Hashevolution · Hashevolution · Jun 2, 2026 · Jun 2, 2026
diff --git a/CLAUDE.md b/CLAUDE.md
@@ -15,20 +15,31 @@ See `docs/ARCHITECTURE.md` for full design principles and non-goals.
   QVT α track. v0.2.0 (Foundation Hardening) and v0.3.0
   (Platform Skeleton) both fully closed; v0.4.0 (Layer 4 Lifecycle:
   T1+T7+T2 first bundle) shipped 2026-05-27.
-- **Active theme**: **α-7 closure (5-tier mode reshape) + Track 2c
-  collaboration parallel + bidi gate security fix** (three concurrent
-  tracks as of 2026-06-02). α-7 PR #680 draft (graph top-K + threshold +
-  bucket-(d) phrases). Baseline showed -0.107 abst_f1 regression on
-  null queries at gemma4 production tier (outside noise); 5-tier
-  remeasurement in flight (background) to test mode-pattern reshape
-  across 5 tiers (1b/4b/12b/27b/e4b). Track 2c (Ali Arabic adversarial
-  18 cases) PR #681 draft — preserved imports + design + schema remap
-  + sweep runner + 29-unit-test bidi audit. Bidi runtime gate PR #682
-  draft — `core/input_normalization.py` + `routes/query.py` wire +
-  29 tests. α-6 cycle closed 2026-06-01 PM (PR #678).
-  Sequence: α-7 closure → α-8 typed-filter (evidence-of-absence
-  preservation rule R1-R5 baked in) → Phase 3b proper on α-8 baseline
-  per Option C (2026-06-02 decision). Domain pilot (v0.5) still gated.
+- **Active theme**: **α-8 ontology typed-filter** (R1-R5 evidence-of-
+  absence preservation, architectural fix for α-7's wrong-knob). Phase
+  A (PR #688 `fcf343d`) + Phase B (PR #689 `6d6698e`) landed
+  2026-06-02:
+  - `core/ontology.py` +5 horizontal types (event/date/location/
+    quantity/project) + 6 relations (OCCURRED_AT/HAPPENED_ON/
+    LOCATED_IN/INVOLVES/MEASURED_AS/WORKED_ON)
+  - `core/graph_typed_filter.py` new module + `JAMES_DISABLE_TYPED_FILTER`
+    flag (disable-polarity, default OFF = filter ON)
+  - Matrix cell `C_rag-ontology` between `C_rag-graph` (now pre-α-8
+    baseline with flag forced) and `C_rag-full`
+  - `pipeline_context.build_unified_context` prepends typed entity
+    summary BEFORE graph context (byte-additive)
+  - 137 tests pass; ruff clean
+  - Default behavior post-merge: typed filter ACTIVE in `/query/`
+
+  Phase C N=3 baseline at M_M background launched 2026-06-02 (ETA
+  2.2h). Phase C 5-tier remeasurement + Phase D closure analysis +
+  `docs/ARCHITECTURE.md §5.7` typed filter section = 다음 세션.
+  α-7 cycle closed as REJECT 2026-06-02 (PR #680, 10th wrong-fix-
+  averted). α-6 closed 2026-06-01 PM (PR #678). Track 2c Phase 4
+  M_M sweep complete 2026-06-02 (0 breach / 11 resisted / 6 partial /
+  1 manual_review — α-8 acceptance test target poison_01 identified).
+  Sequence resumes: α-8 Phase C+D → 5-tier remeasurement → Phase 3b
+  proper on α-8 baseline. Domain pilot (v0.5) still gated.
 - **Strategic frame**: We are still building a **mother platform**;
   v1.0 vertical domain branching remains the only authorised
   specialisation point. The 2026-06-01 strategic discussion narrowed
@@ -116,7 +127,8 @@ See `docs/ARCHITECTURE.md` for full design principles and non-goals.
