hispanic origin | NCHS life tables 2010–2024 (Hispanic, NH white, NH black); NHIS–LMF (`other_non_hispanic`) | NHIS–LMF `hispanic` and `black_non_hispanic`; NCHS AIAN-and-Asian composite 2019–2024; Cristia Tables 3–6 | CBO 41681 |
+
+None of the sources reported beside is an unused holdout: this job read
+them, and SM2 claims no out-of-sample check (§10, `EXPOSURE_UNRESOLVED`).
+
+## 4. Mappings to SM1's `log_relative_risk`
+
+SM1's coefficient is a **conditional** log relative risk against the
+dimension's reference category, for one sex, single age 0–120 and path
+year, added across the three dimensions and then normalized within each
+age × sex × year cell (`mortality_ses/specification.py:199–254`,
+`normalizer.py:328–382`). Every source publishes something else: a
+**marginal** ratio between groups, or between a group and its whole
+population, for a band or a single age. This section maps each published
+quantity to a **marginal target** on a stated scale. Section 6 turns the
+marginal targets into conditional coefficients. Each mapping below is
+**[frozen here]**; each assumption carries its label into every report of
+the fit.
+
+**Common scale.** A target is `T = ln(H_g / H_ref)`, where `H_g` is
+category `g`'s average annual probability of death over the target's ages,
+weighted by the category's population at each single age, and `H_ref` is
+the same for the reference category. Probabilities are annual `q` at the
+start-of-year age that SM1 and the baseline use (`baselines/base.py:19–28`;
+`mortality_ses/model.py:171–184`). This crude form matches the OACT and
+NCHS education targets, which divide one category's crude rate within a
+band by another's (Study 129 p. 5; NVSR 74(4) Table 14).
+
+**NHIS targets: an age-standardized hazard ratio [frozen here].** NHIS's
+estimator holds attained age fixed within each band through single-year
+age terms (§4.5), so its hazard ratio is age-adjusted. A crude model-side
+ratio would load any difference between two categories' age mixes onto
+the category's coefficient. For an NHIS target the implied value is
+therefore
+
+`T_NHIS = ln( Σ_a W(a) μ̄_g(a) / Σ_a W(a) μ̄_ref(a) )`,
+
+summed over the band's single ages, where `μ̄_g(a)` averages
+`μ = −ln(1 − q)`, computed row by row from SM1's normalized `q`, over
+category `g`'s rows at age `a` (within the target's subpopulation, rows
+weighted by `w`), and `W(a)` is `R_ref(t)`'s total population of that sex
+at age `a`: one fixed set of weights for both categories. Above the ACS
+age top code, `W(a)` comes from §6.2's spread of top-coded records, so the
+85+ band runs on single ages 85–120. The target side needs no weights:
+under assumption M-H1 (§4.5) NHIS's ratio holds at every age of its band,
+so any fixed standard returns it. The sources read here give no reason to
+prefer another standard; NHIS's own person-time would make the implied
+value depend on survey records, while `R_ref(t)` is the population the
+coefficients act on.
+
+### 4.1 OACT relative mortality ratios (lifetime earnings)
+
+- **M-E1, the target.** For sex `s`, band `B`, year `t` and quintile
+  `k ∈ {q1, q2, q4, q5}`: `L_E(k) = ln RMR_k − ln RMR_q3`, from the printed
+  RMRs of Tables 7–12 or the digitized RMRs of Figures 7–12. Labels map
+  Lowest → `q1`, 2nd → `q2`, 3rd → `q3`, 4th → `q4`, Highest → `q5`.
+  Each RMR divides a quintile's rate by the rate of all retired workers in
+  that sex and band (Study 129 p. 5), so the "Total" denominator cancels
+  in `L_E` and the contrast equals the log ratio of the two quintiles'
+  rates. Its printing interval uses the transformed endpoints of both
+  RMRs (§9.1).
+- **M-E2, a rate ratio as a ratio of annual probabilities (assumption).**
+  Study 129 divides deaths by exposure years, and "Deaths receive a full
+  year of exposure at the age when the death occurs" (p. 5; examples in
+  Appendix B, pp. 13–14). That convention makes each published rate an
+  initial-exposure estimate of an annual probability of death, so this
+  record reads a ratio of two quintiles' rates as a ratio of their annual
+  `q`. It is still an assumption, not a source identity: the source does
+  not call its rates probabilities, and its exposure counts months by age
+  at last birthday.
+- **M-E3, dynamic rank versus frozen rank (assumption).** OACT sorts the
+  retired workers with exposure in each year by AIME within sex and
+  single year of age and cuts quintiles there (pp. 3–4), so the study ranks
+  survivors anew at every age; the study notes that people in high quintiles
+  at younger ages move into lower quintiles at older ages (p. 3). SM3's
+  attribute is a frozen partial-career rank (S0.1 §1.2). The fitted
+  coefficients are contrasts between dynamic-rank groups; applying them to
+  frozen-rank attributes is assumption M-E3. No source here identifies the
+  difference between the two contrasts.
+- **M-E4, beneficiaries versus the whole population (assumption).** OACT's
+  population is retired-worker beneficiaries with a PIA, without
+  WEP/totalization or older-formula records and without anyone previously
+  entitled to disability benefits (pp. 3, 11); the 62–64 group consists of
+  workers who claimed before their NRA and are likely less healthy than
+  others that age (p. 6). Applying their contrasts to everyone, including
+  people who never claimed, is assumption M-E4. In the fit, OACT targets
+  apply to the reference population's Social Security recipients aged
+  62–89 (§6.2), the nearest public stand-in; SM1 itself holds DI-origin
+  persons at the baseline `q` (`model.py:124–160`).
+- **M-E5, age basis.** OACT assigns exposure by attained age, month by
+  month (p. 5); SM1 applies the year-`t` `q` at the start-of-year age. The
+  mapping takes band `[a1, a2]` to SM1 ages `a1` to `a2` unchanged and
+  accepts the half-year offset; sensitivity T-E4 (§5) shifts every band
+  down one year.
+- **M-E6, band to single ages.** One coefficient per quintile × sex ×
+  band × knot, constant across the band's single ages; the reference
+  population supplies the within-band composition (§6).
+
+### 4.2 NCHS death rates by education
+
+- **M-D1, the target.** For sex `s`, band `B ∈ {25–34, 35–44, 45–54,
+  55–64}` and year `t`: `L_D(less_than_high_school) = ln q_LTHS − ln q_HS`
+  and `L_D(some_college_or_more) = ln q_SC+ − ln q_HS`, where each `q`
+  comes from the printed rate `m` (per 100,000, divided by 100,000) as
+  `q = m / (1 + m/2)`. That conversion assumes deaths spread evenly over
+  the year of age, the convention `data/life_table.py` uses for its
+  closing rate (lines 21–28). The 25–64 and age-adjusted columns are not
+  targets, because they repeat the band information.
+- **M-D2, the pooled category.** "Some college or collegiate degree"
+  pools SM's `some_college` and `college`. The model's pooled `q` over
+  both categories, weighted by the reference population, matches its
+  target, so NCHS identifies one combination of the two coefficients.
+- **M-D3, the split.** NHIS–LMF supplies
+  `ln(h_college / h_some_college)` at attained ages 25–64 by sex, pooled
+  over deaths 2009–2019 and held constant over time (§4.5).
+- **M-D4, ages 65 and over.** NHIS–LMF supplies marginal hazard ratios of
+  `less_than_high_school`, `some_college` and `college` against
+  `high_school` at attained ages 65–74, 75–84 and 85+, by sex, pooled and
+  held constant over time. An NHIS hazard `μ` maps to an annual probability
+  as `q = 1 − exp(−μ)` (constant force within the year).
+- **M-D5, degrees versus years (assumption).** NCHS and NHIS classify by
+  highest diploma or degree; SM3's PSID attribute counts years (<12, 12,
+  13–15, 16–17; S0.1 §1.2). A GED holder or a person with twelve years
+  and no diploma can fall in different SM categories under the two
+  definitions (S0.1 §2.3). This record maps degree-based sources by
+  degree (§8.2) and reports the difference with every result.
+- **Not stated.** NCHS computes category rates from stated education and
+  counts "not stated" deaths only in the totals (Table 14 footnote 3), so
+  the fit adjusts nothing for them and reports their share as a
+  data-quality figure.
+
+### 4.3 NCHS life tables by race and Hispanic origin
+
+- **M-R1, the target.** For sex `s`, single age `a` from 0 to 99, and
+  year `t`: `L_R(g, a) = ln q_g(a) − ln q_NHW(a)` for `g ∈ {Hispanic,
+  NH black}`, from the six-decimal `qx` of the tables in §3.3.
+- **M-R2, age basis.** NCHS's `qx` is the probability of dying between
+  exact ages `x` and `x + 1` (NVSR 75(5) p. 2), the baseline's convention;
+  the mapping is the identity.
+- **M-R3, ages 100–120.** The "100 and older" interval holds no single-age
+  `q`, so it supplies no contrast target (§5). The e65 targets alone read
+  it, for NCHS's closing of the open interval: `e100 = T100 / l100` from
+  the printed `Tx` and `lx` of that row (§9.2).
+- **M-R4, `other_non_hispanic`.** NHIS–LMF supplies the marginal hazard
+  ratio against `white_non_hispanic` at attained ages 25–44, 45–64,
+  65–74, 75–84 and 85+, by sex, pooled and held constant over time (fit).
+  Beside it, for 2019–2024, the record reports `ln` of the ACS-weighted
+  average of the AIAN and Asian NH `q` over the NH white `q`, labeled
+  "partial group: no NHPI, multiracial or other people".
+- **M-R5, bridged and single race (assumption).** 2010–2017 contrasts are
+  bridged-race; 2018 onward are single-race; Hispanic origin does not
+  change (NVSR 75(5) pp. 1–2, 54). SM's categories follow the single-race
+  definitions (§8.2), and every path labels the 2017–2018 break. **A
+  possible 2023–2024 change (label).** In NVSR 75(5), the 2024 tables,
+  the notes under Tables 4 (p. 21), 5 (p. 23), 7 (p. 27), 8 (p. 29) and
+  9 (p. 31), that is Hispanic total and male and AIAN total, male and
+  female, read "Updated classification ratios were applied"; the note
+  does not appear under Table 6 (Hispanic female, p. 25), and NVSR 74(2)
+  and 74(6) carry no such note. The 2024 Hispanic tables may therefore
+  rest on a revised misclassification adjustment. P-PRIMARY holds the
+  2024 level after the last knot (Q6, §7.1), so every path labels a
+  possible 2023–2024 break for `hispanic`, and the AIAN-and-Asian
+  composite beside the fit carries the same label for 2024. The fit
+  reports the 2023–2024 change in the Hispanic contrasts beside the path
+  and adjusts nothing for it.
+- **M-R6, ratios to the total population.** For the e65 target only
+  (§9.2), each group's `q` divides by the NCHS total-population `q` at the
+  same age and year, and each group's open-interval `e100` divides by the
+  total's.
+- **Properties the source states and every report repeats:** the
+  misclassification adjustment, the Medicare supplement at 66–99 for
+  three groups only, modeled old-age mortality for the other groups, and
+  smoothing above about 85 (§3.3).
+
+### 4.4 Cristia's odds ratios (sensitivity and beside only)
+
+**Log odds ratios never enter as log relative risks.** This section cites
+Cristia's prose by printed page; the PDF page runs two higher (S0.1), and
+the tables carry PDF pages (§3.1). Cristia's models predict death in the
+next year with the bottom quintile as reference (printed p. 18). Tables 5–6 hold, for each sex and age range, a column with age
+and birth year only (columns 1, 3, 5) and a column that adds race,
+marital status and education (columns 2, 4, 6); the third specification
+in Figure 2 (race and marital status) appears only as a plotted line.
+
+- **M-C1, the reference population [frozen here].** For sex `s` and
+  Cristia age range `b`, the reference population is the single ages in
+  `b`, weighted by the stationary population `Lx` of TR 2026's
+  historical period `q` for 1993 (the midpoint of Cristia's 1983–2003
+  observation years; if the pinned historical file starts later, its
+  earliest year, reported), with equal quintile shares. The ranges are 35–49,
+  50–64 and 65–75. Tables 5–6 print the first range as "30-49", but the
+  sample keeps ages 35–75 only (printed p. 8); this record reads it as 35–49 and
+  reports the misprint.
+- **M-C2, marginal earnings contrasts (columns 1, 3, 5).** At each age
+  `a` in `b`, solve the intercept `α(a)` so that
+  `0.2 × Σ_k expit(α(a) + ln OR_k) = q_base(s, a, 1993)`, with
+  `OR_q1 = 1`. Then `p_k(a) = expit(α(a) + ln OR_k)`, and the band contrast
+  against `q3` is `ln` of the `Lx`-weighted mean of `p_k` over that of
+  `p_q3`. Quintiles map Bottom → `q1` through Top → `q5`.
+- **M-C3, conditional contrasts (columns 2, 4, 6).** At a named reference
+  profile (white, married, high school graduate, `q3`), solve `α(a)` so
+  that the profile's probability equals `q_base(s, a, 1993)`; compute each
+  category's probability at the profile with only that category changed;
+  take contrasts against SM's reference categories (`q3`, `high_school`,
+  `white_non_hispanic`). Cristia's education reference is less than high
+  school, so the contrast against high school is the difference of two
+  predicted log probabilities.
+- **M-C4, category mapping (assumption).** Cristia's "White", "Black" and
+  "Other Race" map to `white_non_hispanic`, `black_non_hispanic` and
+  `other_non_hispanic`, and "Hispanic" to `hispanic`; Tables 3–4 list the
+  four as one "Race and Ethnicity" block, and this job did not confirm
+  that they are mutually exclusive.
+- **M-C5, marginal group ratios (Tables 3–4).** Directly age-standardized
+  one-year mortality ratios against all same-sex persons in the band
+  (printed p. 14). Contrasts against a reference category are differences of
+  their logs; they are rate ratios already and need no conversion.
+- **M-C6, uncertainty.** Tables 5–6 print robust z statistics in
+  parentheses, not standard errors (table notes), so
+  `SE(ln OR) = |ln OR| / |z|`; a contrast whose printed z or odds ratio
+  leaves this undefined carries no standard error. Tables 3–4 print no
+  uncertainty. Marital status stays in columns 2, 4 and 6, which adjust
+  for a dimension SM1 does not model.
+
+### 4.5 The NHIS–LMF estimator
+
+**[frozen here]** for the fit job; this job opened no NHIS record.
+
+- **Persons.** Sample adults interviewed 2009–2018 with `ELIGSTAT = 1`,
+  weighted by `SA_WGT_NEW` divided by 10 (ten pooled years).
+- **Time at risk.** From the middle of the interview quarter
+  (`SRVY_YR`, `INTV_QRT`) to the middle of the death quarter (`DODYEAR`,
+  `DODQTR`) for `MORTSTAT = 1`, or to 31 December 2019.
+- **Attained age.** `AGE_P` plus elapsed time. Because NHIS suppresses
+  birth dates, the estimator splits each person-year across attained ages
+  on the assumption that birthdays fall uniformly through the year. A person
+  with `AGE_P = 85` (85 or older at interview) contributes only to the
+  85+ band, the one band whose lower bound such a person surely meets.
+- **Model.** For each sex, band and dimension, a survey-weighted Poisson
+  regression of deaths on that dimension's category indicators and
+  single-year attained-age indicators, with log person-years as offset.
+  In the 85+ band the age terms are indicators of attained ages 85–89,
+  90–94 and 95+ for persons interviewed below 85, and one indicator per
+  completed year since interview for persons top-coded at interview. The
+  exponentiated category coefficients are the marginal hazard ratios.
+- **M-H1, proportional hazards within a band (assumption).** The model
+  gives each category one hazard ratio per band, so it assumes that ratio
+  holds at every attained age in the band. In the 85+ band the assumption
+  covers ages 85–120, including the ages that NHIS sees only through its
+  95+ indicator and through persons top-coded at interview. §4's common
+  scale compares this ratio with an age-standardized model ratio, and
+  §11.7 tests that the two agree whenever the model's hazards are
+  proportional.
+- **Missing values.** `EDUC1` 97–99, and any race code that §8.2 leaves
+  unclassified (none in the primary; `HISPAN_I` 10 and 11 under
+  sensitivity X-R1), drop out of that dimension's estimate only, and the
+  fit reports their counts.
