diff --git a/.gitignore b/.gitignore
index 5d2d2e9..71fe2f4 100644
--- a/.gitignore
+++ b/.gitignore
@@ -5,3 +5,4 @@ composer.lock
 .idea/
 .vscode/
 graphify-out/
+.DS_Store
diff --git a/CHANGELOG.md b/CHANGELOG.md
index 693710e..a80d997 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -14,7 +14,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 - `tests/TurkishCorpusQualityTest.php` with aggregate FP/FN/coverage gate
 - `docs/dictionary-expansion-policy.md` — criteria for future Turkish dictionary entry PRs
 - Inline Turkish lexicon taxonomy in `docs/dictionary-expansion-policy.md` — four supported categories (`profanity`, `insult`, `sexual`, `abbreviation`), severity guidance, Batch 1 acceptance policy, corpus requirements, and maintainer curation workflow
-- `docs/batch1-tr-lexicon-research.md` — Batch 1 Turkish lexicon research and curation document scaffold (structure only; research not yet performed)
+- `docs/batch1-tr-lexicon-research.md` — Batch 1 Turkish lexicon research and curation document (40-candidate pool, maintainer evaluation, 15 approved entries for PR 3)
 
 ### Changed
 
diff --git a/docs/batch1-tr-lexicon-research.md b/docs/batch1-tr-lexicon-research.md
index 2f5ee2f..4ea6411 100644
--- a/docs/batch1-tr-lexicon-research.md
+++ b/docs/batch1-tr-lexicon-research.md
@@ -4,7 +4,7 @@ This document supports v0.4 Batch 1 Turkish lexicon expansion. It is a planning
 
 For expansion policy, taxonomy, and quality gates, see [`docs/dictionary-expansion-policy.md`](dictionary-expansion-policy.md).
 
-**Status:** Structure prepared — research not yet performed.
+**Status:** Maintainer curation complete — 2026-07-04. Ready for PR 3 implementation.
 
 ---
 
@@ -32,20 +32,26 @@ This document covers **research and curation only**.
 
 ## Research Method
 
-Research for this document will use **agent-reach** in a follow-up step for high-level Turkish usage pattern research. No research results are recorded yet.
+Research was performed in four parallel agent reports (no additional web research in this curation step):
 
-### Intended research contexts
+| Research input | Focus |
+|----------------|-------|
+| Agent 1 — gaming/chat research | Gaming, Discord, Twitch, live-chat abbreviations and obfuscation |
+| Agent 2 — general internet research | Forums, social media, general Turkish internet insult/profanity lanes |
+| Agent 3 — FP review | False-positive quality review (binding tie-break) |
+| Agent 4 — normalization review | Normalized-key collision and variant mapping review |
 
-| Context | Focus |
-|---------|--------|
-| Turkish gaming/chat usage | Broad profanity and abbreviation patterns in live chat |
-| Twitch/YouTube comment style | Common evasion and obfuscation in video comments |
-| Turkish Reddit/forum-style discourse | Frequency vs niche slang; standalone token usage |
-| Common Turkish internet abbreviations | High-confidence short forms; collision risk |
-| Obfuscation/evasion patterns | Leetspeak, separators, Turkish character folding |
-| False positive and normalized collision risks | Substring neighbors; overlap with existing seed entries |
+The individual agent reports were temporary maintainer research artifacts. They are summarized here and are not intended to be committed as long-lived repository documents.
 
-Research output at each stage must follow the [Research discipline](#research-discipline) rules below.
+### Tie-break priority
+
+When agents disagree, classifications follow this order:
+
+1. **FP review** (Agent 3) — binding on standalone-token and normalization-collision FP
+2. **Normalization review** (Agent 4) — binding on normalized-key uniqueness and variant mapping
+3. **Research evidence** (Agents 1 + 2) — frequency, usage context, obfuscation patterns
+
+Existing seed entries (`amk`, `aq`, `siktir`, `orospu`, `mal`) are baseline coverage — not re-proposed as new candidates.
 
