feat: cross-language Q&A (Operation Sandwich) + Windows quickstart — all default-OFF

Three workstreams, full suite 2482 passed, experimental paths default-OFF. #000055 — Windows quickstart without make tasks.py (pure-stdlib runner) + make.bat shim + .gitattributes; README Windows section rewritten. Quickstart needs only Python 3.10+ (no make/bzip2/curl/bash). Mirrors the Makefile quickstart subset; drift-pinned by tests/test_tasks_runner.py. #000001 §7 Phase 0 — deterministic cross-language guard arborist/qa/crosslang.py: non-English signal (¿/¡/non-ASCII) + an es function-word stoppack. Fail-closed to UNGROUNDED before retrieval/LLM (mirrors the quantifier reject-DAG) when no content token survives, else strips es stopwords from the retrieval query only. English path byte-identical by construction. Default OFF (crosslang_guard_enabled). Measured: the anarcocapitalismo field case 10.4s -> 1.6s. #000056 — Operation Sandwich (cross-language grounding) arborist/qa/mt/: opus-mt es/fr/ru<->en, lazy per-pair memoised singleton (fixes the 88%-engine-error concurrency defect), manifest-pinned, [mt] extra; entity_mask wrapper. Sandwich = translate query in (retrieval + LLM prompt) -> English answer -> UNTOUCHED verifier grounds English-vs-English -> translate the verified answer out as display-only (banner-labelled, zero grounding). question_hash + verifier_policy_hash invariant; MT engine identity binds into RetrievalPlan, not governance. CLI --crosslang-translate / make XLANG_MT=1. Default OFF; entity_mask default OFF (measured net-negative at bench scale). Fan-out bench (bench/*.py): Spanish ~0% -> 71% grounded vs the real no-support baseline; the round-trip predictor was tried and refuted; the entity-mask lever failed at scale (corpus-title anchoring untried). CLAUDE.md: cross-language bright-line convention + module map. Pre-existing modified diagram files are intentionally excluded.
2026-05-18 12:12:23 -04:00 · 2026-05-18 12:12:23 -04:00 · 2c98fc964e
commit 2c98fc964e
parent b711215f11
32 changed files with 4173 additions and 29 deletions
--- a/docs/TICKETS.md
+++ b/docs/TICKETS.md
@ -111,6 +111,8 @@ Newest first. Update on every open/close.

 | ID       | Title                                          | Status                | Opened     | Directive |
 |----------|------------------------------------------------|-----------------------|------------|-----------|
+| #000056  | Operation Sandwich — cross-language grounding via query+display MT | **implemented & landed 2026-05-17 · default-OFF** (fox: "call it operation sandwich, create a new ticket and finish it"). Mechanism + bright line **verified live end-to-end** (real Hermes + real opus-mt: es query → English answer+verifier → es display; `answer_text` English, `display_answer` Spanish additive, `question_hash`/`verifier_policy_hash` invariant; 6 tests + full suite 2477 passed, 0 regressions). **Fan-out measured (§9, n=1, 75 q):** EN baseline 85% → es+sandwich 71% = **−14pp cost**; transitions PRESERVED 31 / DOWNGRADE 16 / LOST 17 / N/A 11. A deterministic round-trip predictor was tried and **refuted** (12/17 LOST round-tripped CLEAN; another instance of the codified CLAUDE.md bench-maxing lesson — not re-added). LOST taxonomy from the artifact: ≈7 entity-translate (`Boltzmann`→"perntzmann", `Tarsus`→"Tarso"), ≈3 broad-enum, ≈several n=1 noise. **Lever built+validated:** `arborist/qa/mt/entity_mask.py` mask/restore (real opus-mt: `who is Paul of Tarsus?` "Pablo de Tarso"→**"Paul of Tarsus"**); default-ON within the default-OFF sandwich. Caveat: bench is lowercased so cap-detector lift is a **lower bound** (corpus-title anchor = v2). Also: fan-out caught + fixed an 88%-engine-error concurrency defect (per-call model load → memoised singleton + lazy per-pair); French + Russian breads added (manifest, `crosslang_source_lang`). **Lift measured 2026-05-17 (comparator corrected, fox):** the true baseline is the pre-ticket ≈0% (raw es query → song-title noise, UNGROUNDED, 10.4s) — NOT native English. Against that: **the sandwich is a large net win (≈0% → 71% es grounded, proof core never corrupted); the −14pp vs English is the cost of a new capability, not a regression — calling it a "fail" was a comparator error.