Update README.md

README.md

@@ -44,7 +44,7 @@ row["token_classes"]    # BIOES targets, aligned to row["tokens"]
 row["privacy_mask"]     # [{"label": "USER_ID", "start": 30, "end": 38, "value": "..."}, ...]
 ```
 
-Train a token classifier on `(tokens, token_classes)`; report span-level precision / recall / F1 on `validation`.
+Train a token classifier on `(tokens, token_classes)`. Use `validation` span metrics for checkpoint selection and iteration comparison; evaluate release candidates only on the separately maintained Golden Set test split.
 
 <br>
 
@@ -59,7 +59,7 @@ ai4privacy/pii-masking-openpii-1.5m (ko + en)
    ↓ (3) synthesize domain templates × value generators (finance/card/bank/insurance/security)
    │                 · method A: mention-level perturbation of refined ko rows
    │                 · method B: static template × value-generator synthesis (no LLM at runtime)
-   │                 · hard-example injection: O decoys + label-confusion/context pairs + boundary positives (§2.1.1)
+   │                 · hard-example injection: O decoys + label-confusion/context pairs + boundary positives (§2.4)
    ↓ (4) mix        concat + shuffle(seed=42) → BIOES derivation → integrity + RRN audit
 This dataset (train / validation)
 ```
@@ -69,7 +69,7 @@ This dataset (train / validation)
 * **Locally generated values in the value pool** are md5-partitioned per split (train / validation) with empty pairwise intersection, enforced by a disjointness assertion.
 
 ## 2.2. Splits
-| split | rows | ko | en | en-share (replay) | role |
+| split | rows | ko | en | en-share | role |
 |-------|-----:|---:|---:|---------:|------|
 | `train` | 58,494 | 42,072 | 16,422 | 28.1% | training input |
 | `validation` | 14,543 | 10,460 | 4,083 | 28.1% | HPO / checkpoint selection |
@@ -227,41 +227,45 @@ All PII values are synthetic - none are drawn from any real individual's records
 
 | regime | rows | how values are kept safe |
 |--------|-----:|--------------------------|
-| **Inherited values** (`ko-refine` + `en-replay`) | 42,026 (57.5%) | Values come from the upstream **ai4privacy** synthetic corpus (LLM / `faker`-generated, CC-BY-4.0) and are re-labeled. `RRN`/`FRN` and `CARD_NUMBER` are additionally invalidated (§3.2). |
+| **Inherited values** (`ko-refine` + `en-replay`) | 42,026 (57.5%) | Values come from the upstream **ai4privacy** synthetic corpus (LLM / `faker`-generated, CC-BY-4.0) and are re-labeled. Applicable `RRN` and `CARD_NUMBER` values are additionally invalidated (§3.2); `FRN` occurs only on locally generated paths. |
 | **Locally generated values** (`ko-perturb` + `domain`) | 31,011 (42.5%) | Values come from independent local generators. Labels covered by §3.2 additionally use reserved or unassigned formats or deliberately failing checksums. |
 
 The `source_dataset` buckets are not equivalent to these value-provenance regimes. The `openpii-1.5m-ko` bucket mixes refined rows with 9,409 perturbed rows whose PII is locally re-generated. Consequently, the source buckets total 51,435 `openpii-*` rows (70.4%) and 21,602 `domain` rows (29.6%), while the value-provenance table above correctly reports 57.5% inherited and 42.5% locally generated rows.
 
-## 3.1. Why this is not personal data
-Under Korea's Personal Information Protection Act (PIPA), a value is personal information only if it (1) matches a real individual's real value **and** (2) singles that individual out, alone or via easy linkage. The claim fails on two independent grounds:
+## 3.1. Why this dataset is designed not to contain personal data
+Under Korea's Personal Information Protection Act (PIPA), personal information is information about a living individual that can identify the individual directly or through easy combination with other information; the statutory definition also includes pseudonymized information. This dataset's safety position therefore rests on provenance - rows are designed not to be information about any real individual - rather than on a simplified test that a synthetic value can never resemble a real value. See [PIPA Article 2](https://www.law.go.kr/LSW/lsLinkCommonInfo.do?chrClsCd=010202&lsJoLnkSeq=1030669293).
 
 * **Not sourced from real records.** Every value is synthetic - generated by our rule/`faker` pipeline, or inherited from the synthetic ai4privacy corpus. There is no original record behind any value, so this is **not pseudonymized personal data**; there is nothing to re-identify.
-* **No source linkage even on a coincidental match.** If a synthetic value (say a `faker` mailbox at `gmail.com`) happens to equal a real one, the rest of the row - name, address, phone - is independently synthetic, so the row does not reproduce a coherent source record. Separately, 27,738 of 73,037 rows (38.0%) contain at least one value governed by the checksum/Luhn controls or domain-only structural controls in §3.2. The remaining rows rely on synthetic provenance and absence of source linkage, not on a claim that every value is structurally impossible.
+* **No source linkage even on a coincidental match.** If a synthetic value (say a `faker` mailbox at `gmail.com`) happens to equal a real one, the rest of the row - name, address, phone - is independently synthetic, so the row does not reproduce a coherent source record. Separately, 31,848 of 73,037 rows (43.6%) contain at least one value governed by the checksum/Luhn controls or locally generated structural controls in §3.2. The remaining rows rely on synthetic provenance and absence of source linkage, not on a claim that every value is structurally impossible.
 