 ---
 
@@ -62,11 +68,51 @@ Research output at each stage must follow the [Research discipline](#research-di
 
 ## Stage 1 — Candidate Pool
 
-Approximately **30–40 candidates** will be collected here after research. Table is empty until research is performed.
-
-| ID | Candidate term | Category | Likely severity | FP risk | Collision risk | Usage evidence summary | Corpus plan | Notes |
-|----|----------------|----------|-----------------|---------|----------------|------------------------|-------------|-------|
-| | | | | | | | | |
+**40 unique candidates** merged from Agents 1 and 2, evaluated by Agents 3 and 4. Seed duplicates excluded.
+
+| ID | Canonical term | Category | Proposed severity | Confidence | FP risk | Normalization notes | Research source(s) |
+|----|----------------|----------|-------------------|------------|---------|---------------------|-------------------|
+| C01 | salak | insult | low | high | low | Key `salak` stable; leet `s4lak`, repetition, separators converge | A1, A2, A3, A4 |
+| C02 | aptal | insult | low | high | high | Key `aptal` unique; `aptalca` is different token | A1, A2, A3, A4 |
+| C03 | gerizekalı | insult | medium | high | low | Key `gerizekali`; reject `gerzek` as second row | A1, A2, A3, A4 |
+| C04 | dangalak | insult | low | high | low | Key `dangalak` unique | A1, A3, A4 |
+| C05 | yavşak | insult | medium | high | medium | Key `yavsak`; ASCII `yavsak` same key; zoological homonym | A1, A2, A3, A4 |
+| C06 | şerefsiz | insult | medium | high | low–medium | Key `serefsiz` after fold | A1, A3, A4 |
+| C07 | pezevenk | insult | medium | high | low | Key `pezevenk`; `pzvnk` truncation not covered | A1, A2, A3, A4 |
+| C08 | andaval | insult | low | medium | low | Key `andaval` unique | A1, A3, A4 |
+| C09 | gavat | insult | medium | high | low | Key `gavat` unique | A2, A3, A4 |
+| C10 | dalyarak | insult | medium | high | low | Key `dalyarak` unique | A2, A3, A4 |
+| C11 | piç | insult | medium | high | high | Key `pic` — **collision with English `pic`** | A1, A2, A3, A4 |
+| C12 | kahpe | insult | medium | high | low | Key `kahpe` unique | A1, A2, A3, A4 |
+| C13 | kaltak | sexual | medium | high | low | Key `kaltak` unique | A1, A3, A4 |
+| C14 | sürtük | sexual | medium | high | low | Key `surtuk` after fold | A1, A2, A3, A4 |
+| C15 | yarrak | sexual | high | high | medium | Key `yarrak`; defer `yarak`/`yrk` as separate rows | A1, A2, A3, A4 |
+| C16 | yarak | sexual | high | medium | medium | Key `yarak` ≠ `yarrak` — duplicate intent | A1, A3, A4 |
+| C17 | amcık | sexual | high | high | medium | Key `amcik`; `amcıq` → `amciq` not covered | A1, A2, A3, A4 |
+| C18 | göt | sexual | medium | high | high | Key `got` — **collision with English `got`**; `6öt` leet gap | A1, A2, A3, A4 |
+| C19 | puşt | sexual | medium | medium | low | Key `pust` after fold | A1, A3, A4 |
+| C20 | kaşar | sexual | low | medium | high | Key `kasar` — **collision with cheese homograph** | A1, A3, A4 |
+| C21 | fahişe | sexual | medium | medium | low–medium | Key `fahise`; historical-descriptive standalone use | A1, A3, A4 |
+| C22 | sokayım | profanity | high | high | high | Key `sokayim`; innocent `sokmak` conjugation homonym | A2, A3, A4 |
+| C23 | siktirgit | profanity | medium | medium | low | Key `siktirgit` distinct from seed `siktir` | A2, A3, A4 |
+| C24 | lan | profanity | low | high | high | Key `lan` unique; ubiquitous particle standalone | A2, A3 |
+| C25 | sokarım | profanity | high | medium | high | Key `sokarim`; innocent `sokmak` conjugation | A2, A3 |
+| C26 | ananı | insult | high | medium | medium | Key `anani`; phrasal fragment | A2, A3 |
+| C27 | aminakoyayim | profanity | high | medium | low | Long phrase token; poor standalone dictionary shape | A2, A3 |
+| C28 | soktugum | profanity | high | low | medium | Morphological family too broad | A2, A3 |
+| C29 | hassiktir | profanity | medium | medium | low | Key `hassiktir`; elongation of seed `siktir` | A2, A3, A4 |
+| C30 | ibne | — | — | high | — | Slur-domain; v0.4 excluded | A1, A2, A3, A4 |
+| C31 | sktr | abbreviation | medium | high | low | Key `sktr` ≠ `siktir`; requires own row | A1, A3, A4 |
+| C32 | hsktir | abbreviation | medium | medium | low | Key `hsktir`; depends on `sktr` precedent | A1, A3, A4 |
+| C33 | amq | abbreviation | medium | medium | medium | Key `amq`; 3-char; redundant with seed `amk`/`aq` | A1, A3, A4 |
+| C34 | gerzek | insult | low | medium | low | Key `gerzek` ≠ `gerizekali`; semantic duplicate of C03 | A1, A3, A4 |
+| C35 | pic | insult | medium | high | very high | ASCII duplicate of C11; same key `pic` | A2, A3 |
+| C36 | amcığa | sexual | high | low | medium | Agglutinative fragment | A1, A3 |
+| C37 | sikik | profanity | medium | medium | medium | Root overlap with seed `siktir` | A1, A3 |
+| C38 | götlalesi | insult | medium | low | low | Niche compound meme insult | A2, A3 |
+| C39 | godoş | insult | medium | medium | low | Dated in younger cohorts | A2, A3 |
+| C40 | kezban | insult | low | medium | medium | Stereotype insult; given-name collision edge | A2, A3 |
+| C41 | oç / mk / sg | abbreviation | — | high | very high | ≤2 characters; acronym storm — Batch 1 automatic reject | A1, A2, A3, A4 |
 