 
 | Purpose | File |
 |---|---|
-| **🟢 NEXT SESSION ENTRY (read this first)** | **`docs/handovers/v0.4-next-session-entry-2026-06-02-post-tracks.md`** |
+| **🟢 NEXT SESSION ENTRY (read this first)** | **`docs/handovers/v0.4-next-session-entry-2026-06-02-evening.md`** |
+| Previous entry (2026-06-02 post-tracks — now sealed) | `docs/handovers/v0.4-next-session-entry-2026-06-02-post-tracks.md` |
 | Previous session entry (α-7 launch — sealed) | `docs/handovers/v0.4-next-session-entry-2026-06-02-alpha7.md` |
 | Previous-previous (Phase 3a launch — sealed) | `docs/handovers/v0.4-next-session-entry-2026-06-01-PM.md` |
 | Track 2c integration design memo | `docs/design/v0.4-track-2c-arabic-adversarial-integration.md` |

diff --git a/docs/handovers/v0.4-next-session-entry-2026-06-02-evening.md b/docs/handovers/v0.4-next-session-entry-2026-06-02-evening.md
@@ -0,0 +1,175 @@
+# JAMES — Next Session Entry (2026-06-02 evening)
+
+> Cold-start orientation for the next Claude Code session. Sealed
+> 2026-06-02 evening after α-8 Phase A+B landed and Phase C M_M
+> sanity launched in background.
+
+## 0. TL;DR (30 sec read)
+
+α-7 closed as **REJECT** (10th wrong-fix-averted, PR #680). α-8 architectural
+fix landed in **two PRs the same day**:
+
+- **PR #688** `fcf343d` — Phase A: ontology +5 horizontal types + typed
+  filter module + flag (`JAMES_DISABLE_TYPED_FILTER`)
+- **PR #689** `6d6698e` — Phase B: matrix cell `C_rag-ontology` +
+  pipeline_context typed-prefix overlay
+
+Default behavior post-merge: typed filter **ACTIVE** in `/query/`. Phase C
+N=3 baseline at M_M kicked off in background 2026-06-02 (ETA ~2.2h);
+output log `reports/research-runs/alpha8-phase-c-sanity.log`, per-cell
+JSON in `reports/research-runs/qvt-ablation-cells/`.
+
+Other parallel landings (5 PRs merged this session):
+- **PR #684**: 30-error ruff F-class cleanup (main CI debt cleared)
+- **PR #685**: Bucket-(d) 11 oracle phrases (gemma3 family refusal coverage)
+- **PR #686**: Track 2c Ali Arabic adversarial Phase 1-3 + α-8 R1-R5 update
+- **PR #687**: Track 2c Phase 4 M_M sweep results (0 breach + poison_01
+  acceptance target)
+- **PR #682**: Bidi gate runtime input normalization (Track 2c X3 response)
+
+Total: 8 PRs landed + 2 PRs closed (#680 REJECT, #681 superseded by #686).
+
+## 1. First commands when you sit down
+
+```powershell
+cd C:\Project\James-RAG-Evol-v010
+git status
+git log --oneline -10
+
+# Check Phase C background run status (running or completed?)
+type "C:\Users\karu-\AppData\Local\Temp\claude\C--Project-James-RAG-Evol-v010\c47eaf4c-729a-45b2-a903-1825b34fc205\tasks\bt5gdm2jw.output"
+
+# Inspect per-cell JSON outputs if completed:
+ls reports\research-runs\qvt-ablation-cells\
+
+# If Phase C run was killed / incomplete, restart:
+python scripts/qvt_ablation_matrix.py `
+    --sector-cells C_rag-graph,C_rag-ontology `
+    --tiers M_M --n-runs 3 --resume
+```
+
+## 2. State of α-8 (4-phase plan)
+
+| Phase | Status | Notes |
+|---|---|---|
+| **A** ontology + typed filter module + flag | ✅ PR #688 merged | core/ontology.py 14.9 KB / core/graph_typed_filter.py 10.7 KB |
+| **B** matrix cell + pipeline_context wiring | ✅ PR #689 merged | C_rag-ontology between C_rag-graph (forced flag=1) and C_rag-full |
+| **C** N=3 baseline (M_M) | ⏳ background launched 2026-06-02 | log/cells in reports/research-runs/; verify completion before next launch |
+| **C-extend** 5-tier remeasurement | ⏸️ pending | M_XS/M_S/M_M/M_L/M_XL × n=3 × 2 cells = 30 cells ≈ 6-10h compute |
+| **D** closure analysis + ARCHITECTURE.md | ⏸️ pending | Per-question-type cross-tab + R1-R5 acceptance check (multihop null + Ali poison_01/04) |
+
+## 3. Acceptance tests for α-8
+
+Per design memo §2.4:
+
+1. **Multihop_RAG null query test**: re-run `scripts/research/audit_12b_null_query_refusal_shape.py` against post-α-8 baseline at M_M. Pass if null-query abst_f1 stays ≥ α-6 baseline level (0.609). Fail if regression (= same R1+R2 gap as α-7).
