# Portable synthesis archive

Run from this directory with Python 3.12 or later:

    python 07_analysis_code/replay_whole_synthesis_bundle.py --tier verify
    python -m pip install -r requirements.txt
    python 07_analysis_code/replay_whole_synthesis_bundle.py --tier headline
    python 07_analysis_code/replay_whole_synthesis_bundle.py --tier native
    python 07_analysis_code/replay_whole_synthesis_bundle.py --tier diagnostics
    python 07_analysis_code/replay_whole_synthesis_bundle.py --tier portable-substate

Verify uses the Python standard library. Headline reintegrates the saved original-likelihood grid and checks every reported probability, interval and Bayes factor in the completion receipt. Native additionally refits every reference-bias literature/local family from the archived contrast inputs and checks its results. Counts are determined by the supplied files, not hard-coded. The compact archive also includes all held derived source tables and nested result files used for the explicit native-unit register mappings, in their original bytes. Their full source rows or JSON pointers, specified units, uncertainty types and overlap constraints appear in held_native_diagnostics.csv; they are not added to the shared-mean likelihood.

Diagnostics refits the Montreal and California cohort specifications from the supplied cohort tables, including covariance and calibration alternatives. It does not infer unpublished counts or redigitize a primary figure. The calendar-aligned local estimates, conditional power scenarios, workbook adjudication and state-bootstrap receipts are included for inspection; their raw-panel replay is not promised by this compact tier.

The historical conversion companion includes80 scenarios and their source/repair tables. Impossible inverse point values are unavailable, not clipped zero-risk estimates; a valid interval may still reach zero. The original Excel workbook is not in this compact archive, so its formula/cache audit requires the full held project. The missing-OR repair preserves all risk ratios and finite odds ratios; before and after CSVs and the exact18-field change record are supplied.

The serialized-flag audit supplies its checked Canadian exports, processed panels and UK population cache, with source and implementation hashes. It matches2,016 Canadian source records and preserves the382-row analysis panel. Explicit true/false decoding prevents malformed-input selection errors; the held input flags were already valid, so no estimate changes. UK population subtraction was rejected by the geographic-identity guard. Source audit replay uses the full held-project paths and original metadata bytes; this portable package does not promise a fresh WHO mortality extraction or a full synthetic-control rerun.

The persistent workbook manifest names all seven historical workbooks and294 sheets even after review classifications change. Its latest row-reconciliation outputs and the ten-row Direct Individual companion are included. The latter distinguishes an OR, an HR and a historical mixed-measure summary from risk ratios and adds no independent evidence. The original workbooks remain outside this compact archive; validating their bytes, sheet names and cached cells requires those held originals.

The Numerical Ledger and Manuscript Table companion traces24 displayed rows to12 historical results through132 source-field checks,132 cross-display checks and20 native-result checks. Withdrawn PCV mappings remain unavailable as usable individual estimates even when an old numerical conversion is retained in a source file. The ADDM six-site uptake SD is not the statewide SD. These display audits add no independent likelihoods. The NHIS summary companion also traces28 rows and532 fields to its parent estimate table, preserving adjusted and unweighted sensitivity records.

Three historical register dashboards have199 linked-cell, held-source and inventory-count comparisons. One cached p=0 is an empty-source-reference error; its companion value is unavailable. The original workbooks are preserved, not regenerated. The generator uses exact family identities and fails on missing or duplicate matches, including the later replacement of the historical mixed-measure literature pool. An18-case in-memory formula-engine receipt and its exact JavaScript sources are included. Re-executing that test requires the bundled Artifact Tool environment; the compact Python tiers verify supplied bytes but do not execute the spreadsheet engine. The main historical summary has268 single-SUMIFS lookups that each match one finite numeric source token. Those tokens are stored as text in the original workbook; this numerical crosscheck is not a native Excel recalculation certification. The176 formulas outside that initial screen are covered by the extension described below.

The extended main-summary companion independently calculates all444 formulas, including the176 outside the single-result lookup screen. It explicitly decodes numeric strings and Boolean flags and rejects unsupported grammar, missing aggregate values, empty source matches, misaligned ranges and circular dependencies. This is bounded source arithmetic, not native Excel recalculation. Another companion traces42 literal numerical displays and all634 dictionary entries (4,438 fields) to held sources. The schedule summary omitted nine events without an exact year; they remain in the source and are explicitly listed in the companion, not treated as zero effects. Original workbooks are unchanged and are not included in the compact archive. Narrative claims and native-engine numeric-text behavior retain separate audit limits.

