@konneal/engine 0.1.0

package/docs/REDESIGN-NORMATIVE-RAG-ETSI.md

@@ -0,0 +1,485 @@
+# Normative RAG Redesign — Learning from the ETSI Experience
+
+*Primary source: Al Masoud, Arazzi, Germani, Nocera (Univ. Pavia + CEI),
+"Exploring Structural Complexity in Normative RAG with Graph-based approaches:
+A case study on the ETSI Standards", arXiv:2604.09868, Jan 2026.
+Secondary primaries only where they corroborate a specific claim (cited inline).
+This supersedes the structure/graph sections of RESEARCH-SOTA-2026.md and
+ROADMAP-SOTA.md for retrieval design. 2026-08-27.*
+
+---
+
+## 0. Why this paper is the authority for us
+
+It is the only published empirical study of RAG **on industrial standards**
+(ETSI EN 301 489-X, ≈50 documents, ≈3000 pages, 800+ synthetic Q&A with
+witness-string golden chunks). Co-authored with CEI (Italian Electrotechnical
+Committee) — a standards body, not a vendor lab. Their problem statement is
+ours word-for-word:
+
+> standards are highly cross-referential, semi-structured, and make use of
+> specialized formal vocabulary… dense text embeddings… overlook [precise
+> keywords] in favor of fuzzier semantic similarity… parent-child (and
+> sibling-to-sibling) dynamics between chunks [are] a relevant factor.
+
+They evaluate **lightweight, low-latency** modules only — no GNN training,
+no agentic loops in the measured stack. That matches our Cloudflare /
+cost-first constraint. Their conclusions are measured, not hyped. We treat
+them as load-bearing.
+
+---
+
+## 1. Their Information Model (the contract we must match)
+
+```
+IG = (IU, E)          InfoUnits + Edges
+IU node = {title, body}
+E = P ∪ C             Parthood ∪ Citation
+```
+
+| Relation | Semantics | Cardinality |
+|---|---|---|
+| **P (parthood)** | document→section, section→subsection; section code prefix (`4.1` parent of `4.1.2`) | tree (≤1 parent) |
+| **C (citation)** | internal (same doc, section-only mention) or external (doc name ± section) | many-to-many |
+| Ancestry | transitive closure of P; same publication ⇔ share common ancestor | derived |
+
+Construction rules that matter:
+
+1. **ToC → ordered InfoUnits**, then P from section-code prefix geometry.
+2. **Oversized sections may split** — children of the original section
+   (Structured + Chunks). Flat chunking alone is the baseline they beat.
+3. **Tabular sections are exempted from being chunked.** Atomic units.
+4. **Empty-body InfoUnits are valid** (a section title that immediately
+   opens a subsection; a top-level document node).
+5. **Citation resolve heuristic:** section-only → internal; doc-name-only →
+   document; both → that section of that document.
+6. They explicitly flag the **IEC Smart Standards** machine-readable
+   framework as the upgrade path when content is no longer PDF-scraped —
+   i.e. our Metanorma/#592 direction, named from the standards world.
+
+---
+
+## 2. Their modules and measured outcomes (the only numbers that matter)
+
+Eight configurations × three K values. Variants: Vanilla (flat 300-word
+chunks) vs Structured (official section hierarchy) vs Structured+Chunks
+(hierarchy, split when oversized). Modules stacked on top.
+
+| Module | What it does | Measured effect on ETSI corpus |
+|---|---|---|
+| **Structure preservation** (P in the index) | section hierarchy as the chunk boundary | **↑ precision, ↑ MRR; recall ≈ flat (not significant)** |
+| **Structured + Chunks** | hierarchy + split oversized sections | **overall best compromise** |
+| **BM25 prefilter → dense rerank → RRF** | sparse first (exact jargon), dense for meaning, fuse | **↑ precision, ↑ MRR; recall unchanged** (expected — fusion reorders, doesn't recover) |
+| **Embedding smoothing** `em ← α·em + (1-α)·mean(N(u))` | blend each node embedding with parent/children/siblings/citation neighbors; lightweight alternative to GNN graph-embeddings | **slight ↑ recall** — the only module that moved recall |
+| **Neighbor expansion + graph re-rank** (seed → neighbors, hub penalty) | post-retrieval graph walk | **not effective** — scoring function inadequate for the domain |
+
+Their conclusion, verbatim force:
+
+> the methods we have tested mostly improve the **precision** of retrieval,
+> and **structure-preservation with chunking** seems to be the overall best
+> compromise.
