# Phase 69 — boundary reconstruction and a smaller-issuer extension

## Result

All remaining phase68 LEV and SIZE discrepancies can be numerically reconstructed by applying the observed released bounds to the supplied-lag candidate formulae. This is an in-sample, post-hoc explanation, not independent confirmation of the original preprocessing recipe.

| Released indicator | Lag candidate | Unclipped exceptions | Exceptions at lower bound | Exceptions at upper bound | Matches after clipping |
|---|---|---:|---:|---:|---:|
| LEV | lag_liab / lag_asset | 135 | 135 | 0 | 12,653/12,653 |
| SIZE | log(lag_asset) | 180 | 18 | 162 | 12,653/12,653 |

Bounds: LEV [0.026000464, 0.871121839]; SIZE [23.62073049, 30.62155284]. Each lower-bound exception has a candidate below the lower bound, and each upper-bound exception has a candidate above the upper bound. No nonfinite pairs were present. Comparisons retain phase68 rounding sensitivity rtol1e-5, atol1e-8; source-statement comparisons below use their separate phase67 tolerances.

The bounds are observed minimum/maximum values of the same released indicators. Discovery preceded the recorded phase69 hypothesis; this is not a held-out validation or preregistered test. Perfect reconstruction therefore does not establish why the bounds were chosen, which percentile or reference sample was used, whether clipping happened before sample selection, or whether these are appropriate predictors in a new study. No arbitrary tolerance sweep, source-data overwrite or inferred statutory tax-base assignment is performed.

The phase68 135 and 180 exceptions are no longer numerically unexplained under this candidate transformation. Their original processing provenance remains unresolved. The two exception counts may overlap in issuer-years and must not be added as a unique-row count.

## Original filing extension

A third issuer was added: Dong Wha Pharmaceutical, stock000020. It is the smallest numerical stock code in the panel and has an accessible original separate filing. Its size is substantially below the two previous issuers; this is not a legal SME classification or a random size-stratified sample. All three annual columns (2015–2017) were selected before comparing them to panel values.

Primary source: [DART receipt20180618000261](https://englishdart.fss.or.kr/dsbh002/viewer.do?rcpNo=20180618000261). The downloaded page identifies an unamended annual report with separate financial statements, annual columns2017/2016/2015 and amounts in KRW, unlike the million-KRW units in the earlier sample. The selected accounting fields and cash-tax outflow were manually transcribed from their corresponding statements; snapshot SHA256 and column locators are recorded in `data/metadata/residual_plan_v69.json`.

For the three added issuer-years, assets, liabilities, equity, pretax and net income all agree after the declared currency conversion and precision tolerance (15/15 monetary checks). GETR and CETR agree with independently sourced tax expense and cash-tax payments divided by pretax income (6/6 rate checks). The hypothesis that raw tax is cash taxes paid fails in all three (0/3).

Together with the unchanged phase68 sample, this is nine issuer-years in three companies: 45/45 monetary and 18/18 annual-rate checks pass. The raw-tax-as-cash-tax hypothesis fails in all nine. The extension is still purposive and tiny relative to 1,754 firms; its pass rate is not a population accuracy estimate. Report-version and accounting-scope limitations from phase67 remain applicable. No firm rows from the licensed dataset are published; the metadata contains independently collected public filing facts and results contain diagnostic counts/booleans.

## Reproduce

1. Acquire the pinned phase65 main CSV.
2. Run `python scripts/audit_residuals_v69.py --download` to acquire/check the pinned original filing and rebuild results.
3. Inspect `data/processed/residual_audit_v69.json`.
4. Run `python -m unittest discover -s calibration -p 'test_*v6*.py'`.

The input and filing checksums must match; a changed source stops the run for review. Raw sources remain ignored and are not public research assets.

## Next research decision

Use this evidence to document candidate lag/bound transformations rather than silently treating LEV/SIZE as current-year raw ratios. Obtain an independent preprocessing specification before claiming those construction rules are verified. A productive next pilot can be built directly from the public filing reference sample, with transparent separate-statement units and unadjusted annual tax ratios, explicitly as a small descriptive case study. It must not be used to infer a reform effect, national tax elasticity or a domestic taxable-income threshold.

Full-panel descriptive approval, original preprocessing certification, causal approval and fiscal calibration remain false. Phase65 release/license inconsistencies are not resolved by numerical reconstruction.