The two power/acquisition summary companions include16 formula checks and122 source comparisons. Saved model-native power flags are not current causal identification or individual-risk power. Eighty retained Swedish acquisition records reconcile the original799 payload hashes with the720-hash pre-restoration base. Analysis3500 exactly explains the historical805-versus809 count:keep-last hash/path deduplication discarded counted aliases for four NJ discipline workbooks totaling64,884 bytes. All four remain in the current inventory. The corrected generator preserves distinct URLs and counts each eligible hash once. Raw checks in the full project verify those four workbooks and ten NCES archives; compact replay does not reacquire them. Five guide sheets have187 row dispositions and148 source/claim checks. Three guides incorrectly say three report-year clusters; the source and corrected companion say thirteen clusters for three vaccine additions. The dated generator is corrected, but original workbook snapshots are preserved. These audits add no independent likelihoods.

The acquisition-restoration companion preserves the original757-row CSV and the corrected837-row table, plus a byte/format-checked80-file restoration receipt. All original base fields are unchanged. The current inventory now contains799 unique counted hashes, including the79-hash Swedish extension. Rebuilds retain this reviewed extension and the old directory-scanning entrypoint no longer promotes every discovered file to valid data. Four Swedish coverage CSVs use explicitly specified cp1252. Fifty-six JSON files decode, eight XLSX containers pass CRC checks, and twelve older XLS files pass only an OLE signature check; their worksheet cells are not certified by this check. The receipt separately lists258 later files absent from the legacy table, including129 retrieval sidecars. The3481 crosswalk identifies114 already-consumed outcome series,3 other executed inputs,1 manual anchor source,9 metadata/query files and2 contextual reports. Their raw payloads are not included in this compact package, so full source replay requires the held project. The nine overlapping missing-component fits and108 posterior summaries are included as input/output JSON and source-code companions; the diagnostics tier refits its nine models from supplied cells and reproduces all108 summaries. They add no independent synthesis weight. The omitted-row zero convention remains unverified, and diagnosis sums are not unique-patient outcomes.

This is not full raw-source replay. The likelihood grids have already been profiled/fitted from underlying cells; these commands do not reconstruct those likelihoods. The replay selects numerical functions from the archived analysis scripts using Python's abstract syntax tree, rather than executing their original acquisition/full-project entry points. Manifest verification hashes all supplied inputs and expected outputs before these functions independently recompute the numerical results. Historical register rows are an evidence inventory, not independent observations or a common estimand. Native family fits are distinct targets and must not be pooled again as independent evidence. The headline is an exponentiated shared mean under specified assumptions, not a universal individual risk ratio. Article PDFs, restricted data, and full historical raw-data archives are deliberately excluded.

Primary article acquisition URLs and locators appear in PRIMARY_SOURCE_ACQUISITION.csv and the literature disposition table; held-snapshot hashes appear in the reproduction manifest. Obtain those materials from the publisher or indicated public source subject to its terms; do not redistribute copyrighted PDFs. A fresh publisher download can differ in bytes (e.g., access-date watermarks). Original analysis scripts are archived for inspection, but entry points requiring excluded raw files are not runnable in this compact package. The listed tiers are the supported portable commands. Machine-specific root prefixes in text metadata are normalized to relative paths; the bundle manifest hashes these delivered files, while execution receipts retain the original execution hashes. Report links to historical files outside this compact archive are not supplied here.

To reacquire the five primary documents used in the Jain/Lee and Taylor/Farrington source audits, run `python 07_analysis_code/acquire_synthesis_primary_sources.py`. It never overwrites an existing source, verifies file signatures and compares hashes against the 2 October 2026 UTC snapshots. Hash differences require source review and may reflect publisher watermarks. Taylor's article is contained on PDF pages 302–305 of the ERIC collection; the command verifies the complete collection, not a newly extracted subset. This command does not acquire every historical paper or raw dataset. The Korean supplement's 13 tables, 288 ratio records and main/supplement discrepancy crosscheck are included as extracted CSVs, with shared-participant and repeated-result identifiers preserved. The 30 North Thames case-series and age-at-diagnosis contrasts share one participant series and are retained individually, not pooled as independent studies.