+- **Variance.** §9.3.
+
+## 5. Age support and tails
+
+SM1 needs a coefficient for every category, sex and single age 0–120 in
+every path year (`specification.py:171–197`). No source covers all of
+them. **A source gap never becomes an empirical zero effect**: every age
+outside a source's support takes a named extrapolation, fixed below
+**[frozen here]** before any fit, and each sensitivity is a registered
+alternative reported beside the primary, never a replacement for it.
+Extrapolations act on the fitted conditional coefficients (§6), not on
+the marginal targets.
+
+| Dimension | Ages the fit's sources support | Primary rule below | Primary rule above | Registered sensitivities |
+|---|---|---|---|---|
+| Lifetime earnings | 62–89 (bands 62–64 to 85–89, with backcasts before 2015 and 2020, §3.1) | `TAIL_E_HOLD_62_64_DOWN`: ages 25–61 take the 62–64 coefficient; ages 0–24 take the age-25 value | `TAIL_E_HOLD_85_89_UP`: ages 90–120 take the 85–89 coefficient | T-E1: ages 25–61 take the 65–69 coefficient instead, which avoids the early-claimer selection the source describes for 62–64 (Study 129 p. 6). T-E2: Cristia age shape (below). T-E3: a linear taper of each log contrast from its 85–89 value at age 89 to zero at age 110, and zero above, labeled "assumed convergence, not a finding" (the source reports narrower spreads at older ages, pp. 3, 8, but no rate). T-E4: every OACT band shifted one year younger (M-E5). T-B1: in place of `BAND_BACKCAST_HOLD_FIRST`, a band takes, at each knot before its first fully supported year, the targets of the next younger band fully supported at that knot: 75–79's for 80–84 and for 85–89 at 2010–2014, and 80–84's for 85–89 at 2015–2019, labeled "band backcast from a younger band" |
+| Education | 25–64 (NCHS bands); 65–74, 75–84 and 85+ (NHIS–LMF) | `TAIL_D_HOLD_25_DOWN`: ages 0–24 take the 25–34 coefficient | The 85+ band runs to 120 | T-D1: ages 65–120 take the 55–64 coefficients, with no NHIS input, to show how much the 65+ coefficients rest on NHIS |
+| Race and Hispanic origin: `hispanic`, `black_non_hispanic` | Single ages 0–99 | None needed | `TAIL_R_HOLD_99_UP`: ages 100–120 take the age-99 coefficient; the fit reads no contrast from the "100 and older" interval | T-R1: a taper from the age-99 value to zero at age 110, labeled as T-E3 |
+| `other_non_hispanic` | 25–44, 45–64, 65–74, 75–84, 85+ (NHIS–LMF) | `TAIL_O_HOLD_25_DOWN`: ages 0–24 take the 25–44 coefficient | 85+ runs to 120 | T-R2: for 2019–2024, the partial AIAN-and-Asian composite (M-R4) replaces NHIS at every age, labeled "partial group" |
+
+- **T-E2, Cristia's age shape.** For each sex and non-reference quintile
+  `k`, let `c_b(k)` be Cristia's marginal log contrast against `q3` in age
+  range `b` (M-C2). Ages 50–61 take the 62–64 coefficient times
+  `c_50–64(k) / c_65–75(k)`, and ages 25–49 times
+  `c_35–49(k) / c_65–75(k)`. For range `b`, the unweighted least-squares
+  slope through the origin of `c_b(k)` on `c_65–75(k)` over the sex's
+  four non-reference quintiles, `Σ_k c_b(k) c_65–75(k) / Σ_k c_65–75(k)²`,
+  replaces a quintile's ratio when `|z| < 2` for that quintile's 65–75
+  odds ratio, and every ratio stays within `[0.5, 3]`; this record
+  fixes both guards to keep a ratio of small numbers from driving the
+  sensitivity.
+- **Ages under 25.** NCHS leaves out ages below 25 because people that
+  young may not have finished their education (NVSR 74(4) p. 127); SM3's
+  earnings proxy counts income from age 22 (S0.1 §1.2); and SM3's
+  population starts at 50.
+  The held values below 25 exist because SM1's table must be complete; a
+  frame that classifies a person under 25 on either dimension must
+  register that use first (§8.3). `R_ref` classifies no one under 25 on
+  either dimension: it holds them at the reference categories (§6.2), so
+  these held values reach no one in the fit, and SM2 models no earnings or
+  education gradient below 25.
+- **The baseline's top.** Ages 100–120 multiply a baseline `q` that is
+  already large, and SM1 caps each probability at one and conserves the
+  cell's deaths by recomputing the factor over the uncapped rows
+  (`normalizer.py:338–356`). SM1 logs capped cells, and §9.1 checks the
+  logs.
+
+## 6. Joint route
+
+### 6.1 The identification problem
+
+The sources publish marginal contrasts: OACT's earnings ratios do not
+hold education or race fixed; NCHS's education rates do not hold earnings
+or race fixed; NCHS's life tables do not hold earnings or education
+fixed. SM1 adds conditional log relative risks. When two dimensions vary
+together within a cell, the model's implied marginal contrast for one of
+them equals its own coefficient **plus** a composition term from the
+other dimensions' coefficients, weighted by how the other dimensions'
+categories spread across its categories. Adding marginal contrasts as
+coefficients therefore reproduces the marginal targets only in a
+population where the three dimensions are independent within every cell.
+Correlation can make the sum overshoot or undershoot; which one depends
+on the joint distribution, and no one can tell in advance (S0.1 §2.3).
+Turning marginal targets into conditional coefficients needs a
+**reference joint distribution** of the three attributes by sex and single
+age, and an assumption that the effects add on the log scale.
+
+### 6.2 The reference population `R_ref(t)` [frozen here]
+
+For each knot year `t` (2010–2024), `R_ref(t)` is the weighted ACS 1-year
+PUMS population of that year: every person record the file holds,
+including group-quarters records where the file carries them (the fit job
+records which it does), with these attributes:
+
+- **Sex and single age.** `AGEP` gives single ages 0–99, top-coded by
+  state: every record at or above its state's threshold carries the state
+  mean of those ages (2024 top-coded-values document, "Explanation of
+  Topcoding"; each year's thresholds and means sit in that year's
+  top-and-bottom-coded-values CSV, which the fit job records; a year
+  whose CSV the fit job cannot obtain counts as a missing ACS year under
+  `SOURCE_UNAVAILABLE` (iii), §10). A record at
+  or above its state's threshold stands for every age from the threshold
+  to 120: the fit spreads its weight over those single ages in proportion
+  to the stationary person-years `Lx` of the TR 2026 `q` for that sex and
+  year, and each spread row keeps the record's other attributes. Below the
+  threshold, `AGEP` is the single age.
+- **Education and race and Hispanic origin.** From the ACS education,
+  Hispanic-origin and race items through the crosswalk in §8.2; education
+  at ages 25 and over only.
+- **Persons under 25 [frozen here].** `R_ref` assigns no lifetime-earnings
+  or education category below age 25, the record's own limit (§5, §8.3).
+  Every person under 25 holds `q3` and `high_school`, the reference
+  categories, whose log relative risk is zero, and the diagnostic column
+  records `reference_below_25` on both dimensions (§8.1): a label, not a
+  classification. Race and Hispanic origin classify at every age.
+- **A lifetime-earnings quintile proxy `Q̃`.** At ages 62–89, among
+  persons with positive Social Security income, `Q̃` is the weighted
+  quintile of that income within sex and single age; the construction splits
+  weight that straddles a quintile boundary so each quintile holds exactly
+  one fifth.
+  This is the public stand-in for OACT's AIME quintile (M-E3, M-E4): both
+  rank people within sex and single age. OACT ranks by AIME, which orders
+  people as their PIA does within a birth year (Study 129 p. 2, note 2);
+  the ACS ranks by the Social Security income each person reports, which
+  departs from the person's own PIA in the ways the last bullet of this
+  list sets out.
+- **The dependence model.** For each knot year `t` and sex `s`, on that
+  year's persons with positive Social Security income aged 62–84, for
+  each age group `G ∈ {62–69, 70–79, 80–84}`, a weighted ordered logit
+  `P(Q̃ ≤ k | d, r, a) = expit(c_k(a) − λ_D,G[d] − λ_R,G[r])`,
+  with `λ_D[high_school] = λ_R[white_non_hispanic] = 0` and cut points
+  by single age, every parameter indexed by `(s, t)`. Ages 85 and over
+  use `G = 80–84`, and ages 25–61 use `G = 62–69`; persons under 25 hold
+  the reference categories (above). §6.6 names this copying as assumption
+  M-J1.
+  - **One fit per knot year [frozen here].** `R_ref(t)`, `Q̃`, the income
+    amounts, the cut points and the replicate weights all belong to one
+    ACS year, and §6.3 calibrates one knot at a time. A fit per year keeps
+    each knot's inputs within one file, and a replicate refit then
+    re-estimates exactly what the point fit estimates. `λ` enters the
+    calibration as an input, not an unknown, so §6.3's system stays
+    square.
+  - **Weights and ties [frozen here].** The fit weights each record by
+    `PWGTP` and ranks by `SSP` as the file holds it: each 1-year file
+    carries a single `ADJINC` factor (data dictionaries for 2010–2019 and
+    2021–2024), so a year's ranking needs no adjustment. Records with
+    equal `SSP` at the same sex and single age form one block. The
+    construction splits the block's cumulative-weight interval at each
+    quintile boundary it spans, and every record in the block takes the
+    same fractional membership of each quintile, in proportion to the
+    overlap; record order inside a block changes nothing. A record enters
+    the likelihood once for each quintile it belongs to, weighted by
+    `PWGTP` times that fraction. Every dictionary read here marks `SSP`
+    as rounded, and the 2024 top-coded-values document lists it among the
+    variables top-coded at a state mean above a state threshold, so blocks
+    occur.
+  - **Optimizer [frozen here].** Newton–Raphson on the weighted
+    log-likelihood with analytic gradient and Hessian; each age's cut
+    points enter as `c_1(a)` and the logs of the gaps
+    `c_k+1(a) − c_k(a)`, so they stay ordered. The start is `λ = 0`
+    with cut points `logit(0.2)`, `logit(0.4)`, `logit(0.6)` and
+    `logit(0.8)` at every age. Each step halves until the log-likelihood
+    does not fall, at most 30 times, and the fit takes at most 100 steps.
+    It has converged when every parameter moves by less than
+    `1 × 10⁻⁸` and the largest gradient component, divided by the total
+    weight, falls below `1 × 10⁻¹⁰`. Anything else is
+    `SOURCE_UNAVAILABLE` (iv) (§10).
+  - **Replicates [frozen here].** Inside each ACS replicate refit (§9.3),
+    the replicate weight replaces `PWGTP` in every step above: `Q̃` and
+    its blocks, the ordered logit from the same start under the same
+    rules, the cut points and `P(d, r | s, a, t)`. A replicate whose
+    ordered logit fails stops the procedure, as the point fit's failure
+    does.
+- **The joint weights.** At sex `s`, age `a` and year `t`,
+  `w(k, d, r) = P(d, r | s, a, t) × P(k | d, r)`: the first factor is the
+  ACS weighted share among all persons; the second comes from year `t`'s
+  ordered logit for sex `s`, with cut points `c_k(s, a, t)` recomputed so
+  that, at ages 62–89, every quintile holds exactly one fifth of the
+  subpopulation with positive Social Security income (OACT cuts its
+  quintiles at fifths of the records at each age, though their exposure
+  shares can differ slightly, pp. 3–5), and at ages 25–61 and 90–120 one
+  fifth of the whole population; below 25 every person holds `q3`. Each
+  recomputed cut point solves its share equation by bisection until the
+  share lies within `1 × 10⁻¹²` of its target. Persons without Social Security income at 62–89 take
+  `P(k | d, r)` from the same model. The fit
+  evaluates OACT targets on the subpopulation with positive Social
+  Security income (its own `P(d, r | s, a, t)`, the same `P(k | d, r)`);
+  every other target on the whole population.
+- **What the proxy is not, and where it is weakest.** A benefit amount is
+  not an AIME, and Study 129 itself shows how they part. The amount
+  depends on when the worker claims: the NRA is the age of first
+  entitlement "without an age-based reduction" (p. 6), and "workers who
+  can wait to retire may do so to receive a higher monthly Social Security
+  benefit" (p. 9). A person "may be entitled as a retired worker on their
+  own record and as a spouse on another record", and where the spouse
+  benefit is the higher, "The total benefit is the same amount as the
+  full spouse benefit" (p. 8, note 1), so the total follows the other
+  record, not the beneficiary's own AIME. Women
+  "often became dually entitled", so that their AIME "may not have closely
+  aligned with their family income level" (p. 8). The ACS item also
+  counts more than a retired-worker benefit: in every year 2010–2024 its
+  definition includes survivor benefits, permanent disability insurance
+  payments made by the Social Security Administration, and railroad
+  retirement (next bullet), none of which OACT's AIME ranking sees (M-E4);
+  Study 129 does not discuss survivors. **The proxy is therefore weakest
+  for women:** for a dually entitled woman, `Q̃` ranks an amount set on
+  her spouse's record, not her own AIME. JR-C-W (§6.5) shows how much
+  women's coefficients rest on it, and §6.6 lists what the gap leaves
+  unidentified.
+- **The income item's acceptance rule [frozen here].** This round read
+  the item in the subject definitions for every year 2010–2024
+  ("Income in the Past 12 Months", item 4; 2010 p. 79, 2024 p. 95) and in
+  the data dictionaries for 2010–2019 and 2021–2024 (§13.6). The
+  dictionaries define `SSP` as "Social Security income past 12 months"
+  for persons 15 and over. The subject definitions hold two wordings, one
+  for 2010–2016 ("Social Security income includes Social Security
+  pensions and survivor benefits, permanent disability insurance payments
+  made by the Social Security Administration prior to deductions for
+  medical insurance, and railroad retirement insurance checks from the
+  U.S. government. Medicare reimbursements are not included.") and one for
+  2017–2024, headed "Social Security or Railroad Retirement income", which
+  reads "Social Security and U.S. railroad retirement pensions and
+  survivor benefits" in place of "Social Security pensions and survivor
+  benefits" and is otherwise the same. Both describe the same benefit
+  types, which this record takes as the proxy's stated content, not as a
+  failure. For each year it uses, the fit job checks the year's subject
+  definition and `SSP` dictionary entry: a file whose SHA-256 equals the
+  one §13.6 lists passes, and any other copy passes when the item's text
+  matches the frozen wording for its years, line breaks and spacing aside.
+  A year that fails, or whose documents the fit job cannot obtain (2020's
+  dictionary is one, §8.2), takes `λ` from the nearest passing year, the
+  earlier year on a tie, keeps its own file for everything else, and
+  carries the label "income-item definition borrowed from" that year. If
+  no year passes, `SOURCE_UNAVAILABLE` (vi) applies (§10).
+
+### 6.3 Joint marginal calibration [frozen here]
+
+For each knot `t` and sex `s`, the unknowns are:
+
+| Block | Unknowns | Targets |
+|---|---|---|
+| Lifetime earnings | 4 non-reference quintiles × 6 bands = 24 | 24 OACT contrasts (M-E1), including backcast copies (§3.1) |
+| Education, 25–64 | `less_than_high_school` and `some_college` per NCHS band (8), plus one split `δ_s`, with `college = some_college + δ_s` in every band | 4 bands × 2 NCHS contrasts (M-D1, M-D2) plus the NHIS split (M-D3) = 9 |
+| Education, 65+ | 3 non-reference categories × 3 bands = 9 | 9 NHIS contrasts (M-D4) |
+| `hispanic`, `black_non_hispanic` | 2 × 100 single ages = 200 | 200 NCHS contrasts (M-R1) |
+| `other_non_hispanic` | 5 bands | 5 NHIS contrasts (M-R4) |
+
+That is 247 equations in 247 unknowns per knot and sex. Coefficients at
+ages outside these blocks follow §5. Below 25, `R_ref` holds everyone at
+`q3` and `high_school` (§6.2), so its rows there vary by race and
+Hispanic origin only, and the race coefficients at ages 0–24 carry no
+earnings or education composition. At ages 25–61 they and the education
+coefficients come out net of an earnings composition that rests on the
+copied dependence (M-J1, §6.6). The fit computes each target's implied
+value by expanding `R_ref(t)` into
+rows, one per (quintile, education,
+race) combination at each single age, with weight `w`, applying the
+candidate coefficients and SM1's `normalize_probabilities` cell by cell
+against the baseline `q`, and averaging the normalized probabilities over
+the target's ages and subpopulation: crude for OACT and NCHS targets, and
+age-standardized hazards for NHIS targets (§4, common scale).
