# MRI-003 — Verification record

## Before implementation · 21 September 2026

The project inventory read 315 supplied non-cache files (266 distinct byte contents, 8,755,388 bytes), including archived duplicates, source/code files, structured data, archive members, workbook XML, the SQLite schema, and the MRI-001 preview image. Git internals, Python caches, and Finder metadata were excluded. MRI-001/002, their evidence conventions, the landing page, and the complete MRI-003 packet were inspected before implementation.

The named `/research/war-of-the-worlds-part-1/` folder was absent. The supplied packet was found at `War_of_the_Worlds_Part_I/`. No relocation or new research acquisition was needed. The site root is not a Git repository; the separate correction-engine repository was not modified.

All 19 entries in the supplied `MANIFEST.json` matched before running the check.

Executed the requested command from that packet directory:

```sh
python code/check_galton.py
```

Runtime: Python 3.10.9. Exit status: 0. All script assertions passed.

| Output | Reproduction result |
| --- | --- |
| `arithmetic_checks.json` | Identical text and values |
| `table_II_check.csv` | Identical values |
| `table_III_margin_check.csv` | Identical values |
| `table_I_check.csv` | All rounded values identical; five unrounded fields differ in final floating-point digits |
| `table_III_literal_implementation_NOT_exact.csv` | All printed, rounded, and residual values identical; 55 unrounded fields differ in final floating-point digits |

Largest unrounded absolute difference: approximately `1.10845e-12` counts. This is a platform-level floating-point difference, not a changed historical finding. The original result bytes were restored against the supplied SHA-256 hashes after comparison. The supplied code, input data, and analysis parameters were not altered. The differences and existing limits were reported before prototype implementation.

## Existing limits retained

- Single-pass illustrative propositions, not complete or double-coded evidence.
- One-analyst numerical transcriptions; no locally bundled original scans.
- No independent reconstruction of the upstream Booth data.
- Table III is expressly not reproduced exactly; header and row-label readings remain distinct in the data and are not used as a headline discrepancy.
- No Figure/Table IV reconstruction, historical person-level dataset, policy effect estimate, or civic-worth score.

## Repeatable implementation checks

Run `python3 mri/003/code/verify.py` from the site root. It checks source preservation, arithmetic in an isolated temporary copy, deterministic artifact generation, unchanged original data, printed blank states, source/claim references, duplicate HTML IDs, and every local page link.

## Browser inspection

Checked the local site in the Codex browser at desktop width and at 390px and 320px:

- All five view links select one view and maintain the selected navigation state. Browser Back returns to the preceding view.
- Contributor selection, proposition selection, and the original-record source panel work. Shaw remains a written contributor, with unknown attendance.
- Mechanism selection and the show/hide names toggle work. The measurement view has no contributor-name leakage while names are hidden; original attributions remain available through the source panel.
- Table I displays the ten preserved counts. Table II selection traces the published allocation (for example F+G: 200 → 200 → 150 to S + 50 to R), with whole-class totals explicitly distinguished from highlighted contributions.
- All five numerical stages work. Table III’s residual count and incomplete-replication label remain visible in Output.
- Native criterion and direction disclosures expand, including keyboard Enter activation. Escape closes the source panel and returns focus to the opening link.
- The final layout has no document-level horizontal overflow at 320px or 390px. A Table II grid overflow found during narrow-stage testing was corrected; its wider table scrolls inside its own keyboard-focusable region.
- The static page loads with zero scripts, five visible views, all 25 original records, all five numerical stages, and six cited source entries. The complete 31-entry source ledger is linked and included in the reusable JSON.
- No browser warning or error messages were captured during the completed checks.

The repeatable verifier passes under Python 3.11.4 as well as the initial Python 3.10.9 arithmetic run. Floating-point differences vary by runtime, while the summary and rounded results remain identical. JavaScript syntax passes `node --check`.

Only two pre-existing site files changed: the root README and the MRI landing page, which now links this local prototype. MRI-001, MRI-002, the book, and all original packet files retain their original bytes. No external sources are downloaded by the page; source links are available for reader inspection. Nothing was deployed or published.

## Focused visual-analysis pass · 21 September 2026

The existing five-view structure is retained. Three seed-linked pathways were added to view 2; the numerical figures in view 3 were refined into a divided normal curve and directly labeled paired bars; one four-step evidence-boundary graphic was added to view 5. No findings or research inputs were changed.

- Each pathway resolves to one original seed record (G01-02, H01-01, W03-01). Its source panel preserves the original coding and qualification. The names toggle also controls pathway attribution.
- Table I has ten probability areas and nine dividers at the supplied probable-error boundaries. Its ten displayed counts match the reproduced rounded values. The finite drawing window affects only the illustration; the counts include the infinite tails.
- Table II displays the exact paired series `[100,150,250,250,150,100]` and `[89,161,250,250,161,89]` on a shared zero baseline and scale. The caption identifies constructed, normalized rates and distinguishes them from individual data or validation of civic worth.
- The concluding sequence marks Table I and Table II arithmetic as checked, while retaining the need for criteria, evidence, and judgment. Its caption explicitly notes that judgment also enters category selection, grouping, and assumptions.
- Inspected the figures at desktop width, 390px, and 320px. Mobile pathways and the concluding process become vertical; no document-level horizontal overflow was found. Table II's existing wider mapping table retains its own scroll region.
- The static version has zero scripts, all five visible views, three pathways, ten class values, six paired comparisons, and four concluding steps. Source access and the names toggle were also checked in the interactive version.

The original packet and `FINDINGS.md` remain byte-identical to the snapshot taken before this pass. The standard verifier and JavaScript syntax check pass. No dependencies, animation, deployment, or publication were added.