Open project_results/whole_project_synthesis_3377/index.html for the result register and interpretation. The archive preserves relative directory paths. Generated verification/replay output goes into replay_outputs and is not part of the fixed manifest.

The diagnostics tier reproduces three Swedish interaction batteries from supplied outcome, population-component and exposure snapshots: historical geography, reporting geography and omission of calendar-period proxies. It does not re-extract the excluded original county PDFs or API exports. Seventy-two pairwise fits,18 recodings,900 permutation draws and600 recovery draws are overlapping diagnostic calculations, not72 independent studies. Six restricted common-effect summaries are reintegrated from a supplied likelihood; their low effective sample size remains flagged, and this does not regenerate the original importance-sampling caches. The new Orebro primary report is linked in its metadata; reacquire it with `python 07_analysis_code/sweden_hepb_timing_sources_3486.py --acquire-primary --extract`. The full source-audit and analysis entry points still require the complete project, not just this compact bundle.

The diagnostics tier also refits48 Swedish control-bias sensitivity models from supplied control rows. Three affected controls are replaced and three become unavailable under complete-component selection. These fits are bias-distribution diagnostics, not a new joint autism result. The original15-context control calibration and the17-context headline have distinct scopes. Full joint control-calibrated refits require the original likelihood caches; any supplied likelihood-only replay is explicitly labeled as such.

The Swedish LSS companions retain the corrected municipality panel, nine municipal result tables per input version, the complete synthetic-control results, 48 likelihood summaries and a 92-row current-register overlay. The source audit compares 60,610 fields across 19 original workbooks. The affected municipalities remain outside the 205-municipality complete-record sample, so the balanced likelihood and synthetic-control results are unchanged apart from likelihood roundoff; the broader municipal coefficients do change. Eight pretrend tests reject. These service-use diagnostics are not new autism observations or individual-risk bounds. The compact tiers verify these companion files by hash but do not execute their source-workbook extraction or full municipal/synthetic-control fits. Those entry points require the original XLS/XLSX workbooks and full project; Python xlrd, openpyxl, matplotlib and cvxpy are additional full-project dependencies.

An auxiliary source check reproduces76 national age totals,437 recipient-age weights and4,017 municipal Kolada rate records. Kolada uses the same underlying Socialstyrelsen source and is not independent replication. There are295 Kolada rates without usable workbook rates, including two invalid-reporting municipality-years excluded from analysis. Three fresh workbook/API comparisons reproduce held discrepancies. Publisher quality declarations establish2020–2022 nonresponse scope; Tranås2022 workbook zero placeholders are unavailable, not measured zeros. The archive contains120 overlapping workbook/Kolada fits,240 event-study rows,20 assignment diagnostics and168 annual descriptive summaries on available and matched samples. These are native service-use sensitivities, not autism or individual-risk effects. Fits use2012–2019 windows;2020–2025 rates enter descriptions only. The compact tiers hash-verify these files; full3493 source extraction requires the excluded PDFs/XLS/XLSX, and3494 fits require the supplied panels plus its full dependencies. Use `python 07_analysis_code/sweden_lss_source_context_3493.py --acquire` in the full project to retrieve the nine source checks. Unexplained differences in dissemination remain.

The diagnostics tier also reproduces the municipal/group/national disposition of53,952 held Kolada records, including49,935 aggregates. Current metadata identifies45,864 group records;4,060 records have unavailable group membership and11 are national totals. Those unresolved memberships are not inferred. The N28891 metadata reports no organizational-unit data. Group means are checked against current members as a dissemination diagnostic, not certified historical membership. These rows add no independent likelihoods. Two additional saved control-calibration likelihoods supply12 more residual-bias summaries, for18 total. The1,048,576-draw-per-scramble sensitivity and the coarser quadrature check remain Monte Carlo-limited; the compact replay reintegrates the saved likelihoods but does not regenerate their large core caches.

Pennsylvania diagnostics replay486 model terms per input version (historical and source-complete), including444 PennData specifications,36 federal specifications and six state/federal interaction terms. They use supplied model-ready inputs, not fresh worksheet extraction. All2,412 county rates reconcile with archived count/enrollment cells. The current register links127 repeated displays to their corrected parent fits;18 conditional power scenarios retain their explicit13-cohort assumption, separate from17-cohort source scenarios. These are not demonstrated individual-risk bounds. An independent age-analysis extraction already contains the same804 county-years and is numerically unchanged. Original historical tables remain distinguishable from corrected columns. Retrieve the twelve county workbooks using `python 07_analysis_code/acquire_pennsylvania_county_sources.py`; its source URLs and snapshot hashes come from two retained acquisition ledgers. Downloads never overwrite existing files. A changed publisher hash requires review, not silent substitution.