 **Category values:** `profanity`, `insult`, `sexual`, `abbreviation`
 
@@ -78,33 +124,230 @@ Approximately **30–40 candidates** will be collected here after research. Tabl
 
 ## Stage 2 — Maintainer Evaluation
 
-Every Stage 1 candidate must be classified here with a short rationale.
+Every Stage 1 candidate classified below. Tie-breaks applied per Agent 3 (FP) then Agent 4 (normalization).
 
 | Candidate ID | Status | Rationale |
 |--------------|--------|-----------|
-| | | |
-
-**Allowed statuses:**
-
-| Status | Meaning |
-|--------|---------|
-| **Recommended** | Strong fit for Batch 1; passes selection rules |
-| **Deferred** | Potentially valid; postponed (risk, collision, evidence gap) |
-| **Rejected** | Does not meet acceptance criteria |
+| C01 | **Recommended** | Both research agents + FP review agree; dominant everyday insult usage; standard clean corpus sufficient (see [A08 re-evaluation](#a08--salak)) |
+| C02 | **Deferred** | Agent 3 binding: high meta-linguistic/educational standalone FP; Agent 1 deferred; normalization OK but FP gate not achievable |
+| C03 | **Recommended** | Multi-source high confidence; unique key; reject `gerzek` sibling per normalization one-key rule |
+| C04 | **Recommended** | Agent 1 + FP + normalization aligned; no innocent homograph |
+| C05 | **Deferred** | Agent 3 binding: zoological standalone homonym (yavşak balığı); Agent 1 recommended overridden |
+| C06 | **Recommended** | Agent 1 + FP aligned; formal-news adjective trade-off documented, acceptable |
+| C07 | **Recommended** | Cross-source convergence; unique key; low FP |
+| C08 | **Recommended** | Agent 3 promoted over Agent 1 deferral — fills insult slot with clean token shape |
+| C09 | **Recommended** | Agent 2 + FP + normalization aligned; Ekşi salience; unique key |
+| C10 | **Recommended** | Agent 2 + FP aligned; long token, low FP |
+| C11 | **Deferred** | Agent 3 binding: `piç` → `pic` normalization collision with English `pic`; Agent 2 recommended overridden |
+| C12 | **Recommended** | Both agents + FP aligned; categorized as `insult` (character betrayal register) |
+| C13 | **Recommended** | Agent 1 + FP + normalization aligned |
+| C14 | **Recommended** | Agent 3 FP overrides Agent 2 ethics deferral; low lexical FP |
+| C15 | **Recommended** | Dominant sexual slang; canonical over `yarak`; obfuscation gaps documented in corpus |
+| C16 | **Deferred** | Normalization: `yarak` ≠ `yarrak`; duplicate moderation intent — cover via C15 obfuscation corpus |
+| C17 | **Recommended** | Both agents + FP aligned; distinct from seed `amk`/`aq` |
+| C18 | **Deferred** | Agent 3 binding: `göt` → `got` English collision; 3-char; `6öt` leet pipeline gap |
+| C19 | **Deferred** | Agent 1 ethics defer + borderline slur-adjacent; no urgent Batch 1 need |
+| C20 | **Deferred** | Agent 3 binding: `kaşar` → `kasar` cheese homograph; normalization collision |
+| C21 | **Recommended** | Agent 3 optional slot candidate; historical-descriptive FP manageable with extra clean frames |
+| C22 | **Rejected** | Agent 3 binding: innocent `sokmak` conjugation homonym — FP gate unreachable; Agent 4 normalization OK irrelevant |
+| C23 | **Deferred** | Compound token; semantic overlap with seed `siktir`; matcher design clarity first |