+>
+> Neighbor expansion… does not seem to be effective, indicating that more
+> work is needed…
+>
+> Smoothing slightly improves recall… promising way of embedding relational
+> information into the index.
+
+Future work they name: multi-hop synthesis across distant passages / across
+documents; hybrid solutions for that class; evaluation on broader Q&A with
+more complex link structure.
+
+---
+
+## 3. Audit of OIML SMART AI against the ETSI stack
+
+### 3.1 What we already get right (keep, do not rebuild)
+
+| ETSI requirement | Our status | Evidence |
+|---|---|---|
+| Structure-preserving chunk boundaries | ✅ clause-boundary sections from Metanorma HTML/adoc | `ingest/parse.py`, `chunk.py` |
+| Structured + Chunks (split oversized) | ✅ `split_long` on section text | `chunk.py` |
+| Dense + lexical RRF | ⚠️ partial — see 3.2 | `hybrid.ts` + `pipeline.ts` post-rerank RRF |
+| Cross-encoder rerank | ✅ bge-reranker-base | beyond their Granite-only dense |
+| Document-level citation / supersession graph | ✅ successor, part_of (doc↔family), defines | `ingest/graph.py`, D1 registry |
+| Edition/authority steering | ✅ derived active edition (status field lies) | beyond ETSI scope; matches 2026 freshness scoring |
+| Contextual enrichment | ✅ 99.4% corpus | Anthropic recipe; they have none — we are ahead on context-in-embedding |
+| Machine-readable source path | 🟡 hand-rolled; #592 is the contract | they name IEC Smart Standards; we have Metanorma |
+| Generation-side grounding | ✅ quote anchors, CRAG, reflection, faithfulness judges | out of their retrieval-only scope |
+| Don't over-invest neighbor expansion | ✅ graph lane is terminology-only (`defined_terms→defines→doc_numbers`) | aligns with their negative result |
+
+### 3.2 Gaps the ETSI evidence says are material (ordered by measured impact)
+
+#### G-ETSI-1 — Lexical retrieval is a post-filter, not a prefilter  【CRITICAL】
+
+**ETSI:** BM25 is the *first* query-time stage — high-recall, low-latency
+prune of the full corpus on exact keywords (standards jargon), *then* dense
+rerank, *then* RRF.
+
+**Us:** `keywordRank` + `rrfFuse` run **only over the dense top-K already
+retrieved** (`pipeline.ts` after Vectorize). If dense never recalled the
+chunk, lexical cannot recover it. We re-order; we do not recall.
+
+Vectorize has no sparse vectors (platform fact). The fix is not "wait for
+Vectorize" — it is a **corpus-wide inverted index we own**:
+
+- D1 FTS5 over `chunk_id + text` (or a dedicated terms table), OR
+- build posting lists at ingest into KV/R2, query in-worker.
+
+BM25 (or BM25-ish with real IDF from corpus stats) returns top-N ids → fetch
+vectors/metadata → union with dense top-K → RRF → cross-encoder. This is
+exactly their II-B5→6→7 cascade, and it is the single highest-leverage
+gap their numbers identify for a standards corpus.
+
+#### G-ETSI-2 — No section-level parthood in the retrieval graph  【HIGH】
+
+**ETSI InfoModel:** every InfoUnit participates in P; ancestry is
+queryable; siblings and parents are first-class neighbors for smoothing
+and (attempted) expansion.
+
+**Us:** `clause_anchor` is a **string on chunk metadata**. The D1 graph has
+doc↔family `part_of` and concept `defines`, but **not**
+`clause:R-60-1:2021:4.1 —part_of→ clause:R-60-1:2021:4`. Parent/sibling
+geometry is not traversable at query time and cannot feed smoothing.
+
+Fix at ingest: for every chunk with anchor `A.B.C`, emit
+`part_of` edges up the anchor tree + a `section` node per unique anchor.
+Empty-body section nodes (title-only parents) are valid per ETSI.
+
+#### G-ETSI-3 — No internal citation edges  【HIGH】
+
+**ETSI:** C is half the edge set. "§4.1 (see also §6.3)" and "as required
+by OIML R 76-1:2006 §3.5" are first-class. Resolve heuristic is simple and
+domain-correct.