Independent retrieval of eight New Jersey URLs subsequently found a separate basename-cache error: four2021–2022 discipline files had been attributed the held2022–2023 file hashes. The source register supplies the correct eight URLs and hashes. Analysis3501 extracts1,456 cells, preserving192 suppressed cells and476 genuine zeros. The four reporting margins overlap, and neither year supplies a vaccine effect estimate. Descriptive disciplinary-removal counts are not autism diagnoses or unique-child counts. The compact diagnostics recheck the supplied cell ledger and totals; fresh primary-workbook extraction requires the full project sources. The NJ downloader now uses URL-specific filenames.

The URL/path census covers2,231 top-level CSV headers and150 selected tables. It retains9,190 provenance relationships and37 unsupported-schema tables for review. Twelve multi-URL paths comprise the four NJ collisions, seven catalog/direct-download references, and one Virginia four-page overwrite. The Virginia source-index downloader uses URL-specific names. Four current VDH pages were retrieved; three archived Tableau workbook attempts returned404, which is not proof that no alternative source exists. Current2025–26 SIS coverage bands are not exact percentages, and suppressed/nonreported values are not zeros. These source checks add no independent likelihood. Raw HTML and XLSX acquisition remain outside the compact replay.

The corrected current subnational inventory has837 source records and803 eligible payload hashes. Four earlier-year NJ URL records now point to their independently retrieved bytes. The historical300 table remains unchanged for replay. The37 unsupported field schemas have source-specific dispositions, including parent-key joins for Dutch, Nigerian and Costa Rican references. Search pages, readiness flags, profile fields and analysis references do not become new outcome observations. Compact replay verifies the supplied ledger relationships, not all original source cells or new network acquisitions.

The Ohio school-join companion preserves unknown enrollment for two school-years. Four previously zero aggregate-weight cells are unavailable; every finite coverage value and all outcome values are unchanged. Its source panel and historical join are supplied for portable reconstruction. This is a missing-data handling repair, not an additional effect estimate.

The UAE companion includes its 129 public source payloads: thirteen dataset-detail pages, three search pages and 113 pagination responses. Diagnostics reconstruct all 608 held preview rows, compare all 828 vaccine cells, recover ten older disability records, and refit the three historical and three corrected descriptive screens. Two malformed HTML metadata resources are retained as partial previews; four later metadata pages are not held. Dose details, subtotals and grand totals remain separately labeled. Only details enter corrected analytical sums. The positive nominal MMR–rehabilitation association is retained; these log-change slopes are not autism risk ratios or new independent causal evidence. No network request, form submission or original-output overwrite is needed for this replay.

The UAE official-source extension preserves all48 attempted resource URLs and four records with no download link. Forty-four HTTP200 responses contain the same visible404 page, not new data; four legacy requests failed. Two official replacement pages yield161 administrative context records, eleven center-stock categories, two student composition shares and one2024 supplementary MMR campaign record. The English and Arabic center charts disagree (semi-government45 versus40; sector totals120 versus115; geography total114). All values are retained without imputation. The campaign dose count is an inequality, greater than89,000, not an exact child count or coverage rate. Three EHS PDF sources remain unacquired locally. Compact diagnostics re-extract these held pages without network access. None adds an independent vaccine-effect estimate.

Colorado source diagnostics re-extract 273,656 cells from five held workbooks. They retain 4,808 ambiguously year-labeled cells without assigning years by position, recover two missing unit codes by unique exact names, and verify an unlabeled total against other sheets. Twenty-two percentage/count discrepancies remain distinct from eight differences consistent with numerical precision. The state/federal comparison retains 3,258 relationships without independent likelihood weight. Fourteen Rhode Island encrypted paths resolve to six distinct files and remain unavailable, not zero. XLS/XLSX readers are included in the compact requirements because these source replays need them. No network access or password guessing is used in diagnostics.