+
+- **Solver.** Newton's method on the full system, started from the
+  plain-addition coefficients (JR-A); convergence when every implied
+  target is within `1 × 10⁻¹⁰` of its target in log units. Each step
+  starts from the full Newton step and halves it, up to 30 times, until
+  the sum of squared log residuals falls; a start that would need a 31st
+  halving, or that has not converged after 100 steps, fails to converge.
+  A second start from all-zero coefficients must reach the same solution
+  within `1 × 10⁻⁹`.
+- **Failure.** If either start fails to converge, or the two disagree,
+  the knot's calibration fails. That is a gate failure, not a tuning
+  opportunity (§10, `GATE_FAIL_OR_REGRESSION_FAIL`), and whether it
+  happens is an outcome this record does not predict (row S2-17).
+
+### 6.4 Candidate routes
+
+| Route | What it identifies | Reference distribution it needs | Where it comes from without PSID | What the public NHIS–LMF can and cannot supply | How it can go wrong |
+|---|---|---|---|---|---|
+| **JR-A: plain addition of marginals** | Each dimension's marginal contrast in its own source population; nothing joint | None; it assumes the three attributes are independent within every cell | Not needed | Not needed. It could check the education × race association, which this route ignores | Double counting wherever the dimensions vary together in the same direction of risk, and the reverse where they vary in opposite directions; the size depends on the deployment population and goes unmeasured |
+| **JR-B: education × race joint, earnings independent** | Education and race coefficients adjusted for each other; the earnings coefficient equals OACT's marginal contrast | `P(d, r \| s, a)` | ACS PUMS (§6.2), with `λ = 0` | Can supply the education × race association in its own population as a check; cannot supply anything about earnings | Counts the earnings–education and earnings–race overlap twice |
+| **JR-C: full joint calibration with a public earnings proxy (primary)** | Conditional coefficients that reproduce every marginal target on one named population, given the proxy's dependence and log-additivity | `P(k, d, r \| s, a)` | ACS PUMS with the Social Security income quintile and the ordered logit (§6.2) | Supplies the education targets at 65+, the split and `other_non_hispanic`; cannot supply the earnings dependence | The proxy's dependence differs from the AIME quintile's (§6.6); effects interact rather than add; the dependence in the deployment population (PSID cohorts) differs from the ACS's |
+| **JR-D: Cristia-bridged shrinkage** | The 1983–2003 ratio of mutually adjusted to marginal contrasts in SIPP–SSA data at 35–75, carried to modern marginals | None named; it is implicit in Cristia's sample | Cristia's Tables 3–6 | Nothing | The ratio may not carry across two to four decades and a different population; columns 2, 4, 6 also adjust for marital status; rounded z statistics; it reproduces no modern marginal on any named population, so no e65 check applies |
+| **JR-E: NHIS direct joint model** | Education and race hazard ratios adjusted for each other in NHIS's 2009–2019 population; earnings as JR-B | The NHIS sample itself | NHIS–LMF | Supplies everything for education × race at attained ages 25+; cannot supply earnings, single ages above 84 at interview, or deaths after 2019 | Does not reproduce NCHS's targets (different population, years and reporting); survey precision; the earnings overlap as JR-B |
+| **JR-F: PSID reference population (not used, §12)** | As JR-C, with the deployment population's own joint distribution | The PSID cohort's lifetime-earnings × education × race distribution | — | — | PSID use follows the project's data-access record. This record does not use it; §12 states the rule for the primary |
+
+**JR-D's definition [frozen here].** For dimension `X`, sex `s` and
+Cristia age range `b`, `κ_X(s, b)` is the weighted least-squares slope,
+through the origin, of Cristia's conditional log contrasts (M-C3) on its
+marginal log contrasts (M-C2 for earnings, M-C5 for education and race),
+across the dimension's non-reference categories, with weights equal to
+the categories' shares in Cristia's sample: for race and ethnicity, Table
+1's "Weighted" column (PDF p. 36), which pools sexes and ages; for
+education, the mean of Table 2's two Mortality Sample columns, 1984 and
+1996 (PDF p. 37); for earnings quintiles, equal weights, since each
+quintile holds a fifth by definition. The same weights serve every sex
+and age range, because neither table splits by them. `κ` stays within
+`[0, 1]`. JR-D's coefficients
+are `κ` times JR-A's, and ages outside 35–75 take the nearest range's `κ`.
+
+### 6.5 Choice
+
+**Primary: JR-C.** Sensitivities: JR-A (labeled "plain addition of
+marginals"), JR-B (JR-C with `λ = 0`), JR-C with `λ` scaled by 0.5 and by
+1.5, JR-C-IND and JR-C-W (below), JR-D and JR-E. Reasons:
+
+1. **Q5 asks for one model.** "Education and race and Hispanic origin are
+   further dimensions of the same model" calls for coefficients that hold
+   the other dimensions fixed. JR-A and JR-B set some of that adjustment
+   to zero by construction, and JR-D borrows it from another era.
+2. **It makes S0.1's e65 conformance computable.** S0.1 requires a fixed
+   reference population and marginalization for the e65 target (§2.5).
+   Only routes that calibrate on a named population can reproduce a
+   marginal target there; JR-C reproduces all of them on one population.
+3. **Its inputs are public and current.** The ACS covers every knot year
+   and is the same survey NCHS uses for its education denominators
+   (NVSR 74(4) p. 132).
+4. **It measures what the others assume.** JR-A and JR-B fix the
+   earnings–education/race dependence at zero; JR-C estimates it with a
+   stated proxy, and the `λ` sensitivities and JR-D bracket it.
+5. **Q5's primacy of earnings stays visible.** Every route keeps OACT's
+   contrasts as the earnings targets; no route lets an education or race
+   source move them.
+
+**JR-C-IND, independence wherever the fit copies the dependence [frozen here].**
+Wherever M-J1 copies the dependence (§6.6), JR-C-IND sets
+`P(k | d, r) = 1/5` for every education and race category, and keeps the
+estimated dependence for recipients aged 62–84 only. Its label is
+"dependence outside recipients aged 62–84 set to independence". It shows
+how much the coefficients at ages 25–61 and 85 and over, and the targets
+evaluated on the whole population, rest on the copy.
+
+**JR-C-W, women's dependence from men's [frozen here].** The proxy is
+weakest for women (§6.2). In JR-C-W, women's ordered logit takes men's
+`λ_D,G` and `λ_R,G` of the same knot year and age group, and women's cut
+points follow §6.2's recomputation; men's tables equal JR-C's. Its label
+is "women's dependence borrowed from men". The calibration runs
+separately for each sex (§6.3), so JR-B's women's tables already show
+`λ = 0` for women alone, and JR-C-W adds the other alternative: the
+dependence men show.
+
+### 6.6 What stays unidentified
+
+- **The dependence between OACT's AIME quintile and education or race.**
+  The Social Security income quintile stands in for it. Benefits that do
+  not follow the person's own AIME move the proxy away from OACT's
+  quintile: claiming-age reductions (Study 129 pp. 6, 9), the full spouse
+  benefit that a dually entitled beneficiary receives in total (p. 8,
+  note 1), and the survivor, disability and railroad retirement benefits
+  that the ACS item counts and OACT's AIME ranking never sees (§6.2,
+  M-E4). **The gap is widest for women**, whose AIME the study says "may
+  not have closely aligned with their family income level" (p. 8): for a
+  dually entitled woman, the proxy ranks an amount set on the spouse's
+  record. No public source settles which way these move the estimated
+  dependence; JR-C-W (§6.5) shows how much women's coefficients rest on
+  it.
+- **M-J1, the copied dependence (assumption).** The ordered logit sees
+  only recipients aged 62–84 (§6.2). The fit applies that dependence to
+  persons it never observed: non-recipients at 62–89, everyone at 25–61,
+  recipients at 85–89 (through `G = 80–84`) and everyone at 90–120. No
+  public source shows how lifetime earnings depend on education and race
+  in those groups. The copy shapes other coefficients: at ages 25–61 the
+  education and race coefficients come out net of an earnings composition
+  nobody observes (§6.3), and the targets evaluated on the whole
+  population carry it. JR-C-IND (§6.5) replaces the copy with
+  independence. Below 25 the fit copies nothing: `R_ref` holds everyone at the
+  reference categories, so SM2 models no earnings or education gradient
+  there, and the race coefficients at 0–24 carry no earnings or
+  education composition.
+- **Interactions.** The model assumes the three effects add on the log
+  scale. No public source has all three jointly with modern deaths. The
+  fit job reports NHIS's education × race interaction terms at 65+ as a
+  diagnostic; it does not gate on them.
+- **The dynamic-versus-frozen rank difference** (M-E3), and the
+  degree-versus-years education difference (M-D5).
+- **The earnings gradient below 62 and above 89, and every tail in §5.**
+- **The time paths of the NHIS-based coefficients.** NHIS's pooled
+  2009–2019 contrasts hold constant across the knots.
+- **DI origin.** The fit treats every reference-population member as
+  free; SM1 holds DI-origin persons at baseline in deployment, and S0.1
+  frozen text notes that this agrees with OACT's exclusion "in direction"
+  without proving equal definitions (S0.1 §2.3).
+
+## 7. Paths over time
+
+### 7.1 Primary path P-PRIMARY (Q6) [frozen here]
+
+- **Knots.** Every calendar year 2010–2024, the same fifteen years for all
+  three dimensions. At each knot `t`, each dimension's targets are its
+  latest observed values at or before `t`; before a dimension's first
+  observation it takes that first observation (hold-first backcast); the
+  three then calibrate jointly on `R_ref(t)` (§6.3). The result is one
+  complete coefficient table per knot (§7.5).
+- **After the last knot.** SM1's `HoldLastCoefficientPath` holds the 2024
+  table for every later year and refuses years before 2010
+  (`specification.py:76–106`). That holding is Q6's primary path: the
+  gradient stays at its last observed level.
+- **Last observed year per dimension.** Lifetime earnings 2022 (Study 129
+  Figures 7–12; the last table year is 2020); education 2022 (NVSR 74(4));
+  race and Hispanic origin 2024 (NVSR 75(5)); the NHIS-based targets pool
+  deaths 2009–2019 and stay the same at every knot.
+
+| Knot | Lifetime earnings targets | Education targets, 25–64 | Race and Hispanic origin targets |
+|---|---|---|---|
+| 2010 | Table 10 (2010); 80–84 from 2015 and 85–89 from 2020 by backcast | 2018 national, by backcast (`EDU_BACKCAST_HOLD_FIRST`) | NVSR 63(7), bridged race |
+| 2011–2014 | Figures 7–12; 80–84 and 85–89 by backcast | 2018 by backcast | NVSR 64(11), 65(8), 66(3), 66(4), bridged |
+| 2015 | Table 11 (2015); 85–89 by backcast | 2018 by backcast | NVSR 67(7), bridged |
+| 2016–2017 | Figures 7–12; 85–89 by backcast | 2018 by backcast | NVSR 68(4), 68(7), bridged |
+| 2018 | Figures 7–12; 85–89 by backcast | NVSR 69(13) I–7 | NVSR 69(12), single race |
+| 2019 | Figures 7–12; 85–89 by backcast | NVSR 70(8) I–7 | NVSR 70(19) |
+| 2020 | Table 12 (2020) | NVSR 72(10) I–7 | NVSR 71(1) |
+| 2021 | Figures 7–12 | NVSR 73(8) Table 14 | NVSR 72(12) |
+| 2022 | Figures 7–12 | NVSR 74(4) Table 14 | NVSR 74(2) |
+| 2023 | Held at 2022 | Held at 2022 | NVSR 74(6) |
+| 2024 | Held at 2022 | Held at 2022 | NVSR 75(5) |
+
+Every knot also carries the NHIS-based targets (education at 65+, the
+split, `other_non_hispanic`). "Held" means the fit imposes the same
+published targets again on that year's reference population, so the
+conditional coefficients can still move with the population's
+composition.
+
+### 7.2 Extracting and digitizing the earnings gradient [frozen here]
+
+- **Tables.** Two independent text-layer readers (Poppler's
+  `pdftotext -layout` and a second PDF library) each transcribe Tables
+  1–12 of Study 129 and Tables 4–8 of Study 124; the two transcriptions
+  must agree on every printed string, or the job stops and reports the
+  cells.
+- **Figures 7–12 (pp. 33–38).** The series are vector paths (§3.1). The
+  digitizer decompresses each page's content stream, reads the axis tick
+  labels and their positions from the text layer, fits a linear map from
+  page coordinates to RMR on the y-axis and to year on the x-axis (at least
+  three ticks per axis per panel; a tick fit whose worst residual exceeds
+  0.1 percent of the axis range rejects the panel), assigns each polyline
+  to a legend entry by stroke color and dash pattern, and reads the vertex
+  nearest each integer year (within a quarter year). Each point keeps its
+  page, panel, series and page coordinates.
+- **Validation.** At every table year in which a table prints the band, the
+  digitized RMR must lie within 0.010 of the printed two-decimal value
+  (the printed value's 0.005 rounding plus a plotting tolerance of
+  0.005). There are 210 such points: 21 band-years × 2 sexes × 5
+  quintiles. Every one must pass; if any fails, the figure years are
+  unavailable (§10, `SOURCE_UNAVAILABLE`), and nobody can build P-PRIMARY.
+- **Digitization half-width.** `δ_dig` is the largest
+  `|digitized − printed|` over the 210 validation points, plus 0.005.
+  Each digitized value then stands for the interval `±δ_dig` around it,
+  which §9 treats like a printing interval.
+- **Which values the fit uses.** Printed table values at 2010, 2015 and
+  2020; digitized values at 2011–2014, 2016–2019 and 2021–2022; only for
+  fully supported bands (§3.1).
+- **Coverage.** With the backcasts, every knot has all six earnings
+  bands, both sexes and all four non-reference quintiles; §5 completes
+  ages 0–61 and 90–120.
+
+### 7.3 Registered sensitivities over time [frozen here]
+
+- **P-TABLE (table only).** Earnings targets from Tables 7–12 alone:
+  2010 for knots 2010–2014, 2015 for 2015–2019, and 2020 for 2020–2024;
+  the other dimensions as in P-PRIMARY. Its last earnings level is 2020,
+  so it cannot become the primary path without amending Q6 (S0.1 §2.3).
+- **P-PRE (pre-pandemic).** From the 2020 knot onward, every dimension's
+  targets hold at their 2019 values. Reason: Study 129 reports that the
+  pandemic changed some groups' spreads in 2020 and 2021 and that the
+  changes "seem to be temporary" (p. 9), while Q6's last observed year
+  for earnings, 2022, falls in that period.
+- **E-S1 (reporting-state education).** Education targets at knots
+  2010–2017 from that year's reporting-state tables (§3.2) instead of the
+  2018 backcast, labeled "not national".
+
+### 7.4 The widening alternative P-WIDEN (Q6) [frozen here]
+
+A custom `CoefficientPath` (`specification.py:62–73`) that equals
+P-PRIMARY at every knot through 2024 and, for `t > 2024`, multiplies every
+non-reference **lifetime-earnings** and **education** coefficient, at every
+age and for both sexes, by
+
+`m(t) = min(1 + g × min(t − 2024, 30), 1.5)`, with `g ≥ 0`.
