# Independent verification: 2022 upstream value-added study

**Result: numerical reproduction passes, with two disclosed source-accounting failures.** Checked September 23, 2026. This is independent computational verification by a second AI agent, not external human peer review.

## What was independently reproduced

The verifier parses the raw CSV with Python's standard-library CSV reader, builds its own label mappings and foreign-only export denominators, and solves **164 transposed systems**: 81 origin-specific basic-price VA masks, 81 separate TLS masks, a signed source-input-balance mask, and an absolute-gap diagnostic. It does not import, run, or reuse functions from the study's forward-solve analysis script.

For origin `o`, `(I − A).T h_o = 1_o × (VA / x)`. The originating VA share for exporter-industry `k` is `h_o[k]`; multiplying by that cell's independently calculated gross exports yields the numerator. Taxes and source discrepancies remain separate.

## Tested outcomes

| Test | Outcome |
|---|---|
| Archive/member integrity and archived producer hash | PASS. CSV SHA-256 `fbab8fd2727bda9fd8c6cfee8227816db82d33bef6aa4b90194ca5dcbc5ad14b`; matches the archive's March 20, 2026 hash report. ZIP-member CRC and extracted bytes agree. |
| Official documentation | PASS. Independently downloaded ReadMe and annex hashes agree with the analysis copies. Units, VA/TLS distinction and direct-purchase category checked. |
| Source decimal grid | PASS. All 18,388,461 numeric tokens lie on the 0.0001 USD-million grid. |
| Labels and dimensions | PASS. CSV labels match the official ReadMe; 4,050 production nodes, 81 origins, 50 industries, 486 final-demand columns. Numeric spreadsheet IDs are not used. |
| Necessary values and zero-output treatment | PASS. Required data are finite. All 51 zero-output nodes have zero relevant flows, VA and TLS; retained with zero coefficient columns. |
| Signs | PASS as reported, not clipped: no negative intermediate uses; two negative VA entries, 134 negative TLS entries and 189 negative final-demand entries, all in inventories. |
| Nine gross-export denominators | PASS. Independently summed foreign intermediate plus all six final uses, including direct purchases abroad; checked against output minus domestic use within the source rounding envelope. |
| Transposed solves and complete accounting | PASS. Maximum absolute adjoint residual `1.6653e-15`; all active-node VA + TLS + explicit source-balance coefficients reconcile within `1e-10`. |
| All nine reported result rows | PASS. Maximum monetary difference `7.42e-10` USD million; share difference `5.33e-15` percentage points. |
| All 729 reported cell–origin rows | PASS. Maximum monetary difference `2.92e-11` USD million; share difference `2.85e-14` percentage points. |
| Absolute-gap propagation | PASS. Separate adjoint calculation agrees with all nine forward L1 diagnostics within `1e-6` USD million. |
| Actual chart geometry and labels | PASS. Measured 18 comparison bars and 54 accounting segments from SVG coordinates against independent results, using axis-tick scales; tolerance `2e-6` percentage points. Labels, row order, dates and separate-tax/source-balance disclosure agree. |
| PNG readability | PASS on AI visual inspection. Both figures have readable panels, legends and footnotes without observed clipping. |
| Report, policy brief and publication JSON | PASS. Nine-cell values, ranges, rankings, vintage, provenance and bounded policy interpretation checked; dollar-maximum versus relative-maximum wording clarified. |
| Global output-row balance within pure rounding | **FAIL (source data).** 123 rows exceed the conservative `0.22685` USD-million allowance; maximum absolute gap `238.6058` at `USA_L`. |
| Global input-column balance within pure rounding | **FAIL (source data).** 26 columns exceed the conservative `0.20265` allowance; maximum absolute gap `82.0320` at `CHN_L`. |

The last two failures are not numerical disagreement or missing adjustments that the verifier silently repaired. The published matrix was left unchanged. For the nine bundles, signed propagated input-balance contributions range from USD 0.00787 million to USD 0.04458 million; the maximum relative contribution is 0.00003372% of exports. The maximum absolute-gap propagation is 0.00004788%. These are accounting diagnostics, **not** model-error or measurement-error bounds.

The exhaustive [numerical check records](independent-checks.json), [provenance checks](provenance-checks.json), [chart/output checks](output-and-figure-checks.json), [source-gap diagnostics](source-imbalance-diagnostics.json), and hash-bound [verification receipt](verification-receipt.json) preserve the individual pass/fail evidence.

## Confirmed nine-cell primary shares

Basic-price originating VA divided by the exporter's sector gross exports to all foreign destinations; percent, 2022.

| Exporter | C26 China / U.S. | C27 China / U.S. | C28 China / U.S. |
|---|---:|---:|---:|
| Japan | 7.55 / 2.90 | 6.53 / 2.66 | 3.90 / 2.01 |
| Korea | 9.40 / 3.22 | 9.68 / 3.68 | 7.46 / 3.73 |
| Germany | 6.74 / 3.06 | 5.13 / 2.74 | 3.20 / 2.45 |

These are not product-level bottlenecks, current supplier availability, disruption losses, or unique coalition value added. Gross exports can carry upstream value across borders multiple times. No sampling confidence interval is available from this deterministic accounting exercise.

## Source and convention qualifications

The [World Bank archive](https://reproducibility.worldbank.org/catalog/511/study-description) documents its CSV as the January 2026 update of the 2025 regular OECD ICIO table, observation year 2022. Its producer hash confirms the preserved member, not byte identity with the presently hosted OECD ZIP, which could not be downloaded directly during this run. World Bank verification of its own paper does not validate this study.

The [OECD ReadMe](https://webfs-sti.oecd.org/files/STI-PIE/ICIO/2025/ReadMe_ICIO_small.xlsx) distinguishes basic-price VA from TLS. The [2025 TiVA guide](https://stats.oecd.org/wbos/fileview2.aspx?IDFile=2143f34e-6feb-41a9-abaf-cb52132608c4), pp. 6 and 12, uses VA plus intermediate-product taxes in its decomposition. Here the original protocol's VA-only primary is retained, with a labelled tax-inclusive sensitivity. The [OECD release description](https://www.oecd.org/en/data/datasets/inter-country-input-output-tables.html) also distinguishes regular tables from the extended China/Mexico firm splits underlying published GVC indicators. Published TiVA is therefore not an exact comparator for these primary estimates.

## Reproduce the independent numerical check

Python 3.10.19, NumPy 2.2.6 and SciPy 1.15.3 were used. Install the two pinned dependencies in [requirements-verification.txt](requirements-verification.txt), obtain the hash-identified CSV using the study acquisition instructions, and run the standalone verifier from the directory containing the copied verification files:

```sh
OPENBLAS_NUM_THREADS=4 python verify_independent.py --raw /path/to/2022_SML.csv --study /path/to/study --out independent-run
python verify_outputs.py --study /path/to/study --out independent-run
```

The first command intentionally exits `1` for this source file because the two pure-rounding source-balance tests fail. It nevertheless completes the independent computation and all output comparisons; `independent-checks.json` records exactly those two failures. An execution/parser error instead exits `2`. Do not reinterpret either source failure as an all-pass receipt. The second command returns `0` for the checked output versions.

The requirements and scripts are portable; no private workspace paths, credentials or analysis-code imports are required. The receipt identifies the exact reviewed study-file hashes. Later changes require an appropriately scoped recheck.