+
+**Us:** relaton gives *document-level* bibliographic edges. Clause-body
+mentions of other clauses/documents are not extracted into C. The graph
+cannot answer "what does this clause require me to also read?"
+
+Fix at ingest: citation mention parser over chunk text (section-id /
+docidentifier patterns we already normalize via pubid-ish regexes) →
+resolve against the section-node index from G-ETSI-2 → `cites` edges.
+External unresolved mentions stay as string-valued edge targets (public
+projection rule already forbids leaking ISO titles).
+
+#### G-ETSI-4 — No embedding smoothing  【HIGH — only measured recall lever】
+
+**ETSI eq. (10):**
+`em_j ← α·em_j + (1-α)·mean({em_n : n ∈ N(j)})`
+where N = parent ∪ children ∪ siblings ∪ in/out citations.
+
+This is the **only module that moved recall**. It is offline (index-time),
+O(|E|) once per build, no GNN, no query latency. It is a drop-in after
+G-ETSI-2/3 give us real neighbors.
+
+Implementation: after embed, one or two smoothing iterations over the
+section graph, write smoothed vectors to Vectorize (or a parallel index
+for A/B). α ≈ 0.7–0.85 (keep self-dominant). Do **not** smooth across
+family boundaries or superseded editions.
+
+#### G-ETSI-5 — Tables still flattened  【HIGH — they exempt them】
+
+**ETSI:** "Tabular sections are exempted from being chunked." Atomic.
+
+**Us:** G1 prototype extracts 10k typed tables (adoc-first) into
+`artifacts/table-chunks.jsonl` — **not in the live index**. Prose pipeline
+still linearizes tables into clause text.
+
+Ship G1: upsert `chunk_type=table` rows as first-class InfoUnits with
+parthood to their parent clause. Do not split table rows across chunks.
+Re-enrichment cost ≈$50 is the gate, not the design.
+
+#### G-ETSI-6 — Precision-first retrieval metrics under-weighted  【MED】
+
+**ETSI primary charts:** AP@K, MRR@K, R@K — and their wins are on **AP and
+MRR**, not recall.
+
+**Us:** `retrieval.mjs` hit@5 (recall-ish), golden citation_any, RAGAS
+faithfulness/relevancy/precision (generation-side). We under-measure the
+axis their methods actually move.
+
+Add AP@K / MRR@K on the retrieval suite with witness-style golden spans
+(we can derive from golden cases' expected citations). Report them on
+every retrieval change. Stop treating recall@5 as the only retrieval KPI.
+
+#### G-ETSI-7 — Neighbor expansion scoring  【LOW — do not chase】
+
+**ETSI:** expander + graph re-rank **failed**. Our terminology-only graph
+lane is the correct restrained form. Do not build a general "expand to all
+neighbors" path until a domain-adequate neighbor score is validated on
+our golden set. Smoothing (G-ETSI-4) is the proven way to push relational
+signal *into the vectors* instead.
+
+---
+
+## 4. Redesigned pipeline (target architecture)
+
+### 4.1 Offline (RDP) — once per index build
+
+```
+OII (Metanorma model / adoc / HTML fallback)
+        │
+        ▼
+┌──────────────────────────────────────────────────────────┐
+│ PARSE                                                     │
+│  sections → InfoUnits {id, title, body, anchor, doc_id} │
+│  tables   → InfoUnits chunk_type=table (ATOMIC)           │
+│  formulas → InfoUnits chunk_type=equation (when typed)    │
+│  empty-body section nodes kept                            │
+└──────────────────────────────────────────────────────────┘
+        │
+        ▼
+┌──────────────────────────────────────────────────────────┐
+│ GRAPH                                                     │
+│  P: anchor-tree part_of (clause↔clause, clause↔doc)       │
+│  C: internal + external cites (mention resolve)           │
+│  existing: successor, family part_of, defines             │
+│  documents registry (derived active) unchanged            │
+└──────────────────────────────────────────────────────────┘
+        │
+        ▼
+┌──────────────────────────────────────────────────────────┐
+│ ENRICH (contextual preamble, quality-first lane) — keep   │