** The genuine negative is the **entity-mask lever**: net-negative at scale (es 71→65, fr+mask 47/−38pp; isolated Paul-of-Tarsus win didn't replicate — 3rd bench-maxing-lesson instance), now **default-OFF** (`crosslang_entity_mask=False`); no-mask sandwich is the keeper. Recommendation flipped: **worth continuing (minus the mask)**, not park. Remaining: n=3, fr no-mask, corpus-title anchoring (only untried lowercase-capable detector). CLAUDE.md updated with the durable cross-lang *convention* only. Tasks #14–#17. Phase 1 of the #000001 §7 family; new ticket clears don't-proliferate (fox-directed + distinct Dav1d audience + architectural inflection: a model dependency `[mt]` + a presentation-translation layer — anticipated by #000001 §7's "split the `[mt]` model-distribution work like `[nli]`/vecpack"). **Sandwich:** translate query es→en (retrieval-side, == `--retrieval-keywords`, binds into `retrieval_plan_hash`, NOT `question_hash`) → English answer through the **byte-for-byte untouched verifier** → translate the verified English `answer_text` en→es into a NEW `display_answer` field, banner-labelled, zero grounding (the `_render_audit_label` render-projection pattern). Engine: local `[mt]` extra, Helsinki-NLP `opus-mt-es-en`/`-en-es`, Apache-2.0, hash-pinned, off-repo `~/.arborist/models/mt/`, optional dep, graceful-degrade — mirrors `[nli]`/`ShadowNLI` (#000049) + vecpack (#000051) verbatim; never Hermes-3-8B; not an external API (reproducibility + zero egress + es↔en is the best-resourced pair). Default OFF (`crosslang_translate_enabled`, gated under Phase-0 `crosslang_guard_enabled`); `--crosslang-translate` / `XLANG_MT=1`. Hash invariants (corrected 2026-05-17 — `governance_policy_hash` is sha256 of the *whole* policy, keys.py:182): `question_hash` + `verifier_policy_hash` untouched (user question preserved, verifier byte-identical); `governance_policy_hash` moves like every policy flag → correct cache partitioning by config (not a leak); MT engine identity binds into `RetrievalPlan.mt_*` (run-DAG), not a policy hash. | 2026-05-17 | — |
+| #000055  | Windows quickstart without `make` (`tasks.py` + `make.bat`) | **in progress** — opened 2026-05-16 (fox: "bat files or some shit … avoid needing makefile for windows … we will test the quickstart on windows"). Pure-stdlib `tasks.py` runner mirroring the **quickstart subset** of the Makefile (bootstrap / fetch-cur / ingest-cur-attached / distill ×2 / query / inspect / falsify / burn / bootstrap-crawler / crawl-ingest / stats / verify / search / clean) + a ~10-line `make.bat` shim so `make <target>` works in Windows cmd and `.\make.bat <target>` in PowerShell. Audited the artifact (not the docs): the 2003 dump is opened via stdlib `bz2` (no external `bzip2`); only `fetch-cur` used `curl` (→ stdlib `urllib`); bash `for…&wait` → `subprocess.Popen` fan-out; `arborist` console-script lands at `.venv\Scripts\arborist.exe`. Net: quickstart needs only **Python 3.10+ + sqlite3** — the repo's existing ethos, now true on native Windows. Same `KEY=VALUE` make-style args so documented commands translate 1:1 (one doc form). Makefile untouched, still canonical on POSIX ("keep it as an option"). Found + fixed a README/Makefile discrepancy: README claimed `[dev,html]` bootstrap extras, Makefile installs `.[dev]` — artifact wins. Drift-pinned by `tests/test_tasks_runner.py`. | 2026-05-16 | — |
 | #000054  | Acronym-parens concept extractor (closes the abbreviation→expansion retrieval gap) | **in progress** — Phase 1 (extractor + 481K edges) landed `58027e9`; Phase 2 (consumer-side surfacing — `synonym_expand` rank-and-truncate over the per-token cap, FTS5-`bm25` ordering in `_search_titles`, expanded `accept_tokens` in title-search + core-keyword + title-rerank, `synonym_expand_strict()` for the multiplicative title-purity rerank to exclude noisy `link_reciprocity` edges, tightened extractor regex to `[A-Z]{3,6}` purging 2-letter homonym edges) landed `ce855db`. **End-to-end verified:** `what is a CPU?` → Central processing unit at #1; `what is a GPU?` → Graphics processing unit at #1 EVIDENCE-WARRANTED 1/1; Mount Kilimanjaro / Soviet Union queries unchanged. **bench-qa n=3 limit=5** (2026-05-13T14:24Z): 30/45 STRICT (67%), zero regressions on basics (mona lisa / capital of france / new london bridge each 9/9 STRICT). 