+
+| Element | Value | Source or reason |
+|---|---|---|
+| Origin | 2024 | P-PRIMARY's last knot, so the alternative departs only where no data exist |
+| Earnings slope `g_E(s, B)` | For each sex and OACT band with at least ten annual points in 2005–2019: the least-squares slope, over those years, of `S(t) / S̄`, where `S(t) = L_E(q1)(t) − L_E(q5)(t)` from the annual figures and `S̄` is its mean over the same years. Bands with fewer points (80–84, five points; 85–89, none) take the mean of the 70–74 and 75–79 slopes for the same sex, each computed as above before the zero floor; the floor below then applies to the mean. Ages outside 62–89 take the nearest band's slope | Study 129 calls the gradient's trends "generally, slowly growing" (p. 10); 2005–2019 are the fifteen pre-pandemic years with at least three bands present, and the estimator leaves out the pandemic years because of p. 9 |
+| Negative slopes | Set to zero: no widening for that sex and band, which then equals P-PRIMARY | The alternative is a widening path; a fitted narrowing is not evidence for it |
+| Education slope `g_D(s)` | `g_E(s, 62–64)` | No national education series longer than five years exists (§3.2); 62–64 is the earnings band nearest NCHS's ages |
+| Race and Hispanic origin | `m(t) = 1`, held flat | CBO 41681 says the gaps between whites and blacks "have narrowed somewhat" (p. 2), with a qualifier on p. 1: for men the black–white gap "has been maintained", and black women's larger gains were "halving the gap"; no source here supports a widening race path |
+| Ages | Every age 0–120 (the tails follow §5's rules applied to the multiplied coefficients) | One rule for the whole table |
+| Caps | The multiplier stops growing after 30 years (2054) and never exceeds 1.5 | The source series spans 1995–2022, about 30 years; the alternative carries no trend further than the data it comes from |
+
+The path's `years` are P-PRIMARY's knots plus 2025–2054; every later year
+returns the 2054 table. SM1 revalidates each returned table and still
+normalizes every cell (Q4), so widening moves deaths between groups and
+never adds deaths. The fit job computes the slopes under this
+frozen estimator; this record computes none.
+
+### 7.5 Synchronization
+
+One rule covers every path: all three dimensions share the same knot
+years; at each knot each dimension takes its latest observed targets at
+or before the knot (hold-first before its first observation); the joint
+calibration then yields one complete table holding every dimension ×
+category × sex × single age 0–120, which `RelativeRiskSpec` validates
+(`specification.py:171–197`). No dimension carries its own knot years
+into SM1. **Q11:** SM owns death alignment, so these coefficients assume
+the TR 2026 baseline `q` and nothing downstream aligns deaths a second
+time (design §7.2).
+
+## 8. Schema
+
+### 8.1 Dimensions [frozen here]
+
+This record confirms S0.1's proposed table against
+`mortality_ses/specification.py` (`RiskDimension`, lines 34–59;
+`RelativeRiskSpec`, lines 145–197) and tightens it in one respect: each
+dimension has one unclassified string, and
+the reason a person lacks a category goes in a separate diagnostic column
+`sm_<dimension>_unclassified_reason`, never into the dimension column. SM1
+also treats a missing value as unclassified (`specification.py:224–227`).
+The same column carries `reference_below_25` for a person under 25 in
+`R_ref`, who holds `q3` and `high_school` by rule rather than by
+classification (§6.2). SM1 reads only the dimension and `sex` columns
+(`specification.py:203–224`), so the label changes no risk.
+
+| `name` | `column` | `categories`, in this order | `reference` | `unclassified_values` |
+|---|---|---|---|---|
+| `lifetime_earnings` | `sm_lifetime_earnings_quintile` | `q1`, `q2`, `q3`, `q4`, `q5` (lowest to highest) | `q3` | `("unclassified",)` |
+| `education` | `sm_education` | `less_than_high_school`, `high_school`, `some_college`, `college` | `high_school` | `("unclassified",)` |
+| `race_hispanic_origin` | `sm_race_hispanic_origin` | `hispanic`, `white_non_hispanic`, `black_non_hispanic`, `other_non_hispanic` | `white_non_hispanic` | `("unclassified",)` |
+
+### 8.2 Crosswalks [frozen here]
+
+**Education.**
+
+| SM category | Meaning | NCHS death certificate | NHIS `EDUC1` | ACS | SM3's PSID years (S0.1 §1.2) |
+|---|---|---|---|---|---|
+| `less_than_high_school` | No high school diploma or GED | "Less than high school diploma or GED" | 00–12 | `SCHL` 01–15 (no schooling through "12th grade - no diploma") | Fewer than 12 |
+| `high_school` | Diploma or GED, no college | "High school diploma or GED" | 13–14 | `SCHL` 16–17 ("Regular high school diploma", "GED or alternative credential") | 12 |
+| `some_college` | Some college without a degree, or an associate degree | Within "Some college or collegiate degree" | 15–17 | `SCHL` 18–20 (some college, no degree; associate's degree) | 13–15 |
+| `college` | Bachelor's degree or higher | Within "Some college or collegiate degree" | 18–21 | `SCHL` 21–24 (bachelor's, master's, professional, doctorate) | 16–17 (the reader's upper code) |
+| `unclassified` | — | "Not stated" | 96–99 | None at 25 and over: `SCHL` covers every person aged 3 and over ("bb" marks under 3), and §6.2 holds everyone under 25 at `high_school` | Missing |
+
+The ACS codes come from the data dictionaries for 2010–2019 and 2021–2024
+(§13.6): `SCHL` has the same 24 codes and the same universe in every one
+of them, and `HISP` and `RAC1P` the same code lists (the 2010 and 2011
+dictionaries word `RAC1P` 9 "Two or more major race groups", later ones
+"Two or More Races"). **2020 stays [open]:** its 1-year dictionary sits
+only under the Census Bureau's experimental-data path (§11.6), outside
+what fix round 1 could fetch. The fit job's first step fetches it,
+records its hash, and checks its `SCHL`, `HISP`, `RAC1P`, `AGEP` and
+`SSP` entries against the code lists above and §6.2's frozen texts; if
+they match, 2020 uses this crosswalk, and if they do not, 2020 counts as
+a missing ACS year (§10, `SOURCE_UNAVAILABLE` (iii)). For every year, the
+loader also tests that each code lands in exactly one category. NHIS
+codes come from the 2009, 2016 and 2018 person layouts (pp. 265, 296,
+294).
+
+**Race and Hispanic origin.**
+
+| SM category | NCHS life tables | NHIS (`HISPAN_I`, `RACERPI2`) | ACS | Cristia |
+|---|---|---|---|---|
+| `hispanic` | Hispanic (any race) | `HISPAN_I` 00–11 | `HISP` 02–24 | Hispanic |
+| `white_non_hispanic` | NH white (single race from 2018; bridged 2010–2017) | `HISPAN_I` 12 and `RACERPI2` 01 | `HISP` 01 and `RAC1P` 1 ("White alone") | White |
+| `black_non_hispanic` | NH black | `HISPAN_I` 12 and `RACERPI2` 02 | `HISP` 01 and `RAC1P` 2 ("Black or African American alone") | Black |
+| `other_non_hispanic` | No complete table; AIAN NH and Asian NH (2019–2024) form a partial group | `HISPAN_I` 12 and `RACERPI2` 03–06 (AIAN only, Asian only, not releasable, multiple race) | `HISP` 01 and `RAC1P` 3–9 (American Indian, Alaska Native and other AIAN codes, Asian, NHPI, some other race alone, two or more races) | Other Race |
+| `unclassified` | — | None in the primary; `HISPAN_I` 10 and 11 under sensitivity X-R1 | None: `HISP` and `RAC1P` code every person | — |
+
+`HISPAN_I` 00–11 map to `hispanic` and 12 to the non-Hispanic groups, as
+S0.1 froze it ("00–11 Hispanic, 12 non-Hispanic", audit lines 468–469,
+inside the **[frozen findings]** block that opens at line 453).
+The 2018 layout's note puts respondents who "refused to answer Hispanic
+origin" in 10 and those who "didn't know it" in 11, and adds "When
+Hispanic origin was not ascertained, this variable was also assigned
+"11"" (2018 Sample Adult layout, p. 8). Codes 10 and 11 therefore get the
+same treatment in every run: `hispanic` in the primary, and `unclassified`
+in sensitivity X-R1, which drops both from the race estimates (§4.5). The
+same note places under 12 respondents who "answered 2,7,8, or 9 in
+ORIGIN_I", a variable this record does not read. The public files carry
+no flag showing whether NCHS imputed race or ethnicity (2018 survey
+description pp. 25–26), so the fit cannot set imputed values apart.
+
+**Lifetime earnings.** OACT's Lowest through Highest, Cristia's Bottom
+through Top, the ACS proxy `Q̃` and SM3's PSID quintiles all map to `q1`
+through `q5`. SM3's quintile is `unclassified` when a person has fewer than
+ten observed head or spouse years (S0.1 §1.2).
+
+**How these differ from `estimates/group_breakdown.py`.** That module's
+MINT education dimension (lines 1227–1274) has five bands by years of
+education: `graduate` (more than 16), `bachelor` (16), `associate`
+(14–15), `high_school` (12–13) and `less_than_high_school` (less than 12).
+Its race dimension (lines 1103–1130) uses `hispanic_any_race`,
+`white_non_hispanic`, `black_non_hispanic` and `other_non_hispanic`. SM's
+four education categories are a new crosswalk: SM's `high_school` is 12
+years where MINT's is 12–13, and SM's `some_college` (13–15) and `college`
+(16–17) cut where MINT's bands do not. SM's `hispanic` is the same concept
+as MINT's `hispanic_any_race` under a different string. SM's columns carry
+the `sm_` prefix, SM never reuses MINT's dimension objects, and the fit
+job's tests assert that the two schemes stay separate, so the out-of-scope
+group definitions do not change.
+
+### 8.3 `unclassified_rule` by use [frozen here]
+
+| Use | Rule | Reason |
+|---|---|---|
+| SM2's fit and conformance on `R_ref` | `refuse`, applied after §6.2's under-25 rule and §8.2's crosswalk | Those rules leave no reference-population person unclassified: everyone aged 25 and over carries a `SCHL` code 01–24, everyone carries `HISP` and `RAC1P` codes, everyone under 25 holds the reference on earnings and education, and everyone else at 25 and over takes `Q̃`'s quintile weights (§6.2). A person who still appears unclassified is a construction error, which stops the run and goes into the fit's report |
+| SM2's NHIS estimates | Not through SM1 | §4.5 drops unclassified persons from that dimension's estimate and reports the count |
+| SM3 | `reference` (confirms S0.1 §2.3) | SM3's estimand keeps unclassified persons in the overall ratio and in no group cell (S0.1 §1.1); under `reference` they take zero log risk on that dimension, SM1 still conserves the cell, and the cell log counts them (`specification.py:219–247`; `normalizer.py:17–35`) |
+| Any projection (Track B, the open population) | `refuse` until the frame carries all three attributes and a registration names the rule | `FRAME_ATTRIBUTE_MISSING` (§10) |
+
+A person under 25 classified on lifetime earnings or education needs a
+registration of its own before any use (§5). `R_ref` classifies no one
+under 25 on either dimension: it holds them at the reference categories by
+rule, labeled `reference_below_25` (§6.2, §8.1), and that rule is not such
+a classification.
+
+## 9. Conformance procedure: `SM2_SOURCE_RECONSTRUCTION_AND_CLUSTER_BOOTSTRAP`
+
+**[frozen here]** throughout, unless a step names S0.1. The procedure runs
+on P-PRIMARY under JR-C. Sensitivities run the same steps and report;
+they do not gate. Its order:
+
+1. Pin every input: source files at the hashes in §13 and the fit job's
+   manifest, module versions, the seed (§9.3), and the reference
+   population files.
+2. Build P-PRIMARY's fifteen tables (§7.1) and validate each with
+   `RelativeRiskSpec`.
+3. Run the deterministic reconstruction checks (§9.1).
+4. Compute the e65 residuals (§9.2) and every family member's
+   simultaneous bound (§9.3, §9.4).
+5. Run the invariants (§9.6).
+6. Apply the pass conjunction (§9.5).
+
+### 9.1 Deterministic reconstruction checks (tier A)
+
+These run at the point estimate and carry no Monte Carlo error. Any
+failure stops the procedure (§10, `GATE_FAIL_OR_REGRESSION_FAIL`).
+
+- **A1, exact printing intervals.** For a target built from two printed
+  values `r_k` and `r_ref` with half-width `h`, the interval for
+  `L = ln r_k − ln r_ref` uses the transformed endpoints of both:
+  `[ln(r_k − h) − ln(r_ref + h), ln(r_k + h) − ln(r_ref − h)]`. The
+  procedure never uses the approximation `h / r` as a log tolerance (S0.1
+  §2.5). The half-widths are 0.005 for an RMR; 0.0000005 for a six-decimal `qx`;
+  0.5 for a printed `lx` or `Tx` (used only in §9.2's open interval);
+  0.05 per 100,000 for an NCHS education rate, carried through
+  `q = m / (1 + m/2)`; `δ_dig` (§7.2) for a digitized RMR; and zero for an
+  NHIS estimate this project computes. The model-implied contrast must lie
+  inside the interval widened by **ε = 1 × 10⁻⁶** in log units on each
+  side. ε is the only extra residual tolerance this record sets: it covers
+  the solver's `1 × 10⁻¹⁰` convergence and floating-point accumulation,
+  and it cannot relax e65's limit. Where `r − h ≤ 0` leaves an endpoint
+  undefined, that side of the interval has no bound and the check applies
+  to the other side only.
+- **A2, the source agrees with itself.** For every quintile, band, sex and
+  table year, the interval for `ln(rate_k / rate_q3)` from Tables 1–6
+  (four-decimal rates) must overlap the interval for `L_E(k)` from Tables
+  7–12. A failure means a transcription error, not a source finding.
+- **A3, digitization.** All 210 validation points pass (§7.2).
+- **A4, two transcriptions agree** on every printed string (§7.2).
+- **A5, SM1's cell log on `R_ref`.** At every knot, no cell is
+  infeasible, every feasible cell conserves its deaths, and the fit reports
+  the counts of capped rows (`normalizer.py:17–35`).
+- **A6, coverage.** Every knot's table holds every dimension × category ×
+  sex × age 0–120 once, with zero reference coefficients
+  (`specification.py:171–197`).
+
+### 9.2 Life expectancy at 65 [frozen here; limit frozen, S0.1 §2.5]
+
+- **Reference population and marginalization.** For group `g` of a
+  dimension, sex `s`, knot `t` and single age `a`: `h_g(a)` is the
+  average of SM1's normalized probabilities over `g`'s rows of `R_ref(t)`
+  (the other two dimensions' categories weighted by `w`), within the
+  target's subpopulation: Social Security recipients at 62–89 for OACT
+  targets and the whole population otherwise, and the whole population at
+  every age above the subpopulation's range. `h_all(a)` is the same
+  average over every row of that subpopulation. The fitted group
+  probability is `q_F,g(a) = q_base(s, a, t) × h_g(a) / h_all(a)`.
+- **Earnings targets.** `q_T,g(a) = min(1, q_base(s, a, t) × ρ_g(a))` at
+  every age 0–120, with `ρ_g` the RMR of `g`'s band at the knot (after
+  backcasts and holds; above 89 the 85–89 RMR), and
+  `e65_T,g = life_expectancy(q_T,g[0..120], 65)` under the closing rule
+  below. OACT publishes RMRs, not life tables, so it has no closing of its
+  own, and S0.1 asks for a fixed tail rule (audit lines 643–646).
+- **Race and Hispanic-origin targets keep NCHS's conventions.** S0.1
+  requires "the source's age/closing conventions" for NCHS life-table
+  targets (audit lines 647–648). At single ages 65–99,
+  `q_T,g(a) = min(1, q_base(s, a, t) × ρ_g(a))`, with
+  `ρ_g(a) = q_g(a) / q_total(a)` from the knot's NCHS tables for
+  `hispanic`, `white_non_hispanic` and `black_non_hispanic` (M-R6), and
+  `ρ(a) = HR_b(a) × q_NHW(a) / q_total(a)` for `other_non_hispanic`, with
+  `HR_b(a)` the NHIS hazard ratio of `a`'s band. Survivors run
+  `l(65) = 1` and `l(a + 1) = l(a)(1 − q_T,g(a))`, and person-years
+  `L(a) = (l(a) + l(a + 1)) / 2`, NCHS's equation 17 (NVSR 75(5) p. 61).