+└──────────────────────────────────────────────────────────┘
+        │
+        ▼
+┌──────────────────────────────────────────────────────────┐
+│ EMBED → SMOOTH (α-blend over N(u) for 1–2 iterations)     │
+│         → UPSERT Vectorize (smoothed)                     │
+│  + BUILD inverted index (D1 FTS5 / postings) for BM25     │
+└──────────────────────────────────────────────────────────┘
+```
+
+### 4.2 Online (worker) — every question
+
+```
+query
+  │
+  ├─ understanding (LLM) ────────── filters, terms, complexity, …
+  │
+  ├─ BM25 prefilter (FULL corpus inverted index) ── top N_bm25
+  │
+  ├─ dense Vectorize (smoothed vectors) ± metadata filter
+  │     + multi-query / HyDE fusion (keep)
+  │     + terminology graph lane (keep; restrained)
+  │
+  ├─ UNION candidate ids
+  │
+  ├─ RRF(dense-rank, bm25-rank)          ← now true hybrid recall
+  │
+  ├─ cross-encoder rerank → listwise (member/complex)
+  │
+  ├─ family pin / term boost / edition steer (keep)
+  │
+  ├─ CRAG grade → optional corrective re-retrieve (keep)
+  │
+  └─ generate + quote-anchor verify + reflect (keep)
+         citations follow usedHits (keep)
+```
+
+**Deliberately absent:** general neighbor-expansion re-ranker (ETSI
+negative). Agentic multi-hop remains Phase-3 for the synthesis class they
+flag as future work — not a substitute for the precision stack above.
+
+### 4.3 Mapping to #592 / Metanorma
+
+ETSI §II: the InfoModel is dictated by PDF extraction limits; they
+"envision that this simple Information Model can be extended/swapped…
+when dealing with normative content that is already available in a
+machine-readable format, such as is foreseen by the IEC Smart Standards
+framework."
+
+Our equivalent is the AI-serialization projection
+(`docs/METANORMA-AI-SERIALIZATION.md`):
+
+| ETSI InfoModel | #592 / ChunkRecordV2 |
+|---|---|
+| InfoUnit {title, body} | typed `nodes[]` (clause/table/formula/term/requirement/…) |
+| P parthood | `edges[]` kind=`part_of` + breadcrumb |
+| C citation | `edges[]` kind=`cites` + Relaton keys |
+| section code geometry | model anchors (stable, not OCR-guessed) |
+| tables atomic | `type=table` with columns/rows payload (conformance rule §8.3) |
+
+**Hand-rolled G-ETSI-2/3/5 now is the reference implementation that #592
+must emit.** We do not wait on upstream to close the retrieval gaps; we
+close them with the sources we have, and the projection is the durable
+contract.
+
+---
+
+## 5. Phased execution (eval-gated, ETSI-ordered)
+
+Every phase ships only if AP@K and MRR@K on the retrieval suite move the
+right direction on the golden + paraphrase probes. Recall@K is reported
+but is not the gate (ETSI: structure/RRF don't move it; only smoothing
+does, slightly).
+
+### Phase E0 — Measurement hygiene (1–2 days, no index change)
+1. Add AP@K, MRR@K, R@K to `tests/retrieval.mjs` with citation-span
+   matching (≥75% golden span containment — their protocol).
+2. Snapshot baselines on current production index.
+3. Stop treating hit@5 as the sole retrieval KPI in roadmap language.
+
+### Phase E1 — True hybrid lexical recall (G-ETSI-1) (3–5 days)
+1. Ingest: write `chunks_fts` (D1 FTS5) or posting lists from
+   `artifacts/chunks.jsonl` (enriched text).
+2. Query: BM25 top-N → union dense top-K → existing RRF → rerank.
+3. Gate: AP@K / MRR@K up on jargon-heavy cases (`n_LC`, part numbers,
+   "maximum permissible error", French/German term forms). Exact-term
+   paraphrase probes must not regress.
+
+### Phase E2 — Section graph + smoothing (G-ETSI-2/3/4) (1–2 weeks)
+1. Emit clause-level `part_of` + `cites` at graph build.
+2. Offline smooth embeddings (α sweep {0.7, 0.8, 0.9} × 1–2 iterations)
+   into a side index; A/B against current.
+3. Gate: recall@K up or flat with AP/MRR non-worse (smoothing is the
+   recall lever; do not accept AP regression).
+4. Do **not** enable neighbor-expansion re-rank in this phase.
