# One prescribed changed-burden/results verification Status: COMPLETE — no material findings; main disposition remains separate. This is the same review's input and actual-results phases, not a new historical audit or a release authorization. ## Role and limits Model-performed, not human or external audit. My resumed context includes prior upstream instrument-related coding/packet work and technical assembly, including Circeii's ordinary comparator assembly. I did not author main's current matrix/integration/execution work checked here. This does not independently revalidate my own upstream historical work. No accepted coding journals/source qualifications were reopened. No source/PDF research, new agents, full-suite refit, shared writes, corpus/state integration, commit, promotion or deployment was performed. Authenticated source bindings are not source-body reading. ## Input burden: pass [INPUT_CHECKS.json]() records199 file bindings and independent checks, not a replay of main's PASS: - Frozen675 → completed676 is exactly13 retirements (including the Qin parent) and14 new full60-dimensional typed/integer rows:840 cells. - The1,876 compelled retained cells comprise927 actual changes and949 confirmations. All37,844 untouched retained cells preserve exact score spelling and U/NA, including100 fractions. - All255 final instructions were reconstructed against actual datasets; all70 archived-parent profiles preserve their own original scores, U/NA and metadata. Parent/child active rules, excluded holds, and the four active editorial holds remain distinct. - All14 accepted operation/follow-up/source-limit objects survive intact. Actual inherited and archive date/region joins, two aliases, two WEST_ASIA overrides and five date variants match declared controls. Eligible nominal dates:628 baseline,627 successor; no separately measured duration series is invented. - Literal77-group weights cover170 distinct members. All six program weight vectors match their specified real-row controls. Illustrative influence weights do not duplicate rows, recalculate pseudo-node medians or substitute inflated Horn sample sizes. The spreadsheet skill's read-only route informed exact-string and typed-missingness preservation; no workbook or corpus was rewritten. ## Actual completed results: pass [RESULTS_CHECKS.json]() records633 file bindings. All63 run-input bindings,16 numerical dependencies, six committed adapters, execution-state digest, and538 cache identities/payload hashes match. Current adapters equal commit c36ea2cb. The readout and lossless archive equal commit176f21d26a3eccb792c82f71b623c9d9cc03851f; decompressing the archive gives the original139,466,277 bytes exactly. All565 ordered scenario records were independently reconstructed from actual raw inputs. Their536 unique preprocessed matrix identities, observed eigenvalues, weights, eligible-date subsets, reported retention arithmetic and explicit limits match. All565 actual records are computed; all1,130 reported8/9-factor fits are computed with no recorded fit warnings. These are not independent-experiment counts. The check recomputed2,260 loading comparisons using independent QR/SVD and assignment calculations, with4e-6-degree absolute angle tolerance near numerical zero. Every named diagonal Tucker value and every readout row was checked. Selected numerical reproduction, using the bound environment and direct estimators rather than main's runner: - Baseline and successor8/9-factor loadings: four fits, exact numerical equality. - Literal-program weighted9-factor loadings: exact equality. - Successor settler-colonial-studies battery omission: the sixth selected fit reproduces the9-space comparison (23.561310°). - Four500-draw Horn cases: both canonical60-column cases, weighted program case and selected48-column omission; exact thresholds/margins. - All eight canonical date correlations and both specified rank-residual missingness controls, shared-row score correlations and legacy k6 comparison reproduce. A preliminary algebraically equivalent weighted scale calculation differed by2.32e-8 in a loading. Using the bound floating-point arithmetic order eliminates that difference exactly. Verifier-only metadata-roster/NA-parser corrections and this rounding check are disclosed in FINDINGS; they are not input or runner defects. ## What the results support Canonical Horn retention remains9, but this is not a “no change” verdict. Relative to baseline, successor maximum principal angles are18.689067° in8-space and11.784825° in9-space; minimum named Tucker is0.077522. Legacy k6 reassignment is17.975831% across662 shared rows (ARI0.632360), not an editorial-family classification. The actual record distribution is524 Horn9,29 Horn8 and12 Horn7; separate battery omissions retain6–8. Literal-program weighting keeps676 real rows with total weight583 and effective sample size615.777, and shifts8-space44.671536° versus canonical successor. These sensitivity outcomes remain visible. Retained-factor count, named descriptive coordinates and editorial families are different products; neither angle size nor a preferred count validates historical boundaries. No separately coded642-only Egypt or internal Kitakami phase profiles were manufactured. Nominal date sensitivities are not historical movement durations. ## Binding and reproducibility - Acceptance SHA256: `404a93c084fc034f69caa187790ba6e4d12f8103b36e43568c9c2c2615ab8ba9` - Actual RESULTS SHA256: `91cd3d38aaf3bae80a4d7972c935d020bd48c56e4e78d0cada34104ecd5008f9` - Actual MANIFEST SHA256: `e04666c721603cc5111d0421331be9855254c314364000dbc2fed37037d35100` The saved [input checker]() uses standard-library read-only inspection; the saved [results checker]() performs the bounded selected checks above. Both emit JSON only. The latter was run with `tmp/second-pass-venv/bin/python -B`, all six BLAS/OpenMP thread variables set to1 and `PYTHONHASHSEED=0`. It never imports or executes the full-suite runner and never opens the cache for writing. Results were checked again for unchanged bytes at the end. [FINDINGS.json]() is the finite disposition-facing record. Its material-findings and required-corrections lists are empty. This completes the prescribed verification; main, not this review, decides subsequent disposition.