2026-05-13 — `arborist/concepts/extract.py:acronym_parens_synonym` lands as a new corpus-agnostic extractor in `EXTRACTORS` (`evidence_kind="acronym_parens"`). Scans each doc's lead chunk for `<Multi-Word Phrase> (ACRO)` where the all-caps acronym's letters match the content-word initials of the phrase in order; emits bidirectional synonym edges between the lowercased acronym and each ≥3-char content token of the phrase. Conservative (strict 1:1 initials, function words filtered, repeated definitions deduped per doc). Closes the *retrieval-side* abbreviation gap (`CPU↔central processing unit`, `GPU↔graphics processing unit`, `RAM↔random access memory`, `FBI↔federal bureau of investigation`, `WHO↔world health organization`, …) that `link_reciprocity_synonym` can't reach because the relation lives in body text, not the wiki link graph (Wikipedia represents abbreviation→expansion as a *redirect* — not an edge). Per-shard like all `concept_relations` data; corpus-agnostic so HTML/blogs/textbooks benefit equally. Retrieval-side only — never proof-path. 8 new tests; full suite green. Closes #000050 §2a's CPU/GPU fixture rows *upstream* of vec; the Orwell-shape conceptual-allusion row remains the genuine #000050 justification. Operational follow-up (not code): `arborist concepts derive --extractor acronym_parens` on each shard. | 2026-05-13 | —         |
 | #000053  | Acronym-aware verifier content tokens | **closed · 2026-05-13** — `arborist.qa.evidence._content_tokens` now keeps all-caps 2-3-char acronyms (CPU/GPU/DNA/FBI/USB…) as content tokens instead of dropping every <4-char token; fixes the field case where "what is a CPU?" cited to "CPU design" tripped `TITLE_MISMATCH` spuriously (claim & title share "CPU" but neither registered) — also affects `SUBJECT_TOKENS_ABSENT` (Rule 9), `BARE_NAME_CLAIM`, spotlight-excerpt token pick. Versioned: `content_token_rules: "v2-acronym-aware"` in both default policies + `_VERIFIER_POLICY_FIELDS` → folds into `verifier_policy_hash`, prior cache records orphan on lookup (by design, same discipline as `base_version` / `hyphen_fold_v1`). Monotone toward *fewer* spurious demotes (only relaxes overlap checks, never tightens). 8 new tests; full suite green; `bench-qa-smoke` clean. Does NOT fix the *retrieval* abbreviation→expansion gap (`CPU`→`Central processing unit` = #000050 vec hybrid / `concepts/` synonym edges — the root cause of the satellite-article retrieval). | 2026-05-13 | —         |
 | #000052  | Relevance + coherence meta-cognition (answer-*shape* signals) | in progress — **§3.1 `diagnose_coherence` landed** (lexical, no model: `circular` / `phrase_component_reuse` / `vacuous`; in `arborist/qa/inspect.py`, surfaced via `inspect_cache_key` + `arborist inspect` `· incoherent: <kind>`; 9 tests; demote-policy hook deliberately not wired — advisory only). Joins the `diagnose_deflection` / `diagnose_metaphor_deflection` / `diagnose_title_relevance` / soft-preflight family of read-only, demote-only, never-in-proof-path sidecars; `phrase_component_reuse` catches the motivating field case (a subject quoting a phrase, a predicate reusing one of that phrase's own tokens as a bare `the <token>` referent). **Still open: (2) `diagnose_relevance`** — semantic (not just lexical) "aboutness": does the answer address the question; is each claim about its cited source? Today's checks (subject-anchor token overlap, stemmed title-stem overlap) are *lexical* and a token collision defeats them — a small *aboutness/reranker* model (NOT NLI — entailment ≠ topicality) under #000049 §7's discipline cage verbatim (demotion-only, hash-pinned, `relevance_model_version`→`governance_policy_hash` iff it touches `audit_mode`, shadow-first, `[…]` extra, the §7 #20 haystack lesson — never over the whole context); gated on evidence, travels with #000049's model question. Motivating field case (2026-05-12, fox): the `claim_lattice` query that returned *"the phrase 'Zionist entity' is sometimes used as the entity, referring to the State of Israel"* at `EVIDENCE-WARRANTED-PARTIAL 2/3` — incoherent + token-collision recombination that NLI can't catch (returns *neutral*, not *contradiction*) and both lexical relevance checks waved through. Flags an upstream retrieval ticket (polysemy / title-token-soup) as the root-cause fix, not scoped here. #000049 sibling | 2026-05-12 | —         |
@ -164,8 +166,8 @@ Newest first. Update on every open/close.