+  The open interval closes as NCHS closes it. NCHS sets `∞q100 = 1.0`,
+  `∞d100 = l100` and `∞L100 = T100`, "the sum of the extrapolated Lx
+  values for ages 100–120" (pp. 61–62), so the interval's life expectancy
+  is `e100 = T100 / l100`. The target applies the group's ratio to the
+  baseline as M-R6 does at single ages:
+  `e100_T,g = e100_base × e100_g / e100_total`, where `e100_g` and
+  `e100_total` divide the printed `Tx` by the printed `lx` in the "100 and
+  older" rows of the group's and the total's tables, and
+  `e100_base = life_expectancy(q_base(s, ·, t)[0..120], 100)`. For
+  `other_non_hispanic`, `e100_T = e100_base × (e100_NHW / e100_total) /
+  HR_85+`. An open interval's central death rate is `l100 / T100 =
+  1 / e100`, so this applies the source's ratio of open-interval death
+  rates. Then `e65_T,g = (Σ_{a=65..99} L(a) + l(100) × e100_T,g) / l(65)`,
+  which equals `life_table.py`'s backward recursion from 99 down to 65
+  seeded with `e100_T,g`.
+- **The model side, and the closing rule.** `e65_F,g =
+  life_expectancy(q_F,g[0..120], 65)` for every group: the backward
+  recursion with the constant-terminal-`q` closing at the last age
+  (`life_table.py:17–37, 181–203`), applied to the fitted probabilities,
+  whose race coefficients at 100–120 hold the age-99 value (§5). The
+  module calls that closing "a builder default awaiting ratification"
+  (lines 30–31); this record fixes it for the model side, the earnings
+  targets and `e100_base` only. Target and model therefore close the top
+  differently for race and Hispanic origin, so a mismatch in the tail
+  reaches the residual instead of cancelling. No ruling this record cites
+  prescribes the earlier construction, in which both sides shared the
+  repository's tail: the design's §5.3 sets the 0.1-year limit, and the
+  design mentions the closing only to call it "a builder-default
+  constant-terminal-q closing rule that awaits ratification" (line 143;
+  line 1146 repeats that §2.1 discloses it). So no sensitivity keeps that
+  construction.
+- **The limit.** `|e65_F,g − e65_T,g| ≤ 0.1` year, **the design's
+  limit, which stays at 0.1 year** (design §5.3), for every earnings
+  quintile and every race and Hispanic-origin group, both sexes, at every
+  knot. Education has no e65 target: NCHS stops at 64 and S0.1 applies the
+  limit to earnings groups and NCHS life-table targets; education's
+  targets are contrasts under §9.1 and §9.3.
+
+### 9.3 Uncertainty
+
+S0.1 froze the requirements (§2.5): simultaneous 95 percent bounds,
+Bonferroni over the family; NHIS resampling that respects strata, PSUs
+and family clusters; ACS replicate weights for its denominators;
+shared-source covariance; rounding, random variation, survey sampling and
+linkage sensitivity kept apart; no independent-person bootstrap; no
+complete-count rate treated as free of uncertainty. This record
+implements them as four components, each reported on its own and then
+combined.
+
+| Component | Source | How the engine draws or computes it |
+|---|---|---|
+| U1, rounding and digitization | Every printed value; every digitized RMR | Uniform within the printing interval of §9.1, independently for each printed cell; uniform within `±δ_dig` for each digitized value |
+| U2, random mortality variation | NCHS education | Poisson on the printed death counts: the `1/D` term of Formula 9 (NVSR 74(4) p. 132) and nothing else. Formula 9's ACS term `(SE(P)/P)²` enters through U3's replicate refits, so the ACS variance counts once |
+| | NCHS life tables | Poisson with expected deaths `q_x × P_x`, where `P_x` is the group's population at age `x` in `R_ref(t)`. The life-table reports print no death counts, so this is a registered approximation, used at the Medicare-supplemented ages too. For the open interval (§9.2), Poisson on its central rate `1 / e100` with expected deaths `P_100+ / e100`, `P_100+` being the group's population at 100 and over in `R_ref(t)` |
+| | OACT | **[open]** Study 129 prints rates and RMRs but no deaths or exposures (Tables 1–12), so the source cannot supply this component (row S2-24). It settles when OACT's exposure or death counts by sex × band × quintile × year are available. Until then every OACT e65 member is inconclusive or fails (below) |
+| U3, survey sampling | ACS (reference population and NCHS education denominators) | One refit per ACS replicate, with the variance formula that each year's PUMS documentation gives and the fit job records. NCHS's denominators move with the same replicate: each category's share within sex and band is re-tabulated from that replicate and scaled to NCHS's printed total. The fit job also reports the ratio of its replicate standard errors to NCHS's printed ones (Table VIII, I–19) |
+| | NHIS–LMF | Rao–Wu rescaled bootstrap: PSUs resampled with replacement within strata, separately for each design period: the 2009 layout's `STRAT_P`/`PSU_P` and the 2016 and 2018 layouts' `PSTRAT`/`PPSU`, with the periods' strata kept distinct, as the 2018 survey description directs for pooled years: years within one design period pool as one, and different periods count as statistically independent, with distinct strata (pp. 92–93); the fit job reads which names each intervening year uses. Resampling whole PSUs keeps every family and every person's whole history together, once the fit job confirms from each year's survey description that families nest within PSUs. Where a description does not confirm it, the fit job checks the files instead: a family whose records carry more than one PSU or stratum merges those PSUs, and their strata, into one resampling unit, and the fit reports how many merges it made |
+| U4, linkage and perturbation sensitivity | NHIS–LMF | Linkage: one discrete alternative fit, not draws, with NHIS weights replaced by the original sample-adult weight and ineligible persons dropped; the fit reports its conformance, and it does not gate. Perturbation: **[open]**, because public data cannot assess it. NCHS substituted synthetic data "for follow-up time or underlying cause of death for select records" and did not perturb vital status (LMF description p. 1), and the public data dictionary flags none of the altered records, so no public alternative can isolate the effect. Only the restricted-use linked files, which this project excludes (§3.4), would settle it. Every NHIS target, and every comparison that uses NHIS, carries the label "public-file perturbation not assessed". X-R1, which recodes `HISPAN_I` 10 and 11, is a crosswalk sensitivity (§8.2), not part of U4 |
+
+**What OACT's open U2 means [frozen here].** S0.1 keeps a target whose
+source does not publish the uncertainty it needs "open or inconclusive",
+and it lets no complete-count rate count as free of uncertainty (audit
+lines 661–669). Every Monte Carlo draw refits the model to that draw's
+targets (§9.4), so a source's random variation moves target and fit
+together, but it still reaches an e65 residual through capping,
+composition and tails. Leaving OACT's term out would narrow the earnings
+e65 bounds and make passing easier, so this record applies S0.1's rule:
+
+- **Every OACT e65 member is inconclusive while the component stays
+  open.** The 150 members of e65 by earnings quintile (§9.4) cannot pass:
+  one whose point residual lies inside ±0.1 year is inconclusive, and one
+  whose point residual lies outside fails (§9.5). The component closes
+  only when OACT's deaths or exposure by sex × band × quintile × year
+  reach the project, through a recorded amendment (§11.1).
+- **What that stops.** Under §9.5 an inconclusive member makes lifetime
+  earnings inconclusive, and `WEAK_POWER_OR_VACUITY` blocks that
+  dimension's admission (§10). SM2 passes as a whole only when all three
+  dimensions pass, so **SM2 cannot earn admission while this item stays
+  open**. The fit still runs every step and publishes every member's
+  residual and bound (except what §11.5 withholds for an NHIS target
+  below a floor), and each dimension's status, so education and race
+  and Hispanic origin carry verdicts of their own.
+- **The OACT contrast members keep their verdicts.** Each draw refits to
+  its own targets, so a contrast member checks the reconstruction (§9.4);
+  OACT's random variation does not enter its residual.
+- **A Poisson sensitivity, reported and never a verdict.** Beside each
+  OACT e65 member, the fit reports a bound with an added Poisson term:
+  at each single age of a band, expected deaths equal the target's
+  probability times one fifth of that age's Social Security recipients of
+  that sex in `R_ref(t)`. It carries the label "likely understates OACT's
+  random variation", because `R_ref`'s recipients include beneficiaries
+  that OACT's population leaves out (M-E4), which overstates exposure. It
+  changes no verdict.
+- **Two labels stay.** The lifetime-earnings coefficient intervals that
+  SM2 hands to SM3 and SM4 carry "OACT random variation not included",
+  and the Study 129 versus Study 124 comparison, which the fit reports,
+  cannot trigger `GRADIENT_SOURCES_INCONSISTENT` (§10).
+
+**The engine.**
+
+- **Monte Carlo draws** cover U1, U2 and NHIS's part of U3: `B = 2,000`
+  draws, each a full refit of every knot and sex (§6.3) followed by every
+  family member's residual. The seed is
+  `numpy.random.SeedSequence(2026110101)`, spawned into one child stream
+  per component and source in the order of §3's tables.
+- **ACS replicates** cover the ACS part of U3: one refit per replicate,
+  with the Monte Carlo inputs at their point values.
+- **Component runs, reported only.** One joint Monte Carlo cannot show
+  what each component contributes, so the fit also runs U1, U2 and NHIS's
+  part of U3 one at a time: `B` draws of that component alone, from the
+  same child stream the joint run uses, with every other input at its
+  point value and a full refit per draw. The ACS replicate refits already
+  isolate the ACS part of U3, and U4's alternative is a discrete fit. For
+  every family member the fit reports each component's standard deviation
+  beside the joint one. The component runs set no bound and change no
+  verdict: the bound uses the joint draws and the ACS replicates, because
+  capping and the tails make the response nonlinear, and the components'
+  variances then need not add up to the joint variance.
+- **Combination.** For family member `k`,
+  `SD_k = sqrt(Var_MC,k + Var_ACS,k)`, treating ACS sampling as independent
+  of the other components. NCHS's education denominators are ACS too, and
+  they sit inside `Var_ACS` with the reference population.
+- **Shared-source covariance.** Each draw perturbs each printed source
+  value once, and every target, refit and e65 in that draw uses that same
+  value: the `q3` RMR enters four contrasts; an NH white `qx` enters two
+  race contrasts and every race e65 target; an ACS replicate enters the
+  reference population and NCHS's denominators together.
+- **The simultaneous bound.** `r̂_k ± z × SD_k`, with
+  `z = Φ⁻¹(1 − 0.025 / 6,884) ≈ 4.486`, the two-sided 95 percent
+  Bonferroni bound over the family (§9.4). If a member's Monte Carlo
+  draws have skewness above 1 in absolute value, its bound uses the
+  first-order Cornish–Fisher quantile `z + (z² − 1) γ / 6` in place of `z`.
+- **Monte Carlo error.** For each member, the Monte Carlo standard error
+  of the bound's half-width, `z × SD_MC,k / sqrt(2(B − 1))`, must stay under
+  one tenth of the member's tolerance `τ_k`: less than 0.01 year for e65,
+  and less than `ε / 10` for a contrast. If it is not, `B` doubles, up to
+  32,000; if the rule still fails, the member is inconclusive. Tier A has
+  no simulation error.
+
+### 9.4 The critical family: 6,884 targets [frozen here]
+
+| Block | Count | Tolerance `τ` |
+|---|---:|---|
+| OACT contrasts (M-E1): 63 observed band-years in 2010–2022 (four bands × 13 years, 80–84 × 8, 85–89 × 3) × 2 sexes × 4 quintiles | 504 | ε |
+| e65 by earnings quintile: 5 × 2 sexes × 15 knots | 150 | 0.1 year |
+| NCHS education contrasts (M-D1, M-D2): 5 years × 2 sexes × 4 bands × 2 contrasts | 80 | ε |
+| NHIS education contrasts at 65+ (M-D4): 3 × 3 bands × 2 sexes | 18 | ε |
+| NHIS `college`/`some_college` split (M-D3): 2 sexes | 2 | ε |
+| NCHS race contrasts (M-R1): 2 groups × 100 ages × 2 sexes × 15 years | 6,000 | ε |
+| e65 by race and Hispanic-origin group: 4 × 2 sexes × 15 knots | 120 | 0.1 year |
+| NHIS `other_non_hispanic` contrasts (M-R4): 5 bands × 2 sexes | 10 | ε |
+| **Total** | **6,884** | |
+
+- **Joint-survey targets.** The 30 NHIS targets are in the family; they
+  are the only targets whose estimates this project computes from survey
+  records. NCHS's education rates also rest on a survey, through their
+  ACS denominators (§3.2), and U3 carries that sampling error (§9.3).
+- **One member per published or estimated quantity.** A target imposed at
+  several knots (a backcast copy, a held target or an NHIS target) is one
+  member, tested by its largest absolute residual across the knots that
+  impose it. The e65 targets differ by knot, because the baseline does, so
+  each knot's is a member.
+- **What the contrast members test.** For a contrast, the residual is the
+  model-implied contrast minus the draw's target after that draw's refit.
+  Because §6.3's system has as many equations as unknowns, a contrast
+  member shows whether every refit reproduces its targets: a
+  reconstruction check. The
+  e65 members compare two different constructions (the model on the
+  reference population against the source's ratios on the baseline), so
+  they also test composition, capping and tails. For earnings that test
+  is narrow: with `Q̃`'s exact fifths among recipients at 62–89 (§6.2),
+  each earnings e65 member checks chiefly whether its band's five printed
+  RMRs average to one, plus capping, the spread of composition across the
+  band's single ages, and the tail above 89. The race and
+  Hispanic-origin e65 members also test the model's tail against NCHS's:
+  their targets close the "100 and older" interval as NCHS does, while the
+  model keeps the repository's closing (§9.2).
+- **What the contrast members cost.** The 6,614 contrast members cannot
+  fail by chance, yet they raise the Bonferroni constant for the 270 e65
+  members from `Φ⁻¹(1 − 0.025 / 270) ≈ 3.738` to about 4.486, which makes
+  those bounds about 20 percent wider; the family still follows S0.1's
+  definition (audit lines 657–659), and this record leaves it unchanged.
+- **Not in the family.** The sensitivities of §5, §6.5, §7.3 and §8.2, and the
+  sources reported beside the fit.
+
+### 9.5 Pass conjunction
+
+SM2 earns admission, labeled "calibrated, not validated", only if all
+of these hold:
+
+1. every tier A check passes (§9.1);
+2. every one of the 6,884 members has its simultaneous bound inside
+   `±τ_k`, with the Monte Carlo rule met (§9.3);
+3. every invariant of §9.6 holds;
+4. no condition in §10 blocks.
+
+A member whose point residual lies inside `±τ_k` but whose bound crosses
+it is **inconclusive**; a member whose point residual lies outside fails.
+An inconclusive member makes its dimension inconclusive (e65 by quintile
+and OACT contrasts: lifetime earnings; education contrasts: education;
+race contrasts, race e65 and `other_non_hispanic`: race and Hispanic
+origin). **A bound too wide to judge never passes.** The record publishes
+each dimension's status; SM2 passes as a whole only when all three pass,
+because SM1 needs all three. While OACT's random-variation component
+stays open, every e65 member by earnings quintile is inconclusive or
+fails whatever its computed bound (§9.3), so lifetime earnings cannot
+pass, and neither can SM2 as a whole.
+
+### 9.6 Invariants (design §6)
+
+- **Life tables.** Every fitted group life table (`q_F,g` for each knot,
+  sex and group) satisfies `data/life_table.py`'s invariants:
+  `period_life_table` and `life_expectancy` agree, both with `f0 = 0.5`,
+  the uniform-deaths convention the module applies at every other age
+  (`life_table.py:13`); `f0` moves only `L0` and `e0` (lines 143 and
+  173–177), which no SM2 quantity reads; survivorship and
+  person-years are nonnegative and nonincreasing with age; changing the
+  radix changes no life expectancy; raising any `q` cannot raise life
+  expectancy at an earlier age (`life_table.py:39–51`).
+- **Registered ordering.** For each sex and knot: if the earnings e65
+  targets rise strictly from `q1` to `q5`, the fitted e65 must rise the
+  same way; if the targets do not (Study 129 reports exceptions for women
+  at 62–64 in a few years between 1995 and 2010, p. 8), the check reports
+  "source unordered" and does not fail. NHIS's education hazards at 65+
+  follow the
+  same rule (`less_than_high_school` ≥ `high_school` ≥ `some_college` ≥
+  `college` in the targets implies the same in the fit). Race and
+  Hispanic origin carry no registered ordering.
+- **Permutation invariance.** Refitting with every input table's rows
+  and the reference population's rows permuted gives coefficients equal
+  within `1 × 10⁻¹²` in log units.