+| C24 | **Deferred** | Ubiquitous particle standalone; 3-char; requires 8+ filler clean frames |
+| C25 | **Deferred** | Morphological homonym of `sokmak`; same innocent-conjugation FP class as C22 |
+| C26 | **Deferred** | Phrasal maternal insult fragment; phrase-aware matching not in v0.4 |
+| C27 | **Deferred** | Phrase-level token; non-standard dictionary standalone shape |
+| C28 | **Rejected** | Broad morphological family; low standalone precision |
+| C29 | **Deferred** | Elongation of seed `siktir`; obfuscation corpus item, not new row |
+| C30 | **Rejected** | Slur-domain; v0.4 policy exclusion |
+| C31 | **Recommended** | Agent 4 normalization requires separate row; Agent 3 optional slot — promoted with abbreviation corpus gate (6 clean cases) |
+| C32 | **Deferred** | Abbreviation class; defer until C31 passes one release cycle |
+| C33 | **Deferred** | 3-char; redundant with seed `amk`/`aq` family |
+| C34 | **Deferred** | Truncation duplicate of C03; one normalized key per intent |
+| C35 | **Rejected** | Identical normalized key to innocent English `pic`; duplicate of C11 |
+| C36 | **Rejected** | Agglutinative fragment; unstable standalone dictionary shape |
+| C37 | **Rejected** | Root-family overlap with seed `siktir`; suffix-aware matching not in v0.4 |
+| C38 | **Rejected** | Niche compound; low corpus ROI |
+| C39 | **Rejected** | Dated register; declining usage |
+| C40 | **Rejected** | Stereotype insult; given-name collision edge cases |
+| C41 | **Rejected** | ≤2 characters; Batch 1 automatic reject per policy |
+
+**Stage 2 totals:** Recommended **15** · Deferred **16** · Rejected **9**
 
 ---
 
 ## Stage 3 — Approved Batch 1
 
-Maintainer selects **exactly 15 entries** from the **Recommended** group only. PR 3 may only use these approved entries.
+Maintainer selects **exactly 15 entries** from the **Recommended** group. PR 3 may only use these approved entries.
+
+**Category mix:** insult ×8 · sexual ×6 · abbreviation ×1 · profanity ×0
+
+Policy targets (≈6 profanity, ≈3 insult, ≈4 sexual, ≈2 abbreviation) are flexible guidance — see note below.
+
+> **Why Batch 1 adds no new `profanity` entries**
+>
+> Batch 1 intentionally contains insults, sexual terms, and one abbreviation, but no new profanity rows. This is a quality decision, not a category omission.
+>
+> Every profanity candidate evaluated in research failed at least one v0.4 acceptance gate without an acceptable trade-off:
+>
+> - **`sokayım` / `sokarım`** — innocent first-person `sokmak` (insert/plug) conjugations share the same standalone token
+> - **`lan`** — ubiquitous colloquial particle; high standalone-token FP despite high frequency
+> - **`göt`** — normalized key `got` collides with English `got`; 3-character tier escalation unmet
+> - **`siktirgit` / `hassiktir`** — semantic overlap with seeded `siktir`; compounds better handled as obfuscation corpus items
+> - **`aminakoyayim`** — phrase-level token; poor standalone dictionary shape
+>
+> Core profanity coverage for v0.4 already exists in the seed dictionary (`amk`, `aq`, `siktir`, `orospu`). Batch 1 prioritizes net-new insult and sexual lanes where FP = 0 is achievable under exact-token matching. Profanity expansion resumes in a later batch when phrase-aware matching, mixed-language corpus lanes, or 3-character policy exceptions are available.
 