 | #000004  | Directive coverage in bench summary            | closed · `acd1f9c`    | 2026-05-01 | D8        |
 | #000003  | Anchor-class warrant generalization (Module H+)| closed · 2026-05-02   | 2026-05-01 | D6        |
 | #000002  | Reference-Frame Polarity Contract (Module L)   | closed · 2026-05-02   | 2026-05-01 | D3        |
-| #000001  | Retrieval-keywords audit gap                   | closed · 2026-05-02   | 2026-05-01 | D4        |
+| #000001  | Retrieval-keywords audit gap (+ §7 cross-language transforms) | **reopened 2026-05-17 (in progress)** — keyword scope stays closed/landed 2026-05-02 (run-DAG `RetrievalPlan` binding, §5). §7 extends the *same* retrieval-transform-provenance substrate to a sibling: cross-language query bridging. Strict bright line — MT/aliases/stopword-guards *propose* candidates, only source-language evidence *warrants*; English answer verified English-vs-English by the **untouched binary verifier**, Spanish is display-only (zero grounding, banner-labelled). Held strictly this needs **no new `audit_mode` token, no `EVIDENCE-WARRANTED-CANONICAL`, no `surface_language` field, no `providence_cache` column, no new run-DAG stage** (the >10%-back deletion; CLAUDE.md schema-column-unchanged + verifier-stays-binary). **Phase 0 landed 2026-05-17** — `arborist/qa/crosslang.py` (deterministic, no model: `¿`/`¡`/non-ASCII signal, es-v1 stoppack); pre-preflight fail-closed mirroring `quantifier_should_reject` (Merkle-auditable reject DAG, `CROSS_LANGUAGE_UNSUPPORTED`, no LLM) + es-stoppack strip on `retrieval_query` only. Measured: `¿Qué es el anarcocapitalismo?` **10.4 s → 1.6 s** (~6.4×, still honest UNGROUNDED); English control byte-identical (by construction — `guard()` returns None for pure-ASCII). 19 tests; full suite 2470 passed, 0 regressions (incl. #000053/#000054); `bench-qa-smoke` stable anchor 3/3 STRICT. No schema change, no new governance/verifier *fields*, no `RetrievalPlan` change; `question_hash` + `verifier_policy_hash` untouched (the enabling flag moves `governance_policy_hash` like every policy flag — whole-policy hash, correct cache partition; corrected 2026-05-17, authoritative in #000056 §2 #6). `RetrievalPlan` MT extension landed in #000056 (Phase 1). **Feature-flagged default-OFF** (fox request): `policy["crosslang_guard_enabled"]` gates both seams via one point → flag-off reverts byte-for-byte to legacy (clean A/B); surface `arborist query --crosslang-guard` / `make query XLANG=1` / `tasks.py XLANG=1`. Default-flip is a separate bench-gated fox decision (rollout discipline matches #000008/#000011/#000049). 20 tests; full suite 2471 passed. **Collision constraint:** the guard is a specific source-language function-word stoppack, NOT a len≤2 heuristic — a blanket short-token drop regresses #000053/#000054 (`AI`/`ML`/`CPU`/`GPU` are load-bearing); `_FTS5_STOPWORDS`/`_TITLE_STOPWORDS` stay in sync; n=3 English bench, no >5pp regression. **Phase 1 split out → #000056 "Operation Sandwich"** (fox-directed; MT provider decided = local `[mt]` opus-mt, hash-pinned, not Hermes/not-API). Deleted by five-step (recorded in §7.6): raw-Spanish route (measured noise vs `en`), alias-map substrate (hand-curated worse-MT), cross-modal/multilingual-vec/SQD-language/ABCDEFG-5S (no named defect — the retired `arborist/v7/` anti-pattern). | 2026-05-01 | D4 |