+- **SM1.** Each table passes `RelativeRiskSpec`; SM1's cell conservation
+  holds on `R_ref` at every knot; SM1's own test modules pass unchanged.
+
+## 10. Named conditions
+
+This record fixes every fallback below now, before any outcome. **Nobody
+tunes anything so that sources agree** (S0.1 §2.5, frozen).
+
+| Condition | Trigger for SM2 | What it stops | Fallback |
+|---|---|---|---|
+| `GRADIENT_SOURCES_INCONSISTENT` | Two comparable sources disagree beyond SM2's fit tolerance. "Comparable": the same SM category definitions and sex, overlapping ages, calendar windows that overlap at least half the shorter one, and populations that differ at most in institutional coverage. Assessed: (a) Study 129 against Study 124 at table years 1995–2015, bands 62–84; (b) NCHS life tables against NHIS–LMF for `hispanic` and `black_non_hispanic` at 25–84; (c) NCHS education for 2018 and 2019, the two fitted NCHS years inside NHIS's 2009–2019 window, against NHIS–LMF at 25–64, compared year by year as (b) compares, with NHIS's estimator (§4.5) run at attained ages 25–64 by sex and with `some_college` and `college` pooled into one category to match M-D2. Not assessed: OACT against Cristia (different years, ages and earnings measures), and `other_non_hispanic` against the partial composite (different groups). Metric: the difference in temporary life expectancy over the pair's shared ages, each built by applying the source's ratios to the TR 2026 baseline of the year compared: for (a), from 62 to the top of the year's oldest band printed in both studies, per table year, sex and quintile (5 × 2 × 5 = 50 comparisons); for (b), from 25 to 85, per NCHS year 2010–2019, sex and group (10 × 2 × 2 = 40); for (c), from 25 to 65, per NCHS year, sex and contrast (`less_than_high_school` and `some_college_or_more`, each against `high_school`; 2 × 2 × 2 = 8). NHIS's hazard ratios act on the baseline's hazard, `q = 1 − (1 − q_base)^HR` (M-D4's constant-force mapping). Triggered when a comparison's simultaneous 95 percent bound, Bonferroni over these 98 comparisons (`z = Φ⁻¹(1 − 0.025 / 98) ≈ 3.475`), excludes ±0.1 year | Nothing about SM2's admission. OACT stays primary for earnings and NCHS for race and for education at 25–64; the other source of each pair keeps its role beside the fit (§3.6), and (c)'s pooled-category NHIS estimate serves this comparison only; the fit's report publishes the disagreement. A triggered (b) or (c) leaves NHIS's fitted roles (education at 65+, the `college`/`some_college` split and `other_non_hispanic`) in the fit, each labeled "NHIS disagrees with NCHS at overlapping ages", with the triggering comparisons named beside them, and nobody tunes anything to make the sources agree. Pair (a) has no random-variation term (§9.3), so the fit reports a disagreement it finds and it does not trigger | None needed |
+| `SOURCE_UNAVAILABLE` | (i) The digitization fails validation, or the digitizer cannot read a figure; (ii) an NHIS or LMF file is missing or its hash differs; (iii) an ACS year is missing; (iv) the ordered logit does not converge for some knot, sex or ACS replicate (§6.2); (v) any source file's hash differs from §13 or the fit job's manifest (the ACS documents of §13.6 follow §6.2's text rule instead); (vi) no ACS year's income-item documents pass §6.2's acceptance rule; (vii) an NCHS report or table file that §3 lists is missing | (i) P-PRIMARY, and with it SM2's admission; (ii) admission of education and race; (iii) one missing year: nothing, the knot uses the nearest observed ACS year, the earlier one on a tie, labeled; two or more: SM2's admission; (iv) JR-C, and with it SM2's admission; (v) the run; (vi) JR-C, and with it SM2's admission (a year that fails alone only borrows `λ`, labeled, §6.2); (vii) one missing year of a dimension: nothing, that year counts as unobserved under §7.5's rule (the knot takes the latest observed year at or before it; a knot before the first observation takes that first observation), labeled; its contrast members leave the family, and `z` keeps the frozen family size of §9.4; two or more missing years of one dimension: that dimension's admission, and with it SM2's | (i) P-TABLE, computed and published labeled, never substituted; the Q6 question goes to Max (§12). (ii) T-D1 for education at 65+, Cristia's conditional `college`/`some_college` contrast for the split, T-R2 for `other_non_hispanic` at 2019–2024, with knots 2010–2018 taking the 2019 composite by §7.5's hold-first rule, labeled "partial group, held back from 2019", so no knot falls to a default zero; all labeled, none admitted. (iv) and (vi) JR-B, labeled, not admitted |
+| `EXPOSURE_UNRESOLVED` | Any report calls a comparison with a source reported beside the fit out of sample or a holdout, or describes a later role as blind to these sources | That claim | None: every SM2 source is an exposed target (§1) |
+| `FRAME_ATTRIBUTE_MISSING` | A frame that would use SM2's coefficients lacks a dimension's column; the Microcosm frame lacks education (design §7.2) | Use of the coefficients on that frame; closed-cohort work continues, labeled | None. The reference population's missing lifetime earnings is a declared proxy (§6.2), not this condition |
+| `SMALL_CELL_DISCLOSURE` | A published aggregate from NHIS or ACS microdata rests on fewer than 100 unweighted persons, or a published death-based cell on fewer than 20 deaths (§11.5) | Publication of that artifact until the writer suppresses the cell and any complement that reveals it | Suppression |
+| `CELL_CONSERVATION_BROKEN` | SM1's cell log on `R_ref` at any knot shows an infeasible cell, or a feasible cell that does not conserve deaths | SM2's admission | None |
+| `GATE_FAIL_OR_REGRESSION_FAIL` | A tier A check fails; a family member's point residual lies outside `±τ`; a knot's calibration fails to converge or converges to two solutions (§6.3); an invariant fails; an SM1 test regresses | SM2's admission and everything resting on it: SM3's registration on these coefficients and SM4's use of them. The fit publishes the result with no seed selection and no tolerance relief; a new attempt needs a new registered version and enters the lane's attempt ledger (design §8) | None |
+| `WEAK_POWER_OR_VACUITY` | A family member is inconclusive (§9.5), which includes every OACT e65 member while OACT's random-variation component stays open (§9.3); or a report presents the contrast members' pass, which reproduces inputs by construction, as evidence that the gradients are right | The affected dimension's admission, and with it SM2's admission as a whole (§9.5); the claim | None. SM2 makes no 0.90-power claim (S0.1 §2.5). The fit still publishes every residual, bound and dimension status, except what §11.5 withholds |
+
+## 11. Build plan
+
+### 11.1 When the fit job starts
+
+**The fit job starts only after this record merges.** Its brief and its
+first commit cite the merge commit and this file's blob hash at that
+commit (`git rev-parse <merge>:docs/design/mortality_ses_sm2_freeze.md`),
+and it refuses to run if the file at its base differs. Any change to a
+frozen item after the merge needs a recorded amendment and a fresh review
+round, as S0.1's header requires.
+
+### 11.2 What it builds, in order
+
+All new modules sit under `src/populace_dynamics/mortality_ses/`. SM1's
+modules stay unchanged: the `CoefficientPath` protocol,
+`HoldLastCoefficientPath`, `RelativeRiskSpec` and `normalize_probabilities`
+cover every interface SM2 needs (design §5.1: a new module and a new
+registered version per milestone).
+
+1. **`sources.py`.** The source registry: for each input, its publisher,
+   title, release date, URL, retrieval date, byte size, SHA-256, and the
+   tables and pages used. Every reader checks the hash first and refuses a
+   mismatch.
+2. **`extract_oact.py`.** Two independent transcriptions of Study 129
+   Tables 1–12 and Study 124 Tables 4–8, and their reconciliation; the
+   Figures 7–12 vector digitizer with its 210-point validation (§7.2).
+3. **`extract_nchs.py`.** Two independent transcriptions of the education
+   tables (2018–2022, and 2010–2017 for E-S1) and the life tables
+   (2010–2024).
+4. **`nhis_lmf.py`.** Readers for the NHIS 2009–2018 sample-adult and
+   person files and the LMF files, the harmonization of §3.4 and §8.2, the
+   estimator of §4.5 and the Rao–Wu bootstrap of §9.3. Its estimates
+   pass to the fit in memory, inside the one registered run (item 10),
+   and never leave it; only the run's published artifact leaves it, and
+   every NHIS- or LMF-derived cell there meets §11.4 and §11.5.
+   Suppression acts on that output alone, so the fit never consumes a
+   suppressed value.
+5. **`reference_population.py`.** The ACS reader, the crosswalks, `Q̃`, the
+   ordered logit and `R_ref(t)` with its replicates (§6.2). The
+   calibration reads `R_ref(t)`'s full cell weights in memory, inside the
+   same run; suppression applies only to what the run publishes (§11.5).
+6. **`targets.py`.** The mappings of §4: extracted values to targets, with
+   printing intervals and assumption labels.
+7. **`joint_fit.py`.** The calibration of §6.3, and the JR-A, JR-B, JR-D,
+   JR-E and scaled-`λ` routes.
+8. **`paths.py`.** P-PRIMARY, P-TABLE, P-PRE, E-S1 and
+   `WideningCoefficientPath` (§7).
+9. **`conformance.py`.** §9 in full, and the comparisons of §10.
+10. **`scripts/run_sm2_fit.py`**, one registered entry point that writes
+    one new artifact (an exclusive output path, with an abort record if it
+    stops), and **`docs/design/mortality_ses_sm2_fit.md`**, the fit's
+    design note with its invariants.
+
+### 11.3 The extracted-table format
+
+One CSV per source table or figure, and a JSON sidecar holding the
+source's registry entry. Columns:
+
+| Column | Content |
+|---|---|
+| `source_id`, `source_sha256` | Registry key and the file's hash |
+| `table_or_figure`, `pdf_page`, `printed_page` | Locator |
+| `row_label`, `column_label`, `category_source` | Verbatim strings from the source |
+| `category_sm`, `sex`, `age_lower`, `age_upper`, `year` | Mapped keys (§8.2) |
+| `quantity` | One of `rmr`, `death_rate`, `qx`, `rate_per_100000`, `deaths`, `population`, `standard_error`, `odds_ratio`, `z_statistic`, `mortality_ratio` |
+| `printed_string`, `value`, `decimals`, `half_width` | The string as printed, its parsed value, its printed decimals and its printing (or digitization) half-width |
+| `method`, `reader_version` | `text_layer_a`, `text_layer_b` or `vector_digitized`, and the tool version |
+| `transcription_check`, `validation_residual` | Whether the two transcriptions agree; for digitized values, the residual at validation points |
+| `assumption_labels` | The §4–§7 labels that apply (for example `M-E2;BAND_BACKCAST_HOLD_FIRST`) |
+
+### 11.4 How microdata stays out of the repository and out of a job's context
+
+- NHIS, LMF and ACS microdata live outside every repository, worktree,
+  scratch folder and Subfleet job folder, under one data directory that Max
+  names. Code finds it through the environment variable
+  `SM2_MICRODATA_ROOT`, and no prompt to a lane that may open workspace
+  files names that path.
+- No agent displays a record: no `head`, `tail`, `cat`, `less`, sample,
+  `print(df)` or row indexing of a microdata file, and no per-person value
+  in a log, test output, fixture, commit, PR body, issue comment or prompt.
+- Readers catch every exception and re-raise it without row content.
+- Every output that reaches an agent is an aggregate over at least two
+  records; anything published also meets §11.5.
+- Tests use synthetic fixtures built from the layouts, never survey
+  records.
+- Runs on real microdata go through the registered entry point on a
+  machine that holds the data directory, and print aggregates only.
+
+### 11.5 Suppression rule
+
+Every published aggregate derived from NHIS, LMF or ACS microdata rests on
+at least **100 unweighted persons** (the SSA MINT8 Table User Guide's
+"Sample Size Restrictions" minimum that S0.1 cites), and every published
+death-based cell on at least **20 deaths** (NCHS's own threshold for
+showing a rate, NVSR 74(4) p. 132). The writer suppresses a cell below
+either floor together with every complement that would reveal it by
+subtraction, and refuses an artifact that still holds one. This is a
+publication rule, not a precision guarantee (S0.1 §2.5).
+
+**An NHIS target below a floor.** A target whose estimate rests on fewer
+than 100 unweighted persons or 20 deaths stays in the fit and in the
+critical family: the run computes its residual and bound in memory, and
+the artifact publishes its verdict (pass, inconclusive or fail) and
+withholds its value, standard error, bound and residual, together with
+every complement that would reveal them.
+
+### 11.6 Public files to download
+
+| Files | URL | Checked in this job |
+|---|---|---|
+| Study 129, Study 124 | https://www.ssa.gov/oact/NOTES/pdf_studies/study129.pdf, https://www.ssa.gov/OACT/NOTES/pdf_studies/study124.pdf | Bytes in S0.1's source folder at the hashes in §13 |
+| US Life Tables 2010–2024 | Under https://www.cdc.gov/nchs/data/nvsr/: `nvsr63/nvsr63_07.pdf` (2010), `nvsr64/nvsr64_11.pdf` (2011), `nvsr65/nvsr65_08.pdf` (2012), `nvsr66/nvsr66_03.pdf` (2013), `nvsr66/nvsr66_04.pdf` (2014), `nvsr67/nvsr67_07-508.pdf` (2015), `nvsr68/nvsr68_04-508.pdf` (2016), `nvsr68/nvsr68_07-508.pdf` (2017), `nvsr69/nvsr69-12-508.pdf` (2018), `nvsr70/nvsr70-19.pdf` (2019), `nvsr71/nvsr71-01.pdf` (2020), `nvsr72/nvsr72-12.pdf` (2021), `nvsr74/nvsr74-02.pdf` (2022), `nvsr74/nvsr74-06.pdf` (2023), `nvsr75/nvsr75-05.pdf` (2024) | Paths from the NVSR listing; 2010, 2017, 2018, 2019, 2022, 2023 and 2024 downloaded and their table lists read |
+| *Deaths: Final Data*, 2018–2022 | Under the same root: `nvsr69/nvsr69-13-508.pdf` and `nvsr69/nvsr69-13-tables-508.pdf` (2018); `nvsr70/nvsr70-08-508.pdf` and `nvsr70/nvsr70-08-tables-508.pdf` (2019); `nvsr72/nvsr72-10.pdf` and `nvsr72/nvsr72-10-tables.pdf` (2020); `nvsr73/nvsr73-08.pdf` (2021); `nvsr74/nvsr74-04.pdf` (2022) | All downloaded |
+| *Deaths: Final Data*, 2010–2017 (E-S1) | `nvsr61/nvsr61_04.pdf`, `nvsr63/nvsr63_03.pdf`, `nvsr63/nvsr63_09.pdf`, `nvsr64/nvsr64_02.pdf`, `nvsr65/nvsr65_04.pdf`, `nvsr66/nvsr66_06.pdf`, `nvsr67/nvsr67_05.pdf`, `nvsr68/nvsr68_09-508.pdf`, and the internet-table files each report's list of internet tables names | Reports downloaded; internet-table files not located |
+| Public LMF, NHIS 2009–2018 | https://ftp.cdc.gov/pub/Health_Statistics/NCHS/datalinkage/linked_mortality/NHIS_2009_MORT_2019_PUBLIC.dat through `NHIS_2018_MORT_2019_PUBLIC.dat`, and `R_ReadInProgramAllSurveys.R` in the same folder | Names, sizes and dates from the folder listing; no file downloaded |
+| NHIS 2009–2018 data files | https://ftp.cdc.gov/pub/Health_Statistics/NCHS/Datasets/NHIS/ followed by the year | Not listed in this job; the fit job lists each year's folder and records the sample-adult and person file names, sizes and hashes |
+| NHIS 2009–2018 documentation | https://ftp.cdc.gov/pub/Health_Statistics/NCHS/Dataset_Documentation/NHIS/ followed by the year, files `samadult_layout.pdf`, `personsx_layout.pdf`, `srvydesc.pdf` | 2009, 2016 and 2018 downloaded |
+| ACS 1-year PUMS person files, 2010–2024 | The Census Bureau's PUMS directory, https://www2.census.gov/programs-surveys/acs/data/pums/ followed by `<year>/1-Year/`; for 2020, the file with experimental weights under https://www2.census.gov/programs-surveys/acs/experimental/2020/data/pums/1-Year/, which the Census Bureau's 2020 experimental-PUMS page links | **Not checked**: no round fetched or listed a data directory. The fit job confirms each URL and records sizes and hashes before use |
+| ACS PUMS documentation, 2010–2024 | Data dictionaries under https://www2.census.gov/programs-surveys/acs/tech_docs/pums/data_dict/ (`PUMSDataDict10.pdf` to `PUMSDataDict12.pdf`, `PUMSDataDict13.txt` to `PUMSDataDict16.txt`, then `PUMS_Data_Dictionary_<year>.txt`); subject definitions under https://www2.census.gov/programs-surveys/acs/tech_docs/subject_definitions/ (`<year>_ACSSubjectDefinitions.pdf`); top-coded values under https://www2.census.gov/programs-surveys/acs/tech_docs/pums/top_coded_values/. The 2020 1-year dictionary sits only under https://www2.census.gov/programs-surveys/acs/experimental/2020/documentation/pums/ (`PUMS_Data_Dictionary_2020.txt`) | Fix round 1 fetched every dictionary but 2020's, every subject-definitions file 2010–2024 and the 2024 top-coded values (§13.6). 2020's dictionary lay outside that round's cleared path; the fit job fetches it first (§8.2) |
+| TR 2026 death probabilities | Already in the repository behind `data/tr2026.death_probability` (S0.1 §1.1) | — |
+
+### 11.7 Invariants for property-based tests (Hypothesis, invented data)
+
+- Printing intervals contain the printed point, widen as the half-width
+  grows, and swapping the two values negates the interval.