 | Approved ID | Candidate term | Category | Severity | Corpus requirement summary | Approval notes |
 |-------------|----------------|----------|----------|----------------------------|----------------|
-| | | | | | |
+| A01 | gerizekalı | insult | medium | ≥3 clean (+2 meta-linguistic), ≥3 profane, ≥3 obfuscated (ASCII fold, leet, `gerzek` gap note) | Canonical long form; do not add `gerzek` row; normalized key `gerizekali` |
+| A02 | dangalak | insult | low | ≥3 clean, ≥3 profane, ≥3 obfuscated (leet, repetition, separators) | Unique key; teammate-blame gaming signal |
+| A03 | pezevenk | insult | medium | ≥3 clean, ≥3 profane, ≥3 obfuscated (leet, separators; `pzvnk` gap note) | Cross-source convergence |
+| A04 | şerefsiz | insult | medium | ≥3 clean (+2 formal-news frames), ≥3 profane, ≥3 obfuscated (ASCII fold `serefsiz`) | Accept formal adjective trade-off; context out of v0.4 NLP scope |
+| A05 | gavat | insult | medium | ≥3 clean, ≥3 profane, ≥3 obfuscated | Relationship-shaming insult; Ekşi salience |
+| A06 | dalyarak | insult | medium | ≥3 clean, ≥3 profane, ≥3 obfuscated (repetition) | Long token; very low innocent usage |
+| A07 | andaval | insult | low | ≥3 clean, ≥3 profane, ≥3 obfuscated | Lower frequency but clean FP profile |
+| A08 | salak | insult | low | ≥3 clean, ≥3 profane, ≥3 obfuscated (leet, repetition, separators) | Standard ≥4-char insult; see re-evaluation note |
+| A09 | kahpe | insult | medium | ≥3 clean (+1 historical-literary), ≥3 profane, ≥3 obfuscated | Character betrayal insult register |
+| A10 | kaltak | sexual | medium | ≥3 clean, ≥3 profane, ≥3 obfuscated (leet, repetition) | No common innocent homograph |
+| A11 | sürtük | sexual | medium | ≥3 clean, ≥3 profane, ≥3 obfuscated (ASCII `surtuk` fold) | Agent 3 FP overrides ethics deferral |
+| A12 | yarrak | sexual | high | ≥3 clean, ≥3 profane, ≥3 obfuscated (+2 `yarak`/dots/spaced; `yrk` gap note) | Prefer over `yarak`; high gaming + social volume |
+| A13 | amcık | sexual | high | ≥3 clean (+2 fold-variant neighbors), ≥3 profane, ≥3 obfuscated (ASCII `amcik`, leet; `amcıq` gap note) | Distinct from seed `amk`/`aq` |
+| A14 | fahişe | sexual | medium | ≥3 clean (+2 historical-descriptive), ≥3 profane, ≥3 obfuscated | Agent 3 optional slot; lower priority than A10–A13 |
+| A15 | sktr | abbreviation | medium | ≥6 clean (+3 gaming handle/acronym frames), ≥3 profane, ≥3 obfuscated (`s.k.t.r`, leet `5ktr`, `sktr git` frame) | First abbreviation row; proves category in CI; distinct key from `siktir` |
+
+**Approval gate:** Exactly **15** rows present and signed off.
+
+**Maintainer approval recorded:** 2026-07-04 — Batch 1 approved for PR 3 implementation.
+
+### Per-entry approval detail
+
+#### A01 — `gerizekalı`
+
+- **Category / severity:** insult / medium
+- **Rationale:** Top-tier intelligence insult in ranked chat and forum arguments; unique normalized key; both research agents recommend.
+- **Minimum corpus:** 3 clean (+2 quoted/meta frames), 3 profane, 3 obfuscated