+
+### Phase E3 — Atomic tables (G-ETSI-5) (re-ingest window, ≈$50)
+1. Upsert G1 table chunks; `chunk_type=table`; parent `part_of` to clause.
+2. Re-enrich + re-embed + smooth.
+3. Gate: table-value golden cases (R 60 n_LC, R 76 MPE, R 111 E2) AP/MRR
+   and answer correctness; no regression on prose cases.
+
+### Phase E4 — #592 producer path (upstream, parallel)
+1. Keep AI-serialization proposal aligned with the InfoModel above.
+2. Hand-rolled adapters remain the consumer of record until
+   metanorma-document emits the projection.
+3. When available: swap parse source, delete HTML/adoc recovery paths for
+   covered doctypes (OCP: new source = new adapter, not pipeline branches).
+
+### Phase E5 — Multi-hop synthesis (their future work = our G10)
+Only after E1–E3 are green. Deep-research / agentic loop for questions
+that need distant passages or cross-document assembly. Not a retrieval
+substitute.
+
+### Explicitly deferred / rejected by ETSI evidence
+- GNN graph embeddings (smoothing is the lightweight substitute they
+  validated).
+- General seed-neighbor expansion re-ranker (failed on ETSI).
+- Embedding-model upgrade as the primary lever (their stack used Granite;
+  wins came from structure + hybrid + smoothing).
+- Overlap chunking (not in their design; independent 2026 evidence also
+  weak — not reopened here).
+
+---
+
+## 6. What changes in our public claims
+
+- Retrieval quality story shifts from "dense + rerank + graph lane" to
+  **"structure-preserving index + full-corpus lexical prefilter + smoothed
+  dense + RRF + rerank"**, which is the ETSI-validated cascade for
+  standards.
+- Graph story shifts from "expand at query time" to **"parthood and
+  citation live in the index; relational signal is baked into vectors via
+  smoothing; query-time expansion stays restrained (terminology)"**.
+- Tables story is no longer optional polish — **atomic tables are a
+  normative-RAG requirement** (ETSI exempts them from chunking; STC
+  arXiv:2605.00318 independently shows MRR 0.36→0.59 / R@1 0.37→0.75 on
+  tabular structure, corroborating).
+- #592 is not a nice-to-have serialization — it is the **IEC-Smart-Standards
+  equivalent** their methodology section already assumes as the end state.
+
+---
+
+## 7. Sources (primary only)
+
+1. Al Masoud, Arazzi, Germani, Nocera. *Exploring Structural Complexity in
+   Normative RAG with Graph-based approaches: A case study on the ETSI
+   Standards.* arXiv:2604.09868, 2026-01.
+   https://arxiv.org/abs/2604.09868
+2. Guttal et al. *Structure-Aware Chunking for Tabular Data in RAG* (STC).
+   arXiv:2605.00318, 2026-05. (tables atomic / row structure — corroborates
+   G-ETSI-5 magnitude.) https://arxiv.org/abs/2605.00318
+3. Rayo, de la Rosa, Garrido. *A hybrid approach to information retrieval
+   and answer generation for regulatory texts.* 2025. (dense+sparse for
+   regulatory IR — ETSI's cited prior for RRF.)
+4. Cormack, Clarke, Buettcher. *Reciprocal rank fusion outperforms
+   condorcet and individual rank learning methods.* SIGIR 2009. (RRF.)
+5. IEC. *Smart Standards – from a market and industry perspective.* 2024.
+   (machine-readable standards end-state ETSI points at; our #592 analogue.)
+
+Corroborating structure-first work (not standards-specific, not used as
+primary authority here): Xu et al. RDR² arXiv:2510.04293; Yu et al. SF-RAG
+arXiv:2602.13647; SPIRE arXiv:2604.20849; Günther et al. late chunking
+arXiv:2409.04701; Merola & Singh contextual vs late arXiv:2504.19754.
+
+---
+
+## 8. Immediate next action
+
+Phase E0 + E1 start now. No re-ingest required for E0/E1 (FTS can be built
+from existing `artifacts/chunks.jsonl` / live chunk texts). E2/E3 need an
+index rebuild window and explicit go on the ≈$50 re-enrichment for E3.
+
+
+---
+
+## Postscript (2026-08-29): vision-unified answer contract
+
+User decision: unify text and figure understanding in ONE vision-enabled
+answer model instead of a forked vision lane.