+- `q = m / (1 + m/2)` and `q = 1 − exp(−μ)` increase, stay in `[0, 1)`
+  and invert exactly.
+- Every completed table holds each category, sex and age 0–120 exactly
+  once, with zero reference coefficients, and §5's rules reproduce the held
+  values exactly.
+- **Calibration round trip:** invented coefficients and an invented
+  reference population generate targets, and §6.3 recovers the
+  coefficients within `1 × 10⁻⁹`. With `λ = 0` and one dimension varying,
+  JR-C equals JR-A. Permuting rows, or scaling every weight by one
+  constant, changes no coefficient.
+- Each calibrated table passes `RelativeRiskSpec` and SM1's cell
+  conservation on the invented population.
+- `Q̃`'s quintiles hold exactly one fifth of the weight at every sex and
+  single age, ties included; permuting records, inside a tie block or
+  across the file, changes no membership; and the ordered logit, fitted to
+  invented records whose quintile weights equal the model's expected
+  shares under known `λ` and cut points, recovers them within
+  `1 × 10⁻⁶`, converging under §6.2's rule from §6.2's start.
+- When the invented population's attributes are independent, its
+  quintiles hold one fifth each, nothing caps, and the invented source
+  tables come from the same coefficients on the same population, the
+  fitted and target e65 agree exactly for every earnings quintile. For a
+  race and Hispanic-origin group they then differ by the tail alone:
+  `e65_F,g − e65_T,g = l(100) / l(65) × (e100_F,g − e100_T,g)`, with
+  `e100_F,g = life_expectancy(q_F,g[0..120], 100)`, because NCHS's
+  equation 17 and `life_table.py`'s `Lx = lx − 0.5 dx` give the same
+  person-years at 65–99 (`life_table.py:13`; NVSR 75(5) p. 61). The two
+  agree exactly when the invented open-interval ratio `e100_g /
+  e100_total` also equals `e100_F,g / e100_base`. The fitted life tables
+  satisfy `life_table.py`'s invariants.
+- **NCHS-closing identity (M7):** with every invented NCHS ratio equal
+  to one (`ρ_g(a) = 1` at 65–99 and `e100_g = e100_total`), §9.2's race
+  target equals `life_expectancy(q_base(s, ·, t)[0..120], 65)` within
+  `1 × 10⁻¹²`.
+- **NHIS implied value under proportional hazards (M-H1):** when every
+  category's hazard at each age of a band equals a fixed multiple of the
+  reference category's, the age-standardized implied value of §4 equals
+  the log of that multiple for any positive weights `W` and any age mixes
+  of the two categories. A companion example with invented values shows
+  the crude ratio failing the same check when the age mixes differ.
+- The digitizer recovers known values from synthetic vector figures
+  within the plotting tolerance and rejects a figure with inconsistent
+  ticks.
+- The NHIS estimator recovers known hazard ratios from synthetic
+  person-time; the bootstrap never moves a PSU across strata.
+- The suppression writer refuses any artifact with a cell under 100
+  persons or 20 deaths.
+- `WideningCoefficientPath` equals P-PRIMARY through 2024, its multiplier
+  lies in `[1, 1.5]` and stops changing after 2054, and race coefficients
+  never move.
+
+### 11.8 Differential tests
+
+- The two table transcriptions (A4), and Tables 1–6 against Tables 7–12
+  (A2).
+- Newton's solution against a generic root finder on the same equations,
+  on invented inputs.
+- The fit's normalization against SM1's `normalize_probabilities` called
+  directly on the expanded rows.
+- `period_life_table` against `life_expectancy` for every fitted table,
+  both with `f0 = 0.5` (§9.6).
+- §9.2's direct sum for a race and Hispanic-origin target,
+  `(Σ_{a=65..99} L(a) + l(100) × e100_T,g) / l(65)`, against
+  `life_expectancy` on ages 0–100 with the probability at 100 set to
+  `1 / (e100_T,g + 1/2)`. The constant-terminal-`q` closing gives
+  `e_w = (1 − q_w / 2) / q_w = 1 / q_w − 1/2` at the last age
+  (`life_table.py:25`, 170), so the recursion starts from `e100_T,g`; the
+  test draws `e100_T,g ≥ 1/2`, which keeps that probability in `(0, 1]`.
+- The NHIS Poisson estimator against a direct person-year rate ratio on
+  synthetic data with one covariate.
+- The replicate-variance code against a hand formula on synthetic
+  replicates.
+
+## 12. Rules for two boundary cases
+
+Choices inside SM2 are this record's, and it states them. Two cases
+reach past those choices, and the lane's coordinator ruled on both for
+this record.
+
+1. **A failed figure digitization stops SM2; the table-only path doesn't
+   stand in for Q6's primary.** If the digitization fails its validation,
+   SM2 stops and reports, the fit publishes P-TABLE as a labeled
+   sensitivity, and the question goes to Max with the failure's evidence,
+   because using the 2020 tables as the last observed level amends Q6.
+   Nothing waits on this now; SM2's admission waits only if the validation
+   fails.
+2. **SM2's primary never takes its reference covariates from the
+   population of the held-out gate that tests it.** This applies the
+   existing held-out separation and sets no new principle. SM2 stays on
+   public data, so SM3 tests a model fitted without its own population; a
+   reference population built from SM3's population (JR-F) could run only
+   inside SM3's authorized step, as a registered sensitivity. PSID use
+   follows the project's data-access record. If a later
+   calibrate-then-test pair wants a different answer, that call spans
+   gates and goes to Max then. Only JR-F depends on this rule; nothing in
+   the primary path does.
+
+This record amends none of Q4, Q5, Q6 or Q11.
+
+## 13. Files read
+
+### 13.1 Restricted lists
+
+The job read each with `sed -n 'a,bp'` in parts of at most 60 lines, every
+part shown in full, and opened no harness-saved copy of any of them.
+
+| File | SHA-256 | Parts |
+|---|---|---|
+| `EV/RESTRICTED-FILES.md` | `b09a2a56533b6880dd832135a60d315e171fa6850079940d8c1aee795f1a6351` | 13 (lines 1–742) |
+| `EV/coordinator/restricted-files-pending-20261024.md` | `4b4936f05955541d4ad8b09ecf35ff01d73df104b01ae06c3446122e815ab931` | 2 (1–105) |
+| `EV/coordinator/restricted-files-pending-20261025.md` | `9f78e02177b3591c58f7177179fdcb018d4daedf64b909e6dd6824e03e5972ac` | 1 (1–23) |
+| `EV/coordinator/restricted-files-pending-20261025b.md` | `0fb8ba87bd8098062b3a133a2e501c0394ac34716e8ee93f1fb749baabf35a69` | 1 (1–38) |
+| `EV/coordinator/restricted-files-pending-20261025c.md` | `8268e05c8f2e885aa114eb08391db13dc13ea78be503a582b015fda88738774b` | 2 (1–121) |
+| `EV/coordinator/restricted-files-pending-20261029.md` | `c405c38194a3a3535b275d84e458964b9f71ce50fd058339ac57c8c54e025609` | 1 (1–26) |
+
+`EV` is `/Users/maxghenis/microcosm-launch-evidence/dynasim-parity-20260909`.
+No other pending file existed, and all six hashes matched the brief.
+
+### 13.2 Repository files, at `18082859`
+
+| File | SHA-256 | What the job read |
+|---|---|---|
+| `docs/design/mortality_groups_financing.md` | `5ee8d4f5df12c1777503984def7717fbfe8285cb0cf038a15eb9e3fe40051cfe` | Lines 1–152, 304–411 and 467–1022, and its headings |
+| `docs/design/s0_1_source_audit.md` | `486d63c91f95e49b97a1012fb41fe5a6937c94194400bf0e5778bd3835f668fc` | Lines 1–200, 324–707, and its headings |
+| `docs/design/s0_1_power_results.json` | `58569c9667c1c1af0e5d5a2fcc0d0e4cd1ed8002fb44d5fe01370f60f1ac74f3` | Top-level keys only |
+| `docs/design/mortality_ses_sm1.md` | `0000b20134b17df1b522795e742a5557ff8c909b89bfdc955e5ec544315c4990` | Lines 1–100, 273–332 |
+| `scripts/s0_1_power.py` | `c891a1f8391edb593f0d86f6ac9c333e053854347179da5dbafc301b0000679e` | Two lines found by a search for SM2 terms |
+| `src/populace_dynamics/mortality_ses/__init__.py` | `5dfb1958b60144f71a7bb54d8a11848bd37670fb2d82fca873650dcf6bc53dd1` | In full |
+| `src/populace_dynamics/mortality_ses/model.py` | `32dfd76dcd8ef6fe237123c043657289178ca150f4ceede76a92afc3ed9e4cdd` | In full |
+| `src/populace_dynamics/mortality_ses/normalizer.py` | `44e5331a83b63d9a7e5c180af04a3fc4c173da08b90ae8c7d8a88958fe869015` | Lines 1–140 and `normalize_probabilities`'s docstring |
+| `src/populace_dynamics/mortality_ses/specification.py` | `6a698b2a134d106becc2b34b2c5ea9b050b804d202c870333265ddb15503f84f` | In full |
+| `src/populace_dynamics/mortality_ses/step.py` | `08be41603343908df4d6a8d6122aedc4fcdc02fc29865f5c9da9786df24b900d` | In full |
+| `tests/mortality_ses/test_model_step.py` | `259ff37cf36b3fdfebd181fbf448beddd19461f0e5c54825870e4643f2b641b2` | Line count and hash only |
+| `tests/mortality_ses/test_normalizer.py` | `84a4a48710856c5a600c9b50da73be104af8d52f0abb84de14f70e53f2074f25` | Line count and hash only |
+| `tests/mortality_ses/test_properties.py` | `99e881749a3e65ace684038ebdc23d6f8d8fbfb19e76e5b64f467d88d96ad8ee` | Line count and hash only |
+| `tests/mortality_ses/test_specification.py` | `2e736ee32e3dc79ba171b2e41b9d185ef5fac943717853ab04e66b6e1357738a` | Line count and hash only |
+| `src/populace_dynamics/data/life_table.py` | `ab9f23bd155734b219c36f2cade0d1eb926e364ef39b0b91179b9e5ccd5dff92` | In full |
+| `src/populace_dynamics/baselines/base.py` | `52e80c94c0206adcba0b0ff6e78ff1af5d9e12f49d33c76d14051caf29de6fd6` | Lines 1–60, 140–315 |
+| `src/populace_dynamics/engine/di_entitlement_rates.py` | `6000ba7b7039e2359920d62fd7f4c4948d729917a39ef7ad8c28d9f3da3c9ba3` | Lines 1–80, 105–120, 180–215, and a search for its mortality fields |
+| `src/populace_dynamics/estimates/group_breakdown.py` | `8b1821c264be68a00e8f867fb09790ecf52dd6658a129343f4f93171ec939f72` | Lines 1103–1130 and 1227–1274, by one `sed` command. Departure: the hash command also read the whole file's bytes; the job displayed and searched nothing else in it. The hash equals the one S0.1 records at `2195258c` |
+
+Also departing from the brief's list: the same hash command read the
+repository's `CLAUDE.md`
+(`594161147040726b7b36d75814e0336b5fcf601db4b5a23e1a95972005aa78e3`),
+whose text the harness had
+already placed in this job's context; and the job read
+`.git/info/exclude` to confirm that git ignores `.sm2-scratch/`.
+
+### 13.3 S0.1's source folder
+
+`S` is `EV/mortality-groups-financing-20261021/s0_1-sources`. The job opened
+every file below by exact name; each hash matches `S/files.sha256`.
+
+| File | SHA-256 | How read |
+|---|---|---|
+| `MANIFEST.md` | `37b1dbf9c8fe520787dbc69a4accf917b33372b832ab518e94b3dc07515e6892` | In full |
+| `files.sha256` | `a5e5b6171009c62afe4549bcab8c6900cd52af865b041cd160fae1a060d20ed0` | Lines for the files below |
+| `oact_study129.pdf` | `937ec3b6c3109426ba8d87e0956e320bdc29ab6ffa07625ae1b379635d04e414` | Page 27's top lines; image list and content-stream operator counts for pp. 33–38 |
+| `oact_study129.txt` | `c94bebb0f918396f04c990c7f1d495d982f9d7a8503e0db637bd30ee3ac21f90` | Text pp. 1–14 with decimals masked; table and figure layouts masked |
+| `oact_study124.pdf` | `ef66df96b8a21fe59bf4fef79aa48c950d3fff866e6725f4dd5c5f339f10ac38` | Not opened (its text file was) |
+| `oact_study124.txt` | `c8716f9346c6ddacd2f1ee5e35f0657cc5ca28ffe6f51446cb434ce5707aa385` | Text pp. 2–10 masked; table layouts masked |
+| `oact_tbosley_20250114.pdf` | `33a192b5725ef338060044533b7f9c3389dde762ddefdcaa26677bc25d468d46` | First lines of slides 1–4 and 20–23 |
+| `nchs_nvsr74-02.pdf` | `3312964360b9a33663be89248f7e75c39e99f895dd04ca86659115fe1a318421` | Not opened |
+| `nchs_nvsr74-02.txt` | `ccab51d670a1290a69dc07caa8e2dd7138fad6eefca2074bc23089d5063b2707` | Table titles and pages; decimal counts of one table |
+| `nchs_nvsr75-05.pdf` | `43e24a1f249f017fcc505a601eb5738f0797506e35b67c4722c91791777d8842` | Not opened |
+| `nchs_nvsr75-05.txt` | `59a05b2078ca221dd3e04c6490396b9a8b91e24af1f4d8de4868730975797615` | pp. 1–4 masked; table titles; technical notes pp. 53–55 masked |
+| `nchs_nvsr74-04_deaths_final_2022.pdf` | `bb7b356841b202cba6a845315679b42424945a78240414dbe67bd100d428b26b` | Not opened |
+| `nchs_nvsr74-04.txt` | `ec24a46d03bb361e5038659a56c5340ba5d330151d3cc02f2a49b1674bcff1db` | Table 14 (p. 86) and Table VIII (p. 133) masked; technical notes pp. 127, 132–133 masked |
+| `nchs_lmf_public-use-linked-mortality-file-description.pdf` | `2e9cae2ffd17b276d200f9388860e682f995a7210bb004bdba54f63f412e446c` | Not opened |
+| `nchs_lmf_public-use-linked-mortality-file-description.txt` | `f426c84347536f8635fddfef9a72950693cd22ae9fced81cddcc155b11775331` | In full |
+| `nchs_lmf_public-use-linked-mortality-files-data-dictionary.pdf` | `5a62ebb5573030ce448756fc48f0618561d713aa4b48b9b1a2669ff20db7bbd6` | Not opened |
+| `nchs_lmf_public-use-linked-mortality-files-data-dictionary.txt` | `5cdb8645bc99cc224654d11cf993ee23758ac51f0c857757df1483616c9098b5` | In full |
+| `nchs_lmf_public.html` | `0d599231ec9d3450066a9d9537b9a83272746b87336c73e6ea45ca2e8aeb55c3` | Text lines on files and dates |
+| `nhis2018_personsx_layout.pdf` | `186bb42ff139503b2a2263f227ca558dd82fab99a77c25395af66ae0fc366322` | p. 294 |
+| `nhis2018_samadult_layout.pdf` | `9ab1b5345df69c11cff74ee1921abb2d02d3b7e698061c0d08589d00012da9a0` | pp. 6–11 |
+| `nhis2018_srvydesc.pdf` | `62662706d190edd80f7225a4f228169566defe5bf6f3a2a6ae29d3f17c4c1246` | pp. 25–26, 70–72 |
+| `cbo_wp2007-11_cristia.pdf` | `e8f376a6330b320bf778ac50a8831b21fbb7884a9f9b72bea761d720f779524c` | Not opened |
+| `cbo_wp2007-11_cristia.txt` | `199af4eab22c9313bb7cc4e1b4981de7357b93b66b3e3d07a75675e8b7247c46` | Printed pp. 6–12, 14, 18–19 masked; Tables 1, 3, 5, 6 and Figure 2 layouts masked. This list left out Table 4, on which M-C4 relies; fix round 1 read its layout with every number masked and confirmed its race block (§13.6) |
+| `cbo_41681.pdf` | `e22720ad1ca203c5a801e61094c17a01d050b0c57ce929fa7dcb1d09573fb467` | Not opened |
+| `cbo_41681.txt` | `f6d2891687ecb41ac06201e2c0fbd856fe126d0c4ce9376c257875a2cdc973af` | pp. 1–3, lines matching definitions and sources, masked |
+
+"Masked" means a filter replaced decimals and comma-grouped counts before
+display. Some integers still showed: page and year labels, a few
+three-digit counts in Table 14's "not stated" row, and one three-digit
+AIME breakpoint in Study 124's Table 2. None of them appears in this
+record. The first job's report names two more leaks that this paragraph
+left out: the percentage shares in Study 124's Table 3 and the
+percentages in Study 129's text, and the TR 2026 example values in
+`life_table.py`'s docstring. None of those appears in this record
+either.