+
+2. **Track 2c poison_01 + poison_04 retest**: re-run `scripts/adversarial_sweep.py` against M_M with α-8 active (default ON). Pass if poison_01 verdict promotes from `manual_review` to `resisted` (LLM emits explicit "leather jacket not in catalog" statement). Fail if poison_01 stays `manual_review` or worse.
+
+   Phase 4 sweep results (pre-α-8 baseline for comparison):
+   - 11 resisted / 6 partial / 0 failed / 1 manual_review (poison_01)
+   - 0 auto-failure across 18 cases
+   - File: `reports/research-runs/adversarial-sweep-M_M-20260602T084119.json`
+
+## 4. Quality Delta Card framework (Phase D)
+
+Baseline reference (per design memo §1.4): **C_rag-graph** (post-α-7 baseline = pre-α-8 path with `JAMES_DISABLE_TYPED_FILTER=1` forced in sector cell).
+
+Verdict tree:
+- `graded Δ ≥ +0.030` → ⭐⭐ adopt + ARCHITECTURE.md + finding memory promoted
+- `+0.010 ≤ graded Δ < +0.030` → tier-gated adopt (which tiers per cross-tab)
+- `graded Δ ≤ +0.010 OR abst_f1 < α-6 baseline` → ⭐ operational only (reject typed filter as architectural fix; heuristic top-K declared sufficient on this benchmark, memory promoted as such)
+
+Per design memo §1.3 honest framing:
+- Even if Δ exceeds noise, finding is ⭐⭐ partial (typed filter is decades-old technique; MS GraphRAG / LlamaIndex have type hints). JAMES contribution = numbers + routing rule + empirical position.
+- ⭐⭐⭐ candidate elevation requires cross-fixture sanity (multihop + Ali Track 2c crossover) — Phase D incorporates this.
+
+## 5. Phase 3b proper status
+
+Per `feedback_graph_layer_top_k_debt`: publishable graded claim requires α-7 closure → α-8 measurement → cross-family sanity. As of session close:
+
+- α-7 closure: ✅ REJECT
+- α-8 baseline measurement: ⏳ Phase C in flight
+- Cross-family (5 families × n=3 ≈ 45h compute): ⏸️ post-α-8 Phase D
+- Position per design memo §1.4 + 2026-06-02 Option C decision: Phase 3b proper runs **on α-8 baseline**, not pre-α-8.
+
+## 6. Track 2c collaboration status (Ali)
+
+| Phase | Status | File |
+|---|---|---|
+| 1: preserved imports + integration design | ✅ PR #686 | `eval/adversarial/ar_ecommerce-{README-ali,REPORT-provia,provia-results,v1.1-pending}.{md,json,yaml}` + `docs/design/v0.4-track-2c-arabic-adversarial-integration.md` |
+| 2: YAML remap (byte-exact U+202E preservation 18/18) | ✅ PR #686 | `eval/adversarial/ar_ecommerce-v1.1-james.yaml` |
+| 3: adversarial_sweep.py + DSL parser (18 tests) | ✅ PR #686 | `scripts/adversarial_sweep.py` + `tests/test_adversarial_criteria_parser.py` |
+| 4: JAMES M_M sweep | ✅ PR #687 | analysis: `reports/research-runs/track-2c-phase-4-jam-m_m-analysis-20260602.md` |
+| 5: cross-stack report (Ali vs JAMES diff) | ⏸️ post-α-8 acceptance |  |
+| 6: α-8 acceptance integration (poison_01/04 retest) | ⏸️ Phase D |  |
+| 7: Ali cross-stack DM | ⏸️ post-Phase 6 |  |
+
+## 7. Wrong-fix-averted ledger
+
+| Cycle | Count |
+|---|---:|
+| α-5 | 7 |
+| α-6 | 2 (rate-limit + tier override) |
+| α-7 | 1 (top-K rejected at 5-tier scale) |
+| **Cumulative** | **10** |
+
+Additional category catch this session: 1 "cumulative branch contamination" (PR #681 superseded by #686) — recognised pattern; logged in PR #686 description.