 ## Next ID

-`000055`
+`000057`
--- a/docs/tickets/ticket-000001-retrieval-keywords-audit-gap.md
+++ b/docs/tickets/ticket-000001-retrieval-keywords-audit-gap.md
@ -1,15 +1,21 @@
 # Ticket #000001 — Retrieval-keywords audit gap

-**Status:** closed · landed 2026-05-02 (run-DAG binding scope; SQL
-column + audit-events scope deferred per §6 below)
+**Status:** reopened 2026-05-17 (in progress) — keyword scope stays
+closed/landed 2026-05-02 (run-DAG binding, §5; SQL column + audit-events
+deferred per §6). §7 extends the *same retrieval-transform-provenance
+substrate* (`arborist/qa/retrieval_plan.py`) to a sibling transform:
+cross-language query bridging. Doc-only; gated on the MT-provider
+decision (§7.5).
 **Opened:** 2026-05-01
-**Closed:** 2026-05-02
-**Scope:** Design proposal for capturing the `--retrieval-keywords` operator
-hint in the v9.8 audit chain so retrieval is fully reproducible from a
-providence record alone. Doc-only — no code in this commit.
+**Closed:** 2026-05-02 (keyword scope) · **Reopened:** 2026-05-17 (§7)
+**Scope:** Capturing retrieval-side transforms (operator keywords §1–§6;
+cross-language bridging §7) in the v9.8 audit chain so retrieval is
+reproducible from a providence record alone. Doc-only.
 **Audience:** fox + future blackops shifts.
-**Hard constraint:** keywords stay operator-metadata, not part of the user's
-question. Cache-key dimensionality stays at 8.
+**Hard constraint:** retrieval transforms (keywords, MT, stopword
+guards) stay operator/system metadata, never part of the user's
+question. Cache-key dimensionality stays at 8. The verifier stays
+binary; no new `audit_mode` token, no `providence_cache` column.

 ---

@ -288,6 +294,232 @@ in directly) stays in the original proposal as future work.

 ---

+## 7. Phase X — Cross-language retrieval transforms (reopened 2026-05-17)
+
+**Why here, not a new ticket.** MT-to-corpus-language and a
+multilingual stopword guard are *retrieval transforms* — the same
+family as `--retrieval-keywords`. §5 already shipped the substrate:
+`RetrievalPlan` + `retrieval_plan_hash`, run-DAG-bound, outside
+`question_hash`/`cache_key`. A cross-language transform is captured by
+*extending the `RetrievalPlan` fields*, not by inventing a new
+provenance mechanism. Per the don't-proliferate discipline this is
+§000001 Phase X, not #0000XX. (The 2026-05-16 field case:
+`¿Qué es el anarcocapitalismo?` over the en-2003 corpus burned 10.4 s
+assembling title-collision noise — *Así Es El Amor*, *¿Quién es el
+Jefe?* — then correctly returned UNGROUNDED. The fail-closed was
+right; the 10.4 s waste and the missing recall bridge are the gap.)
+
+### 7.1 Hard constraints (the bright line — strict form deletes a design layer)
+
+```text
+MT / aliases / stopword guards  →  may PROPOSE retrieval candidates
+source-language evidence        →  the ONLY thing that WARRANTS
+```
+
+Valid: Spanish query → MT → **English** retrieval → **English** answer
+→ **existing untouched binary verifier** (English-vs-English) →
+existing label, unchanged. Spanish text, if rendered, is **display
+only**, carries zero grounding, and is banner-labelled
+"machine-translated, not verified" — the `_render_audit_label`
+pattern (render projection ≠ proof).