+
+### 13.4 Documents this job fetched
+
+Thirty-three files from `www.cdc.gov/nchs/` and
+`ftp.cdc.gov/pub/Health_Statistics/NCHS/`, with URLs, retrieval date
+(1 November 2026), sizes and SHA-256 in `.sm2-scratch/sources/MANIFEST.md`,
+which the lane moves out of the clone after the job:
+
+- the NVSR index page, its script and its listing CSV, and the NCHS
+  life-table index page;
+- *Deaths: Final Data* for 2010–2021, and the internet tables for
+  2018–2020;
+- *United States Life Tables* for 2010, 2017, 2018, 2019 and 2023;
+- the public linked-mortality FTP folder listing;
+- the NHIS 2009 and 2016 documentation folder listings, sample-adult and
+  person layouts and survey descriptions.
+
+None is microdata. The job opened no ssa.gov, census.gov or other host.
+
+### 13.5 Files not opened
+
+Three times a slow command ran past its time limit and the harness moved
+it to the background, saving its output under
+`/private/tmp/claude-501/-Users-maxghenis-PolicyEngine--mgf-sm2/…/tasks/`
+(a page-mapping loop over Study 129, a fetch-and-convert loop over the
+NHIS layouts, and a search with a Python check); a fourth time a commit
+and push did. The job stopped the first three and opened none of the four
+saved files.
+
+### 13.6 Fix round 1 (3 November 2026)
+
+A second job revised this record to answer independent review r1
+(`EV/mortality-groups-financing-20261021/builds/sm2-review-r1.md`). It
+counts as exposed, like the first, and nothing from it feeds a blind role.
+
+**Restricted lists.** It read the six lists of §13.1 and a seventh,
+`EV/coordinator/restricted-files-pending-20261101.md`
+(`84ab3e1f4c5824a3fd97f8a742a293737deabc680d674724e7e6f2cf4dfc00f5`, 16
+lines), in parts of at most 30 lines, every part shown in full, and opened
+no harness-saved copy of any. All seven hashes matched its brief, and no
+newer pending file existed.
+
+**Documents fetched.** Thirty-five files on 3 November 2026, from
+`www.census.gov/` and `www2.census.gov/programs-surveys/acs/tech_docs/`
+only; none is microdata. `DD` is
+`https://www2.census.gov/programs-surveys/acs/tech_docs/pums/data_dict/`
+and `SD` is
+`https://www2.census.gov/programs-surveys/acs/tech_docs/subject_definitions/`.
+
+| What | URL | Bytes | SHA-256 |
+|---|---|---:|---|
+| Data-dictionary directory listing | `DD` | 35542 | `6353cd1d26a14086aa6ed2f3edc5fcc1b09e60546de0e19910d00409b5944579` |
+| 2010 data dictionary | `DD` `PUMSDataDict10.pdf` | 369558 | `4b4cc0b1adca2abff28e67cf796b6074e45cdf9c8609b88c5e4a45d84fb8fb48` |
+| 2011 data dictionary | `DD` `PUMSDataDict11.pdf` | 360302 | `a0127a8b40d472a11def2efc6dc3a88b8425b707a6eb63f860002c641c054963` |
+| 2012 data dictionary | `DD` `PUMSDataDict12.pdf` | 369966 | `89570a1831bc5c9ee5a72cd89ebdda978ef1f799f1d1353e418f3e1eeb409b1b` |
+| 2013 data dictionary | `DD` `PUMSDataDict13.txt` | 264904 | `14f0061974fc6a1db746679766abe1c4d81afd8e4ef4b5b83579e64d136f98c2` |
+| 2014 data dictionary | `DD` `PUMSDataDict14.txt` | 190795 | `9dd3f70392159c4fa15b55b437ea77796f877b3ec8973561c9762b697abb1f7c` |
+| 2015 data dictionary | `DD` `PUMSDataDict15.txt` | 207251 | `0b6b75d176dcb8572972d11d303ff1f02a11d1b58ca40702508ab3b834dade53` |
+| 2016 data dictionary | `DD` `PUMSDataDict16.txt` | 211482 | `b21a4dd6500b03fe61de92f5433a3b25b635b07492a9806b51db2d8a49e2b89b` |
+| 2017 data dictionary | `DD` `PUMS_Data_Dictionary_2017.txt` | 294848 | `ea5bcde2a07b0c76ac145fac108e9013480b686a9d70afc4e3f4a6512ccbfd69` |
+| 2018 data dictionary | `DD` `PUMS_Data_Dictionary_2018.txt` | 298670 | `843a18e15b6cb7c7326261d478947a1415c4674c62419a8d01946a5a5df4d5fd` |
+| 2019 data dictionary | `DD` `PUMS_Data_Dictionary_2019.txt` | 300134 | `2ecbd86e979aa4e94945012e7537e53c37a4deff7f0435adb000b98b5a50ce77` |
+| 2021 data dictionary | `DD` `PUMS_Data_Dictionary_2021.txt` | 301277 | `53fb9da5c36b68670c54ab12baa13c4767f6b3e43db326585635c5a333090b60` |
+| 2022 data dictionary | `DD` `PUMS_Data_Dictionary_2022.txt` | 301115 | `4a920e7c2a83dc5bede6620e5edf9bed825a8950072271bbbadd800c400ba9e6` |
+| 2023 data dictionary | `DD` `PUMS_Data_Dictionary_2023.txt` | 303015 | `0e58d8edd725d44035c32ee2697acb3963eb39c36caf4096e0ce494c64bb6dc6` |
+| 2024 data dictionary | `DD` `PUMS_Data_Dictionary_2024.txt` | 306867 | `9a98282277aa7f2dafb12b0bdf38b258689d3a1a9c2abbd9b5e2c1ddde6f7b32` |
+| Subject-definitions directory listing | `SD` | 20306 | `a2e01395460524851da9ce5eee71732f61667e56cbf914cc65d6ecbd17117a82` |
+| 2010 subject definitions | `SD` `2010_ACSSubjectDefinitions.pdf` | 688859 | `dd016f630ab0435e2e3edfe21ec5aed5ca5a24391e1efb7a5f022d0332336012` |
+| 2011 subject definitions | `SD` `2011_ACSSubjectDefinitions.pdf` | 700744 | `7eedd173f80fccf700edf83628b4c0d6ae14b1bcc8b7a41f2c5d59a28366cd5e` |
+| 2012 subject definitions | `SD` `2012_ACSSubjectDefinitions.pdf` | 804765 | `61c7199b24f9ac03b46bfaef968ea152e9d4ab427d43772f05cfc48b842da62f` |
+| 2013 subject definitions | `SD` `2013_ACSSubjectDefinitions.pdf` | 813312 | `bcc860028009411b4b45961a7abc50ecc4eaf2cea3e7676da8bbc8fcc3a98d9f` |
+| 2014 subject definitions | `SD` `2014_ACSSubjectDefinitions.pdf` | 1216961 | `bff06e9d1477147c5171aacb452954a554f21d465bbe4e96322dd03c7ca7a048` |
+| 2015 subject definitions | `SD` `2015_ACSSubjectDefinitions.pdf` | 851793 | `881329f214c54e4249165c271107d599b906a96561e9415739ad08f487ffef7d` |
+| 2016 subject definitions | `SD` `2016_ACSSubjectDefinitions.pdf` | 1422620 | `028829c7dc01d5575bd9ecb4bf373f559ad1590be7599ce2de448d66a583203f` |
+| 2017 subject definitions | `SD` `2017_ACSSubjectDefinitions.pdf` | 932466 | `8676e64c48a8519da8648764ad7b326603f12fa1c8a9e0fba9b6965a03e85112` |
+| 2018 subject definitions | `SD` `2018_ACSSubjectDefinitions.pdf` | 1653091 | `842c8188751a22208a71360c8474c770236c313e58196356cb998bb0724f260c` |
+| 2019 subject definitions | `SD` `2019_ACSSubjectDefinitions.pdf` | 1076378 | `ce42270355d9dce11fe8608454402fb80b9b0ce4a4e83ebd55a71f8a117574ed` |
+| 2020 subject definitions | `SD` `2020_ACSSubjectDefinitions.pdf` | 1091755 | `70082ff6739486fa9e7524c34507f59a292d05afcb5fc9f3402a7985ecde5bc5` |
+| 2021 subject definitions | `SD` `2021_ACSSubjectDefinitions.pdf` | 1088859 | `a1db9fb258ff6ff25c5ac4d9a4817eca35f31bb49e238aaad616e6d6f568398e` |
+| 2022 subject definitions | `SD` `2022_ACSSubjectDefinitions.pdf` | 1136140 | `3ba434d7c72ae71c6c8fe4903493cc063a92eca488b89e2e1e490124378caeed` |
+| 2023 subject definitions | `SD` `2023_ACSSubjectDefinitions.pdf` | 1102515 | `6edcfa27b10eaea093867256523cdc6f256a6c7991f4b02823d9034de2e31033` |
+| 2024 subject definitions | `SD` `2024_ACSSubjectDefinitions.pdf` | 1170567 | `6e0c5b1a5d3ebf7e300e583ac6424cbd5f09de50d69e7ef03aa6d070d9c4a7a8` |
+| 2016–2020 PUMS documentation page | https://www.census.gov/programs-surveys/acs/microdata/documentation.2020.html | 368498 | `0893dd7309bb7f88026e3ae23ca8011f127178a8c281c85b6a99a0a6f0bf59c8` |
+| 2020 experimental-PUMS page | https://www.census.gov/programs-surveys/acs/data/experimental-data/2020-1-year-pums.html | 351006 | `3fc562a7536230843d11fc04dd01b491b73772b0fdb99726f0feb1955212e7b1` |
+| 2024 PUMS documentation page | https://www.census.gov/programs-surveys/acs/microdata/documentation.2024.html | 380383 | `b84b42278085eeb4731df9442200a02f207f9f1dd14123b30ee202c54b5fea05` |
+| 2024 top-coded values | https://www2.census.gov/programs-surveys/acs/tech_docs/pums/top_coded_values/2024PUMS_top_coded_values.pdf | 155706 | `d5ff127b7cfa33a8bbea84590472a0ae5806054762ac246471648f38b56196a7` |
+
+From the dictionaries the round read the entries for `SSP`, `SCHL`,
+`HISP`, `RAC1P`, `AGEP`, `ADJINC` and `SEX`; from the subject definitions,
+the Social Security income definition; from the top-coded-values
+document, its variable list and its explanation of top-coding; from the
+pages, their links and the 2020 page's note on experimental weights. The
+2020 1-year dictionary sits under
+`www2.census.gov/programs-surveys/acs/experimental/`, outside the round's
+cleared path, so it fetched nothing there.
+
+**Other files opened outside the repository.** The review r1 file; from
+the review's scratch folder, `m1_invented.py` and `schema_check.py`; the
+first job's brief, report and OACT counts check; in S0.1's source folder,
+`files.sha256` (the lines for the files below), `oact_study129.txt`
+(lines on pp. 2–9, masked), `nchs_nvsr75-05.txt` (pp. 1–2, the table
+notes and technical notes pp. 53–62, masked), `nchs_nvsr75-05.pdf`
+(the text of pp. 57 and 59, masked),
+`nchs_nvsr74-02.txt` (a search for classification ratios),
+`nchs_nvsr74-04.txt` (p. 132, masked),
+`nchs_lmf_public-use-linked-mortality-file-description.txt` (p. 1),
+`nchs_lmf_public-use-linked-mortality-files-data-dictionary.txt` (a
+search for a perturbation flag), `nhis2018_samadult_layout.pdf` (pp. 6–8),
+`nhis2018_personsx_layout.pdf` (p. 294, the `EDUC1` codes),
+`nhis2018_srvydesc.pdf` (converted to text in the round's scratch folder;
+search hits on pp. 8, 16, 70 and 82–94, and pp. 83, 92 and 93 read),
+`cbo_41681.txt` (pp. 1–2, masked),
+`cbo_wp2007-11_cristia.txt` (table titles, and Table 2's layout with
+every digit masked) and
+`cbo_wp2007-11_cristia.pdf` (Tables 1, 2 and 4, PDF pp. 36, 37 and 39,
+every number masked); in the SM2 source folder, `files.sha256`,
+`nchs_us_life_tables_2019_nvsr70-19.txt` and
+`nchs_us_life_tables_2023_nvsr74-06.txt` (searches for classification
+ratios). Each hash matched its folder's `files.sha256`.
+
+**Repository files.** This record; `docs/design/s0_1_source_audit.md`
+(lines 1–60 and 324–707); `docs/design/mortality_groups_financing.md`
+(the lines this record cites, and a search for its mentions of the
+closing rule, which found lines 143 and 1146); `src/populace_dynamics/data/life_table.py`
+(in full); `src/populace_dynamics/mortality_ses/specification.py` (a
+search for the columns it reads, lines 203–224); and `pyproject.toml`
+(lines 118–130, the pytest settings, read before the test run below, a
+file outside the brief's read list).
+
+**What showed despite masking.** A few one- and two-digit counts from
+Table I of NVSR 70(19) and 74(6); number words and one axis label in CBO
+41681's prose on pp. 1–2; in the ACS dictionaries, the `ADJINC` factors
+and the dollar bounds of the income items; and the TR 2026 example values
+in `life_table.py`'s docstring, which the round read in full; and one
+count in NVSR 75(5)'s prose on p. 2, which a search for the life-table
+column headers displayed. None of them appears in this record.
+
+**Code the round ran.** Edit scripts on this file, and one check in its
+scratch folder that ran `life_table.py` on invented life tables to test
+§9.2's formulas (the direct sum against the seeded recursion, the
+identity with every ratio at one, and the tail term of §11.7). It ran
+nothing on real data and computed no outcome. It also ran
+`tests/mortality_ses` serially with `-p no:cacheprovider`: 116 passed.
+Those tests import repository modules beyond SM1's, though not
+`estimates/group_breakdown.py`, and the run displayed none of their
+content.