# Phase 67 — reconciliation against original issuer filings

## Result

Two purposively selected 2018 issuer-years (Samsung Electronics and Hyundai Motor) match separate-statement accounting amounts and both annual ETR measures. The hypothesis that the raw `tax` column represents cash income taxes paid fails in both cases. This resolves one candidate-mapping error in phase 66; it does not clear the full dataset.

| Check | Samsung 2018 | Hyundai 2018 |
|---|---|---|
| Assets, liabilities, equity, pretax and net income vs separate filing | 5/5 pass | 5/5 pass |
| GETR vs public income tax expense / pretax income | pass | pass |
| CETR vs public cash income taxes paid / pretax income | pass | pass |
| Raw `tax` as public cash taxes paid | fail | fail |

No licensed panel rows are redistributed. The reference JSON contains independently collected public financial facts; the result JSON contains comparison booleans only.

## Sources and scope

- Samsung: [DART receipt 20190401004781](https://englishdart.fss.or.kr/dsbh002/viewer.do?rcpNo=20190401004781), FY2018 column in the separate statements (second set, not consolidated). Public pretax income is KRW44,398,855 million, tax expense 11,583,728 million and cash income taxes paid 10,150,465 million. Derived rates are approximately 26.0902% and 22.8620%. The downloaded page identifies an unamended annual report.
- Hyundai: [DART receipt 20220217001012](https://englishdart.fss.or.kr/dsbh002/viewer.do?rcpNo=20220217001012), FY2018 comparative column of the FY2019 separate statements. Public pretax income is KRW709,948 million, expense 335,060 million and cash taxes 617,659 million; derived rates approximately 47.1950% and 87.0006%. The page identifies an amended report; the reason and full amendment history were not established.
- A [Hyundai AGM notice dated March 2, 2020](https://www.hyundai.com/content/hyundai/ww/data/ir/investorNews/2020/0000000910/files/Convening-notice-of-52nd-annual-general-meeting.pdf) explicitly labels its financial statements preliminary before audit. Its FY2018 separate comparative tax expense is 295,007 million and net income 414,941 million, different from the DART receipt above. It is not substituted as the final benchmark. This difference is observed, not explained as a proven restatement.

Sources were downloaded for inspection on October 26, 2026. DART public HTML was accessible through direct retrieval without an API key even when search-service opening returned 502. This is a bounded two-filing check, not a replacement bulk API. Metadata records SHA256 snapshots, original units, receipt IDs and column locators. Snapshots stay under the ignored external directory; users can retrieve originals from the linked primary sources.

## Method and limits

The two familiar issuers and 2018 were selected for diagnostic convenience and reform-year relevance after phase 66 results were known. Selection is neither random nor preregistered. No population error rate is inferred. Public references were transcribed and manually checked against downloaded filings; this is not an automated filing extraction system.

Monetary comparisons convert million KRW to KRW and allow rtol=5e-6, atol=1 KRW for released numeric precision. This is a diagnostic tolerance, not proof of the dataset's rounding algorithm. Rate comparisons retain phase66 rtol=1e-6, atol=1e-8. Phase66 counts used a tighter monetary tolerance and remain unchanged. All five monetary fields pass the declared phase67 rule in both issuer-years.

Positive annual GETR/CETR agree with positive conventional burden ratios for these cases. No sign correction is applied. `tax` is not renamed: its actual meaning remains unverified. LEV and SIZE discrepancies remain unresolved; top-bound concentration is not sufficient to claim a specific winsorization procedure. Separate-statement compatibility here does not establish accounting scope or currency consistency for other rows, and accounting pretax income is not filed domestic taxable income.

## Reproduce and next step

1. Obtain the pinned phase65 input with `python scripts/inspect_firm_data_v65.py --download`.
2. Run `python scripts/reconcile_source_sample_v67.py --download`. Source hash changes stop the run for manual review; existing snapshots are also verified.
3. Run `python -m unittest discover -s calibration -p 'test_*v6*.py'`.

Next: add a predefined diagnostic sample spanning years, issuer sizes and both passing/failing phase66 identities; independently reconcile raw fields and request/locate the exact preprocessing specification before using supplied LEV/SIZE or multi-year ETRs. The general descriptive pilot, causal interpretation and fiscal calibration remain unapproved. License/version conflicts from phase65 remain unresolved.