+- **Normalization:** Author term `gerizekalı` → key `gerizekali`; document `gerzek`/`glk` as detection gaps
+
+#### A02 — `dangalak`
+
+- **Category / severity:** insult / low
+- **Rationale:** Frequent teammate-blame insult; no common innocent sense.
+- **Minimum corpus:** 3 clean, 3 profane, 3 obfuscated
+- **Normalization:** Stable ASCII key; leet `d4ng4lak` converges
+
+#### A03 — `pezevenk`
+
+- **Category / severity:** insult / medium
+- **Rationale:** Stable standalone insult; discord-filter + terlik.js convergence.
+- **Minimum corpus:** 3 clean, 3 profane, 3 obfuscated
+- **Normalization:** `pzvnk` chat truncation not covered — obfuscated limitation note
+
+#### A04 — `şerefsiz`
+
+- **Category / severity:** insult / medium
+- **Rationale:** Rage-chat character insult; distinct normalized key.
+- **Minimum corpus:** 3 clean (+2 formal-news), 3 profane, 3 obfuscated
+- **Normalization:** ASCII fold `serefsiz`; formal adjective usage is documented trade-off
+
+#### A05 — `gavat`
+
+- **Category / severity:** insult / medium
+- **Rationale:** Persistent forum/social relationship-shaming insult; Ekşi 13p proxy.
+- **Minimum corpus:** 3 clean, 3 profane, 3 obfuscated
+- **Normalization:** Key `gavat` unique
+
+#### A06 — `dalyarak`
+
+- **Category / severity:** insult / medium
+- **Rationale:** High-velocity escalatory insult in open comment sections.
+- **Minimum corpus:** 3 clean, 3 profane, 3 obfuscated
+- **Normalization:** Long token reduces abbreviation noise
+
+#### A07 — `andaval`
+
+- **Category / severity:** insult / low
+- **Rationale:** Fills Batch 1 with lower FP than contested short tokens; clean token shape.
+- **Minimum corpus:** 3 clean, 3 profane, 3 obfuscated
+- **Normalization:** Key `andaval` unique
+
+#### A08 — `salak`
+
+- **Category / severity:** insult / low
+- **Rationale:** Default competence insult across gaming and general internet; both agents agree.
+- **Minimum corpus:** 3 clean, 3 profane, 3 obfuscated (standard ≥4-character minimum)
+- **Normalization:** Key `salak` stable; leet, repetition, and separator variants converge
+- **FP re-evaluation (maintainer, 2026-07-04):** Initial curation elevated FP risk because Agent 3 listed *salak balığı* as a zoological homonym. On review, that usage is niche ichthyology terminology — not a high-frequency standalone token in everyday Turkish text. Unlike `yavşak` (actively used as a common fish name in fishing and biology discourse), `salak` in modern usage is overwhelmingly the competence insult. VerbaGuard exact-token matching already prevents embedded-substring FPs (`salaklık`, compound forms). No normalization collision exists. **FP risk reduced to `low`; standard clean corpus applies.** A single optional clean neighbor documenting the obscure fish sense is sufficient if encountered during PR 3 corpus authoring — not a mandatory expanded biology lane.
+
+#### A09 — `kahpe`
+
+- **Category / severity:** insult / medium