+
+## 8. Files added this session (Where to find what)
+
+| Topic | Path |
+|---|---|
+| α-8 design memo | `docs/design/v0.4-alpha-8-ontology-typed-filter.md` (§2.4 R1-R5; §3 integration plan; §4 verdict tree) |
+| α-7 closure analysis | `reports/research-runs/alpha-7-closure-analysis-20260602.md` |
+| Track 2c Phase 4 analysis | `reports/research-runs/track-2c-phase-4-jam-m_m-analysis-20260602.md` |
+| Track 2c Phase 4 raw JSON | `reports/research-runs/adversarial-sweep-M_M-20260602T084119.json` |
+| Bidi audit | `reports/research-runs/bidi-normalization-audit-20260602.md` |
+| α-7 bucket-(d) sub-finding | `reports/research-runs/alpha-7-bucket-d-oracle-phrase-gap.md` |
+| α-6 recovery curve | `reports/research-runs/alpha-6-phase-3a-recovery-curve-20260601.md` (§3.1 has 12b correction annotation) |
+
+## 9. Memory updates this session
+
+New entries (4):
+- `project_alpha_7_closure_state` — α-7 REJECT context
+- `feedback_alpha7_top_k_destroys_positive_contributions` — wrong-knob mechanism
+- `feedback_s4_citation_survives_context_reshape` — ⭐⭐⭐ candidate strengthened
+- `feedback_12b_pure_capability_misclassified_as_plateau` — 12b plateau correction
+- `project_alpha_8_phase_a_b_landed` — α-8 Phase A+B state (this session)
+
+## 10. Open decisions left for next session
+
+1. **Phase C completion**: verify background run completed; if partial, resume with `--resume` flag. If full, proceed to analysis.
+2. **5-tier remeasurement scope**: full M_XS+M_S+M_M+M_L+M_XL × n=3 vs subset (e.g., M_M + M_XL only). Decision driver: Phase C M_M result magnitude.
+3. **Phase D closure timing**: write closure analysis after 5-tier OR after just M_M+M_XL. Trade-off: completeness vs cycle velocity.
+4. **Ali Phase 5 DM share**: wait for Phase 6 acceptance test (poison_01 retest at α-8) OR share Phase 4-5 results immediately for cross-stack discussion.
+5. **Phase 3b proper launch position**: per design memo §1.4 Option C confirmed (post-α-8 baseline). Compute timing depends on Phase C completion.
+
+## 11. Don't-forget guardrails
+
+- **No domain features until v1.0** (CLAUDE.md rule #1) — α-8 5 new types are all horizontal (event/date/location/quantity/project); regulation/transaction/recipe deferred.
+- **20 KB module gate** — `core/ontology.py` now 14.9 KB; further additions need split-first.
+- **Quality Delta Card required** for `core/retrieval|graph|reasoning` PRs (CLAUDE.md rule #2). Phase B PR #689 used `Quality delta: exempt (label: code — A/B oracle infra)` — Phase D PR will land the actual Δ numbers.
+- **Honest framing tier** for every finding (memory `feedback_finding_size_honest_framing`). α-8 ceiling = ⭐⭐ partial.
+
+## 12. Korean handover snippet
+
+이번 세션 (2026-06-02): α-7 REJECT 확정 (PR #680, 10번째 wrong-fix-averted)
++ α-8 Phase A+B 두 PR 한 세션 머지 (#688/#689) + Track 2c Phase 4 sweep
+(0 breach, poison_01 acceptance target 명시) + 5 PR cleanup cascade
+(#683/#684/#685/#682/#686/#687).
+
+α-8 Phase A: `core/ontology.py` 에 horizontal 5 type + 6 relation 추가,
+`core/graph_typed_filter.py` 신규 모듈 + flag 추가. R1-R5 evidence-of-
+absence preservation rule 모두 코드로 박힘. 137 test pass.
+
+α-8 Phase B: matrix cell `C_rag-ontology` + `pipeline_context` typed-
+prefix overlay. Default ACTIVE in production (flag 미설정 시 filter ON).
+
+Phase C N=3 baseline at M_M background 실행 중 — log 위치
+`reports/research-runs/alpha8-phase-c-sanity.log`. 다음 세션 cold-start
+시 완료 여부 확인 → 결과 분석 또는 재실행.
+
+다음 cycle 순서: Phase C M_M 분석 → 5-tier 확장 측정 → Phase D closure
+analysis + ARCHITECTURE.md §5.7 → cross-family Phase 3b proper (45h)
+→ v0.5 도메인 (horizontal moat = 기업 내부지식).