+
+**Model**: `@cf/zai-org/glm-5.3-flash` — natively multimodal (320B/18B
+active), GLM family, flash tier, in the Workers AI catalog. Single answer
+model: prose + tables/equations as text, figures as IMAGE PARTS in the
+message content. Fallback ladder: qwen3.8-27b (text-only; figures degrade
+to their ingest-time description). Price unconfirmed (flash-tier
+expected) — eval-gate before defaulting the member lane.
+
+**Pipeline changes**
+
+1. INGEST (MKO stage 2.5): figure assets → R2 public bucket at stable
+   unit-keyed URLs; FigurePayload.uri rewritten to the canonical URL.
+   One-time vision captioning of EVERY figure into `description`
+   (quality-first lane; moondream 3.1 or glm-5.3-flash; ≈cents for the
+   clean corpus) so text-only tiers and the anon lane still explain
+   figures.
+2. SERVE: buildMessages appends an image part per figure unit among
+   usedHits when the routed model is vision-capable; understanding's
+   figure-presence signal routes to the VL model for members.
+3. OUTPUT contract unchanged: discursive prose + `[[u:]]` references;
+   figure blocks render <img> (canonical URL) + description; references
+   validated against usedHits; payloads served from KV by unit id — the
+   LLM still never re-types table/equation/figure DATA (input yes,
+   output by reference; verbatim quotes remain anchor-checked).
+4. EVAL GATE: new golden cases ("what does Figure A.x show?") must pass
+   with the VL model before the member lane defaults to it; TTFT watched
+   (image tokens add prefill).
+
+**Also from the same review (2026-08-29)**: answer contract v2 (SymGen-style
+symbolic references, arXiv:2311.09188; validation per JSONSchemaBench
+arXiv:2501.10868 — post-hoc, not constrained decoding, per
+arXiv:2405.21047), monthly cost table (CF-only ~$10/mo today; Modal CPU
+micro-service +$110-135/mo is the first justified step if TTFT
+consistency is product-critical; GPUs $580-2,840/mo eval-gated), and the
+MN 116 "consumer references & excerpts" clause to draft upstream.
+
+
+---
+
+## Status board (2026-08-29, end of day — the reconciled plan)
+
+| Plan item | Status |
+|---|---|
+| E0 ETSI metrics (R/AP/MRR + `--repeat`) | ✅ live |
+| E1 full-corpus FTS prefilter | ✅ live (R@5 +9pp) |
+| E2 section graph (parthood/cites/defines in D1) | ✅ live via MKO graph (3,561 section nodes) |
+| E2 embedding smoothing | ⏸ tested, inconclusive at current eval power, rolled back — revisit with `--repeat 3` baselines |
+| E3 atomic typed tables — clean corpus | ✅ live (MKO: 371 table chunks) |
+| E3 atomic typed tables — dirty corpus | ☐ task #20 (~$16 enrichment) |
+| E4 producer-native path (#592/MN 116/MKO) | ✅ MKO live for all 36 clean docs; upstream: preface-walk PR #49, MN 116 spec PR #9, incremental re-ingest PR #48; remaining: OIML pubid flavor, relations embedding, collections, multilingual alignment |
+| E5 deep-research (G10) | ◐ server v1 live (member-gated); UI pending — task #21 |
+| Vision-unified answers (glm-5.3-flash default) | ◐ model live all tiers; image-parts + contract v2 pending — task #22 |
+| Figure assets → R2 + vision captions | ☐ task #23 |
+| Answer contract v2 (SymGen refs + blocks + validation) | ☐ task #22 — the next big quality/rendering step |
+| TTFT pipeline surgery (parallelize/trim pre-generation) | ☐ **URGENT** — measured 25–32s (47s under rate-limit haze); the serial understand→grade chain is the bottleneck, not the model |
+| Modal CPU micro-service | deferred (conditional on TTFT consistency after surgery) |
+| glm-5.3-flash price confirmation | ☐ check spend ledger next day |
+| Turnstile | blocked on user sitekey |
+| 36/224 relaton families missing successor edges | upstream PRs, needs user authorization |
+
+Priorities: 1) TTFT surgery, 2) answer contract v2, 3) figures + captions,
+4) dirty tables, 5) research UI, 6) calm-window `--repeat 3` baseline +
+price check.