+
+Holding this strictly is the >10%-back deletion: there is **no new
+`audit_mode` token, no `EVIDENCE-WARRANTED-CANONICAL`, no
+`surface_language` proof field, no `providence_cache` column, no new
+run-DAG stage.** CLAUDE.md: *schema column stays unchanged; verifier
+stays binary; audit_mode decided by the verifier, never asserted.*
+
+Forbidden (→ the #000049 model-in-proof-path cage, not here):
+Spanish answer → translate-to-English → verify → call it grounded.
+
+### 7.2 Collision constraint (must not regress shipped work)
+
+A noise guard **must be a specific source-language function-word
+stoppack** (es v1: `qué que es el la los las un una de del en y o a`),
+**not** a length / "drop len≤2" heuristic. A blanket short-token drop
+regresses #000053 (acronym-aware verifier content tokens) and #000054
+(acronym_parens edges) — `AI`/`ML`/`CPU`/`GPU`/`DNA`/`FBI` are
+load-bearing 2–3-char tokens. `_FTS5_STOPWORDS` and `_TITLE_STOPWORDS`
+**stay in sync** (CLAUDE.md retrieval-pipeline §9). Bench-gated:
+`make bench-qa` n=3, English suite, no >5pp STRICT-rate regression
+before anything lands.
+
+### 7.3 Phase 0 — grounded design (2026-05-17, real-code pass)
+
+Mapping `arborist/qa/query.py` + `arborist/search/fts5.py` corrected
+four points in the draft. The refinements make Phase 0 *smaller*:
+
+1. **`_FTS5_TOKEN_RE = [A-Za-z][A-Za-z0-9]*` is ASCII-only.** `¿Qué`
+   tokenizes to `Qu`; accented words are truncated before the
+   stopword check. The es stoppack must therefore list ASCII-folded /
+   truncated forms: `qu que es el la los las un una de del en y o a`.
+2. **The measured 9.9 s cost was the search, not context assembly**
+   (`search 9.91s` · `context 0.05s`). The cost was OR-mode FTS5
+   driven by the es function words `es`/`el` (high-DF, full-corpus
+   union scan). So Phase 0's primary lever is **stripping the es
+   stoppack from the retrieval token set**, not the fail-closed —
+   the field case (`anarcocapitalismo` survives the stoppack) is a
+   *content token that misses the corpus*, not a no-content query;
+   stripping `es`/`el` makes the search fast (single rare token, no
+   OR scan) → honest `UNGROUNDED` in ~ms instead of 10.4 s.
+3. **The es stoppack must be language-scoped, never global.**
+   `un`→UN, `la`→LA, `de`→De, `en`→EN collide with the #000053/#000054
+   acronym/abbrev class. Gate it behind a deterministic
+   no-model **non-English signal**: the query contains inverted
+   punctuation (`¿`/`¡`) or a non-ASCII Latin letter. Pure-ASCII
+   English queries never trigger it → the English retrieval path is
+   **byte-identical by construction** (the bench is a no-op by proof,
+   not just by measurement).
+4. **Phase 0 needs no `RetrievalPlan` change.** The explicit
+   fail-closed returns *before* `RetrievalPlan` is built (it reuses
+   the existing `build_reject_run_dag` 3-stage path, exactly like
+   `quantifier_should_reject`). The reason is recorded as a
+   `CROSS_LANGUAGE_UNSUPPORTED` violation in the reject run-DAG
+   (mirroring `BROAD_QUANTIFIER_REJECTED`). Extending `RetrievalPlan`
+   with `detected_language`/`dropped_tokens` is a **Phase 1**
+   concern (when MT retrieval actually runs and the plan is built).
+
+**Phase 0 behaviour, final:** at `query()` entry, if the non-English
+signal (refinement 3) fires: strip the es/multilingual stoppack from
+the retrieval token set for this call only (refinement 1); if **zero**
+content tokens remain, short-circuit to `UNGROUNDED` before retrieval
+via the reject-DAG path (refinement 4); otherwise run retrieval on the
+surviving content tokens (fast — refinement 2) and let the existing
+pipeline return its honest verdict. Adds **no new governance/verifier
+fields** and **no `RetrievalPlan` change**; `question_hash` and
+`verifier_policy_hash` untouched; English path byte-identical.