+- **Rationale:** Character betrayal insult; stable standalone token; low collision.
+- **Minimum corpus:** 3 clean (+1 historical-literary), 3 profane, 3 obfuscated
+- **Normalization:** Key `kahpe` unique
+
+#### A10 — `kaltak`
+
+- **Category / severity:** sexual / medium
+- **Rationale:** terlik.js sexual tier; stable standalone token.
+- **Minimum corpus:** 3 clean, 3 profane, 3 obfuscated
+- **Normalization:** Key `kaltak` unique
+
+#### A11 — `sürtük`
+
+- **Category / severity:** sexual / medium
+- **Rationale:** Common sexual insult in live chat; low lexical FP per Agent 3.
+- **Minimum corpus:** 3 clean, 3 profane, 3 obfuscated
+- **Normalization:** Author `sürtük` → key `surtuk`; ASCII fold in obfuscated corpus
+
+#### A12 — `yarrak`
+
+- **Category / severity:** sexual / high
+- **Rationale:** Dominant sexual slang in FPS trash talk and social comments; prefer over `yarak`.
+- **Minimum corpus:** 3 clean, 3 profane, 3 obfuscated (+2 `yarak` as positive obfuscation)
+- **Normalization:** Do not add `yarak`/`yrk`/`yrrk` rows; document gaps
+
+#### A13 — `amcık`
+
+- **Category / severity:** sexual / high
+- **Rationale:** discord-filter default; high gaming frequency; distinct from seed family.
+- **Minimum corpus:** 3 clean (+2 fold neighbors), 3 profane, 3 obfuscated
+- **Normalization:** Author `amcık` → key `amcik`; `amcıq` → `amciq` not covered
+
+#### A14 — `fahişe`
+
+- **Category / severity:** sexual / medium
+- **Rationale:** Agent 3 optional slot; valid sexual slang with manageable historical-descriptive FP.
+- **Minimum corpus:** 3 clean (+2 historical-descriptive), 3 profane, 3 obfuscated
+- **Normalization:** Author `fahişe` → key `fahise`
+
+#### A15 — `sktr`
+
+- **Category / severity:** abbreviation / medium
+- **Rationale:** Primary gaming abbreviation for `siktir`; Agent 4 requires separate row; fills abbreviation slot.
+- **Minimum corpus:** 6 clean (+3 innocent gaming/acronym), 3 profane, 3 obfuscated
+- **Normalization:** Key `sktr` ≠ `siktir`; spelled-chain `s.k.t.r` converges; first abbreviation proof gate
+
+---
+
+## PR 3 — Maintainer implementation note
+
+PR 3 is delivered as **one feature branch and one merged pull request**. Do not split Batch 1 across multiple GitHub PRs.
+
+Implementation should proceed in **two internal phases** on that single branch:
+
+| Phase | Scope | Exit criteria |
+|-------|--------|---------------|
+| **Phase 1** | Implement approximately half of the 15 approved entries (dictionary rows + corpus fixtures) | Corpus quality verified for implemented entries; full test suite passes |
+| **Phase 2** | Implement the remaining approved entries | Aggregate FP/FN/coverage gates pass; Batch 1 complete |
+
+Suggested Phase 1 split (7 entries — insult + sexual core, lowest FP complexity):
+
+`gerizekalı`, `dangalak`, `pezevenk`, `salak`, `kahpe`, `kaltak`, `sürtük`
+
+Suggested Phase 2 split (8 entries — remaining insults, sexual, abbreviation proof gate):
 