-By label, identifiability breaks down further: **names / addresses / dates / zipcodes** do not single out a person alone (namesakes; a place is not a person); **account-number / user-id / generic-id / secret / I-PIN / CVC / card-expiry** are random values whose owners sit in non-public registries, so a value alone looks nobody up. The residual - a synthetic value coincidentally equal to a real one, with no way to link it - is the accepted, industry-standard risk of any synthetic-PII corpus, identical to the risk the upstream public dataset already carries.
+Individual labels and combinations can be identifying in real-world data, so label type or non-public registry access is not treated as an independent legal basis. The documented structural controls reduce accidental-match risk, while the primary basis remains synthetic provenance and absence of source linkage.
 
 The statements in this section describe **design intent and empirical verification results** (checksum audits, generator-fingerprint checks); they are representations, not a legal warranty - the dataset is distributed as-is (see [Legal Disclaimer](#6-legal-disclaimer)).
 
 ## 3.2. Per-label invalidation
-Two tiers. The first is applied in the refine stage to **all sources** (inherited and synthesized); the second is a **construction property of synthesized (`domain`) rows only** - for inherited rows these labels carry ai4privacy `faker` values and rely on the §3.1 identifiability argument.
+Two tiers. The first is enforced across all source paths for values covered by the public checksum controls. The second is a construction property of locally generated paths (`ko-perturb` + `domain`); inherited `ko-refine` and `en-replay` values for these labels rely on the §3.1 provenance argument.
 
-**Invalidated across all sources** (public checksums - a mathematical guarantee):
+**Checksum/Luhn controls across all source paths:**
 
 | label | mechanism |
 |-------|-----------|
 | `RRN`, `FRN` | old-scheme unification + checksum digit +5 (deliberately fails the check digit; holds regardless of birth band) |
-| `CARD_NUMBER` | Luhn check digit +5, for values ≥13 digits (real-PAN length) |
+| `CARD_NUMBER` | Luhn check digit +5 for values ≥13 digits; shorter values are audited separately as structural residuals |
 
-**Structural, synthesized (`domain`) rows only** (inherited rows use `faker` values here):
+**Structural controls on locally generated paths (`ko-perturb` + `domain`):**
 
 | label | mechanism |
 |-------|-----------|
 | `EMAIL` | RFC 2606 reserved domains (`example.com/net/org`) |
-| `PHONE` | unassigned prefix (leading `0`/`1` for ko; `555-01XX` for en) |
+| `PHONE` | unassigned subscriber prefix (subscriber-number block begins with `0`/`1` for ko; `555-01XX` for en) |
 | `DRIVER_LICENSE` | unassigned region code |
 | `PASSPORT` | unissued prefix (avoidance - may weaken over time) |
 
-`RRN` is unified to the **pre-2020.10 scheme** so invalidity holds regardless of birth band (only the post-2020.10 number-change program leaves a documented residual). `CVC` is never generated standalone - always alongside an (invalid) PAN. The Luhn invalidation targets values ≥13 digits; a handful of ≤12-digit inherited values may still pass Luhn, but 12 digits is below real card-network length, so they are not real PANs. The legacy hangul notation of `DRIVER_LICENSE` (e.g. `서울 96-123456-01`) carries a real region name, so it falls back to the §3.1 linkage argument.
+`RRN` is unified to the **pre-2020.10 scheme** so invalidity holds regardless of birth band (only the post-2020.10 number-change program leaves a documented residual). `CVC` is never generated standalone - it is always paired with a Luhn-invalid PAN.
+
+The current Luhn invalidation targets values ≥13 digits. The audit found 17 inherited 12-digit values that pass Luhn (train 14 / validation 3). ISO/IEC 7812-1 permits PANs shorter than 13 digits, and Luhn validity establishes structural possibility rather than actual issuance. These 17 values therefore remain a documented structural residual and rely on synthetic provenance until the invalidation rule is expanded to the full plausible PAN-length range. See [ISO's PAN-length explanation](https://www.iso.org/news/2016/11/Ref2146.html) and [PCI SSC's Luhn guidance](https://www.pcisecuritystandards.org/faqs/1137/).
+
+The legacy hangul notation of `DRIVER_LICENSE` (e.g. `서울 96-123456-01`) carries a real region name, so it falls back to the §3.1 provenance argument.
 
 <br>
 
@@ -271,7 +275,7 @@ Two tiers. The first is applied in the refine stage to **all sources** (inherite
 * **`PASSPORT` safety is avoidance-based**: the unissued-prefix guarantee can weaken if issuing authorities later allocate those prefixes - unlike the checksum/Luhn labels it is not mathematically permanent (flagged honestly)
 * **`RRN` residual**: pre-2020.10 unification makes checksum invalidity time-independent, but the post-2020.10 number-change program is a residual case not fully covered
 * **`IPIN` value spec pending**: I-PIN values are generated at pilot volume only, pending a finalized value-format confirmation
-* **Inherited values can be real-looking**: 42,026 rows (57.5%) follow inherited-value paths and retain upstream ai4privacy values except for labels re-invalidated in §3.2. These values can use real domains or formats (e.g. `gmail.com` emails, real-format phones, and a few ≤12-digit Luhn-passing card values). Their safety rests primarily on synthetic provenance and absence of source linkage rather than structural invalidation. Upstream annotation conventions and residual noise can also propagate into this portion.
+* **Inherited values can be real-looking**: 42,026 rows (57.5%) follow inherited-value paths and retain upstream ai4privacy values except for labels re-invalidated in §3.2. These values can use real domains or formats (e.g. `gmail.com` emails, real-format phones, and 17 inherited 12-digit card values that pass Luhn). Their safety rests primarily on synthetic provenance and absence of source linkage rather than structural invalidation. Upstream annotation conventions and residual noise can also propagate into this portion.
 
 <br>