+(Correction 2026-05-17: an earlier draft said "no governance change"
+unqualified — false. `governance_policy_hash` is sha256 of the
+*whole* policy (`keys.py:182`), so toggling the `crosslang_guard_enabled`
+flag moves it like every policy flag. That is correct cache
+partitioning, not a leak. Authoritative statement: #000056 §2 #6.)
+
+### 7.4 Phase 1 — MT-to-corpus-language as one soft route (gated)
+
+MT-translated query == `--retrieval-keywords` with an MT front-end:
+folds into `retrieval_plan_hash` exactly like keywords (the §2.4 /
+§A.5 decision verbatim — **A+B, not C**; translated query is
+retrieval metadata, **never** `question_hash`). Governance: the MT
+*engine identity* binds into `RetrievalPlan` (run-DAG), not a policy
+hash; the *enabling flag* moves `governance_policy_hash` like every
+policy flag (whole-policy hash — correct cache partition). **This
+§7.4 sketch is superseded by #000056 (Operation Sandwich), which is
+the authoritative Phase-1 design + the corrected hash invariants
+(§2 #6).** Raw `es` and raw `en`
+questions keep distinct `question_hash` (no cross-language dedup).
+
+### 7.5 The gate (fox decides — nothing in 7.4 lands until then)
+
+MT provider. **Not Hermes-3-8B** (blackops standing rule: 8B unfit
+for human-language translation; cannibalizes coder-agent budget).
+Fork: **Grok API** (flagged candidate; soft-route only;
+provider/version/prompt into `governance_policy_hash`) **vs** a local
+**`[mt]` extra** (NLLB/Marian, hash-pinned like `[nli]`, weights
+off-repo under `~/.arborist/models/mt/<hash>`). Phase 1 stays
+doc-only until chosen.
+
+### 7.6 Deleted by the five-step (named, with the reason)
+
+- **Raw-Spanish retrieval route** — measured pure noise vs an `en`
+  corpus (the song-title case). Revisit *only* when a Spanish/mixed
+  shard exists; not a v1 route.
+- **Deterministic alias-map substrate** — a hand-curated, worse MT
+  with no named maintainer; strictly dominated by a Spanish corpus
+  or the MT route. Cut.
+- **New labels / `EVIDENCE-WARRANTED-CANONICAL` / `surface_language`
+  proof field / two-level proof identity** — label + schema
+  proliferation the strict §7.1 line makes unnecessary.
+- **Cross-modal carriers, multilingual vec, SQD language
+  canonicalizers, a new `language_bridge` run-DAG stage, 5S/ABCDEFG
+  suites** — no named person, no closed defect (five-step step 1).
+  Namespace-reservation now = the retired `arborist/v7/` anti-pattern.
+
+### 7.7 Acceptance criteria (Phase 0 testable now; Phase 1 post-gate)
+
+1a. Cross-language query whose tokens are **all** function words →
+    `UNGROUNDED` **before** retrieval/LLM via the reject-DAG path;
+    `CROSS_LANGUAGE_UNSUPPORTED` violation recorded.
+1b. Cross-language query with a surviving (corpus-missing) content
+    token (the `anarcocapitalismo` field case) → es function words
+    stripped from the retrieval set → **no OR-mode full-corpus
+    scan** (search ≪ the measured 9.9 s) → honest `UNGROUNDED`.
+2. Source-language function words do not drive FTS5 into
+   title-collision noise.
+3. `make bench-qa` n=3 English suite: **no >5pp STRICT-rate
+   regression** (the #000053/#000054 guard).
+4. MT route (when enabled) folds into `retrieval_plan_hash`, **not**
+   `question_hash`.
+5. `governance_policy_hash` changes **iff** MT enabled AND cacheable;
+   unchanged otherwise.
+6. **No** new `audit_mode` token; **no** `providence_cache` column;
+   verifier byte-identical.
+7. Raw `es` and raw `en` of the same question keep distinct
+   `question_hash`.