Portable substate replay checks the hash-pinned 3515–3529 saved companions, validates source-keyed current-register mappings including the immutable historical 369 row/outcome link, and refits one representative Pennsylvania Poisson model from its supplied corrected matched panel. Current Michigan3530v2 saved-result verification is included. The tier does not re-extract raw IDEA, state workbook, or Michigan source files. The149MB source-token ledger and raw source payloads are excluded. PORTABLE_SUBSTATE_3531.json separates delivered-file hashes from original execution pins and lists omitted source paths and retrieval instructions. Full raw replay is not supported by this compact tier.

The current 3520 category replay is separately hash-verified from its supplied v3 receipt:72 attempts,50 fitted, with all30 current corrected402 paired-Pennsylvania fields matching the adapter replay. Its32 changes versus replay-v2 are restricted to paired Pennsylvania and four paired pooled summaries. The older 369 comparison remains unresolved and is not a pass/fail gate. This bounded verification adds no independent likelihood weight and does not rescan raw sources.

The portable-substate tier also checks Louisiana exact-repeat sensitivity and Oregon missing-count bounds. Louisiana retains both scenarios for six outcome/window combinations, but only the six dependent sensitivity rows are added to the register. Four estimates are finite and unchanged; two are unavailable. The Oregon companion independently reconciles all79 state rows:17 groups in2025–26 have wider count and coverage upper bounds, with unchanged lower/midpoint and district values. These compact verifiers read the supplied receipts, tables and pinned analysis code. They do not rebuild excluded raw school files, IDEA files or the NCES archive. Full raw-pin checks are available from the complete project through the companion modules.

The Oregon archived-school companion verifies340 dependent model-attempt records (242 estimated,98 unavailable), five source-frame/vintage summaries and136 conditional power scenarios. Power scenarios are diagnostics, not additional effects. The within-exposure standard deviation is normalized by total model weight; it does not change when every enrollment weight is rescaled. The five autism rows repeated in the broader-outcome table are explicitly linked duplicates. Compact verification checks the frozen005 receipt, saved coefficients, power arithmetic, source locators and native154 parity; it does not reconstruct the omitted large school-source panels.

The Texas county companion verifies32 dependent current fits,16 paired changes and separate historical/current parity checks. It distinguishes canonical county joins, exact-join diagnostics and calendar-alignment sensitivities. Its estimates concern autism as a share of recorded disability categories, not individual autism risk. The added coverage year and county-name repairs do not create independent studies. Raw IDEA, NCES and vaccination workbooks are excluded from compact verification and listed explicitly by the verifier; full raw-pin checks require the complete held project.

The shared-mean support companion recomputes54 nominal likelihood-ratio support sets from both supplied compact likelihood grids. Numerical scrambles estimate the same integral and are not independent studies. Shared normal bias is convolved before profiling the six saved heterogeneity settings. These sets remove only the top-level mean prior: internal nuisance assumptions remain. They are not coverage-calibrated95% confidence intervals, individual risk-ratio bounds or simultaneous bounds on every vaccine effect. The grid maximum is quantized. This diagnostic supplies no additional likelihood weight.

The South African companion retains214 exploratory service-use fits and12 short-panel unavailable results. Native financial-year district joins, missing versus zero observations, administrative vaccination ratios above100, few time clusters, timing variants and multiplicity adjustments remain explicit. Visits are contacts rather than people, and the screening slopes are not individual risks or autism effects. Compact verification uses supplied panels and source contracts; omitted raw workbook and large cache records are listed by the verifier.

The six-state composition summary adds one canonical augmented Texas result to five source-pinned state Autism/EC-disability gradients. Its baseline Texas alternative, leave-one-state-out summaries and assumed-covariance calculations are dependent diagnostics, not additional effect records or shared-mean likelihood blocks. These native composition ratios are not individual autism risks; disjoint states can still share diagnostic and reporting shocks.

The Texas complete-window companion distinguishes14 fits requiring all three exposure-feed years from14 matched-support fits. Fourteen available-window reference replays and seven paired coefficient differences are audit records, not additional estimates. The common complete-window sample has identical baseline and augmented exposures; the added vintage expands support. Recorded panel-key totals must not be substituted for fitted sample sizes after outcome eligibility gates.