-**Approval gate:** Exactly **15** rows must be present and signed off before PR 3 begins.
+`şerefsiz`, `gavat`, `dalyarak`, `andaval`, `yarrak`, `amcık`, `fahişe`, `sktr`
 
-**Maintainer approval recorded:** _Not yet — pending research and curation._
+Phase boundaries are a **workflow convenience** for the implementer — not separate review units. All 15 entries land in one PR with one risk assessment block per entry.
 
 ---
 
@@ -114,7 +357,22 @@ Candidates classified as **Deferred** in Stage 2. Copied here for tracking and f
 
 | Candidate ID | Candidate term | Reason for deferral | Reconsideration condition |
 |--------------|----------------|---------------------|---------------------------|
-| | | | |
+| C02 | aptal | Meta-linguistic/educational standalone FP; Agent 3 binding defer | Expanded meta-linguistic clean lane; FP = 0 proven |
+| C05 | yavşak | Zoological homonym (fish); Agent 3 binding defer | Biology/fishing clean corpus (+4 frames) |
+| C11 | piç | Normalized key `pic` collides with English `pic`; Agent 3 binding defer | Mixed-language TR+EN corpus lane |
+| C16 | yarak | Duplicate intent with C15 `yarrak`; different key | Only if `yarrak` corpus insufficient for recall |
+| C18 | göt | English `got` collision; 3-char; `6öt` leet gap | Mixed-language corpus + pipeline `6`→`ö` evaluation |
+| C19 | puşt | Borderline slur-adjacent; ethics defer | Post-v0.4 slur-adjacent policy |
+| C20 | kaşar | Cheese homograph `kasar`; intent ambiguity | Food-domain clean corpus (+6 frames) |
+| C23 | siktirgit | Compound overlap with seed `siktir` | Matcher compound policy clarified |
+| C24 | lan | Ubiquitous particle standalone; 3-char FP | Standalone particle policy + 8+ clean frames |
+| C25 | sokarım | Innocent `sokmak` conjugation homonym | Phrase-aware matching |
+| C26 | ananı | Phrasal maternal insult fragment | Phrase-aware matching roadmap |
+| C27 | aminakoyayim | Phrase-level token shape | Multi-token expression support |
+| C29 | hassiktir | Elongation of seed `siktir` | Obfuscation corpus-only unless separate row justified |
+| C32 | hsktir | Abbreviation class; depends on A15 precedent | After `sktr` passes one release cycle |
+| C33 | amq | 3-char; redundant with seed `amk`/`aq` | Abbreviation policy exception |
+| C34 | gerzek | Truncation duplicate of C03 | Only if long form insufficient for chat recall |
 
 ---
 
@@ -124,7 +382,18 @@ Candidates classified as **Rejected** in Stage 2. Copied here for audit trail.
 
 | Candidate ID | Candidate term | Reason for rejection |
 |--------------|----------------|----------------------|
-| | | |
+| C22 | sokayım | Innocent `sokmak` conjugation homonym — FP gate unreachable (Agent 3 X01) |
+| C28 | soktugum | Broad morphological family; low standalone precision |
+| C30 | ibne | Slur-domain; v0.4 policy exclusion |
+| C35 | pic | ASCII duplicate of `piç`; English `pic` collision |
+| C36 | amcığa | Agglutinative fragment; unstable standalone shape |
+| C37 | sikik | Root overlap with seed `siktir`; suffix matching not in v0.4 |
+| C38 | götlalesi | Niche compound meme insult |
+| C39 | godoş | Dated register in younger cohorts |
+| C40 | kezban | Stereotype insult; given-name collision edge |
+| C41 | oç / mk / sg | ≤2 characters; Batch 1 automatic reject |
+
+**Seed entries not re-proposed:** `amk`, `aq`, `siktir`, `orospu`, `mal` — already in [`data/tr.php`](../data/tr.php).
 
 ---
 
@@ -132,15 +401,15 @@ Candidates classified as **Rejected** in Stage 2. Copied here for audit trail.
 
 PR 3 cannot start until all items below are satisfied:
 
-- [ ] Candidate pool contains approximately 30–40 entries
-- [ ] Every candidate classified (Recommended / Deferred / Rejected)
-- [ ] Exactly 15 approved entries in Stage 3
-- [ ] No slur entries
-- [ ] No new ≤2 character entries
-- [ ] Every approved entry has a corpus plan
-- [ ] FP risk documented for every approved entry
-- [ ] Normalized collision risk checked against existing dictionary and other candidates
-- [ ] Maintainer approval recorded
+- [x] Candidate pool contains approximately 30–40 entries (40 unique candidates in Stage 1)
+- [x] Every candidate classified (Recommended / Deferred / Rejected)
+- [x] Exactly 15 approved entries in Stage 3
+- [x] No slur entries
+- [x] No new ≤2 character entries
+- [x] Every approved entry has a corpus plan
+- [x] FP risk documented for every approved entry
+- [x] Normalized collision risk checked against existing dictionary and other candidates
+- [x] Maintainer approval recorded
 
 ---
 
@@ -149,6 +418,7 @@ PR 3 cannot start until all items below are satisfied:
 | Document | Purpose |
 |----------|---------|
 | [`docs/dictionary-expansion-policy.md`](dictionary-expansion-policy.md) | Taxonomy, Batch 1 rules, corpus requirements |
+| Internal agent research artifacts | Summarized research inputs for gaming/chat, general internet, false-positive review, and normalization review |
 | [`docs/specification.md`](specification.md) | Dictionary author format |
 | [`CONTRIBUTING.md`](../CONTRIBUTING.md) | Three-PR expansion workflow |
 | [`tests/Fixtures/tr/README.md`](../tests/Fixtures/tr/README.md) | Corpus fixture format |