The Oregon preschool companion reconstructs 18 exploratory attempts from supplied regional source tables and independently recomputes their WLS slopes, CR1 covariance and t-reference intervals, plus 72 leave-one-area-out diagnostics. Sixteen attempts yield estimates; two lack enough waves. Outcomes are autism shares within reported EI/ECSE disability counts, not population autism rates. Capture-inferred and explicitly dated exposure vintages remain separate, as do MMR first-dose and unspecified-dose labels. ALL_REQUIRED is a schedule-coverage composite, not a dose-load measure. Per-10-percentage-point slopes extend far beyond observed within-panel variation; per-one-within-SD contrasts are rescalings of the same slopes. Nine regions and two or three waves do not provide calibrated causal inference. The 54 lag-feasibility records are not extra estimates. Original site-level sources and their raw aggregation are outside the compact replay; all omitted pins remain listed. No independent likelihood weight is added.

The twelve-family source follow-up companion verifies supplied extraction tables, reconciliation records and corrected metadata against its frozen source contracts. It includes Taiwan classification and age-table checks, an institutional sex-label correction, Colorado vintage checks, Swedish source limitations, Virginia and UAE source recoveries, nested-source provenance, and Nigeria reporting-metadata correction. These records do not add effect estimates or independent likelihood weight. Compact verification is not a replay of omitted original PDFs, workbooks or large archives; omitted pins and check-specific limits remain explicit. Reported services, table cells, administrative stocks and overlapping extracts must not be interpreted as additional people or independent studies.

The five-family source-extension companion checks95 descriptive Taiwan age-specific stock rates and their corrected receipt bindings, indexed UAE service-count arithmetic,42 Costa Rican coverage cells with unresolved dates/geographies preserved, a pinned prior594-file Brazilian production-reparse receipt, and three manually transcribed Virginia browser observations. Its verification distinguishes supplied arithmetic from raw-source interpretation and lists omitted source files. It does not independently decode Brazil archives, visually assign unverified UAE chart labels, recover missing Taiwan outcomes, or treat public-browser observations as an independent collection. None of these source records is a new causal estimate or independent likelihood.

The acquired-source audit checks five bounded source families: historical narrative/numeric-text dispositions, a frozen public workbook, South Carolina age/count cells, Connecticut EdSight measurements and margins, and prior BC/Norway duplicate-comparison receipts. The historical register reference is an explicitly archived snapshot, not the current register. Count records, source cells and receipt claims are not independent studies. Compact verification performs supplied-table arithmetic and hash checks; it neither reparses the omitted original workbooks/large register nor recalculates native Excel.

The dependent age-extension companion retains 32 ecological specification attempts: 30 point coefficients and two unavailable rank/saturation cases. Its 1,120 shared-exposure diagnostics are not additional effects. South Carolina age/cohort composition and resident-rate denominators remain separate, as do assumed count dates and contextual infant-policy or eligibility scenarios. Non-positive-semidefinite joint covariance makes primary joint intervals unavailable; state/year-only or exposure-only uncertainty is labeled diagnostic. Portable verification checks saved arithmetic, source-bound reviews, exact sample keys and compact hashes without new estimator fits, NIS microdata extraction or independent likelihood weight.

The source-handling closure checks five final bounded audit families: resident versus civilian age denominators, source-header contracts, NDIS snapshot scope, selected New Jersey workbook cells and selected New York source vintages. It preserves unavailable source cells, geographic and universe inconsistencies, age/date ambiguity and named omitted originals. Full verification mode can hash specified files or bounded source prefixes but does not reparse their source arrays or fit models. These source-handling records do not become effect estimates or new independent evidence.

The dependent-diagnostics companion verifies 300 native parameter-attempt records, not 300 independent studies: New York reported-count calendar/source/suppression sensitivity parameters, older IDEA age-band contextual-policy slopes, and Thompson appendix diagnostics. It retains unavailable rank/sample coefficients and unavailable non-PSD joint intervals rather than replacing them with zero or a state-only primary interval. Native log-count or contextual scales must not be relabeled as individual risk ratios. Its compact command performs supplied-table arithmetic and frozen-hash verification; original HTML/workbook interpretation, model-row source reconstruction, exposure arrays and estimator refitting are outside this command. Every dependent row has zero additional shared-mean likelihood weight.

For omitted original inputs, use each companion contract's exact source paths, publisher URLs and retrieval sidecars in the complete held project. Preserve the specified school/count year, geography, universe and snapshot hash; changed download bytes require source review rather than automatic replacement. Public acquisition routines are opt-in and outside the supported compact verification tiers. Article PDFs, restricted sources, the large historical source-token register and non-compact exposure arrays are not redistributed. No email, form submission, account action or network acquisition is performed by portable-substate. A staged or missing required final companion group fails closed and cannot silently disappear from a release.