+
+### 7.8 Status
+
+**Phase 0 landed 2026-05-17.** New module `arborist/qa/crosslang.py`
+(deterministic, no model: `guard()` + `strip_for_retrieval()`,
+es-v1 stoppack, non-English signal = `¿`/`¡`/non-ASCII letter).
+Two seams in `arborist/qa/query.py`: a pre-preflight fail-closed
+early-return mirroring `quantifier_should_reject` (reuses
+`build_reject_run_dag` → 3-stage Merkle-auditable reject DAG,
+`CROSS_LANGUAGE_UNSUPPORTED` violation, `status=
+cross_language_unsupported`, no LLM), and an es-stoppack strip on
+`retrieval_query` only (LLM / verifier / `question_hash` untouched).
+No schema change, no new governance/verifier *fields*, no
+`RetrievalPlan` change; `question_hash` + `verifier_policy_hash`
+untouched. (The enabling flag moves `governance_policy_hash` like
+every policy flag — whole-policy hash, correct cache partition; see
+#000056 §2 #6 for the authoritative corrected statement.)
+
+Measured (live, `query-dry`): `¿Qué es el anarcocapitalismo?`
+**10.4 s → 1.6 s** (search 9.91 s → 1.26 s, ~6.4×), still honest
+UNGROUNDED; `¿Qué es el?` → `cross_language_unsupported` pre-retrieval;
+English control (`What is anarcho-capitalism?`) byte-identical
+(Anarcho-capitalism #1, unchanged). 19 new tests in
+`tests/test_crosslang_guard.py`; full suite 2470 passed, 0
+regressions (incl. query/dag/claim_lattice/#000053/#000054);
+`bench-qa-smoke` stable anchor (`mona lisa`) 3/3 STRICT. English
+inertness is **by construction** (the signal regex cannot match a
+pure-ASCII no-`¿`/`¡` query → `guard()` returns None → identical
+path), so the bench is a proof, not just a measurement.
+
+**Feature-flagged (fox request, 2026-05-17): default OFF.**
+`policy["crosslang_guard_enabled"]` (default `False`) gates BOTH
+seams via a single point — flag off → `_xlang` is None → behaviour
+reverts byte-for-byte to pre-§7, a clean A/B baseline. Surface:
+`arborist query --crosslang-guard`, `make query XLANG=1`,
+`tasks.py query XLANG=1`. Rollout discipline matches #000008
+(quantifier) / #000011 (soft-preflight) / #000049 (NLI): ships
+default-OFF; the default-flip is a separate **bench-gated fox
+decision**, not autonomous — even though Phase 0 is provably inert
+on English and strictly improves the cross-language case (the
+argument *for* a future flip, deferred to fox). Retrieval-side
+only: the flag does NOT enter `question_hash` or
+`verifier_policy_hash` (user question preserved, verifier
+byte-identical). It DOES move `governance_policy_hash` like every
+policy flag (whole-policy hash, `keys.py:182`) — correct cache
+partitioning by guard state, not a leak; the fail-closed path writes
+no cache row anyway. (Correction 2026-05-17; authoritative: #000056
+§2 #6.) Live
+A/B verified: `¿Qué es el?` flag-OFF → legacy UNGROUNDED;
+flag-ON → `cross_language_unsupported` pre-retrieval. +1 test
+(`test_flag_off_reverts_to_legacy_behaviour`); 20 tests total;
+full suite 2471 passed.
+
+**Phase 1 split out → #000056 "Operation Sandwich" (fox-directed
+2026-05-17).** MT-provider decided: local `[mt]` extra,
+`opus-mt-es-en`/`-en-es`, hash-pinned (not Hermes, not an API). The
+query+display MT sandwich, the `RetrievalPlan` MT-field extension
+(the §7.3-deferred item), and the `[mt]` model dependency all live
+in #000056 — split per the don't-proliferate "distinct Dav1d
+audience / architectural inflection" criterion (a model dependency +
+a presentation layer), exactly the split this section anticipated.
+#000001 §7 stays the Phase-0 home.
+
+---
+
 ## Appendix A — Architectural review (2026-05-01, Asia/Kuala_Lumpur)

 > Fox-supplied review expanding §2-§4 with axiomatic framing,
--- a/docs/tickets/ticket-000055-windows-quickstart-no-make.md
+++ b/docs/tickets/ticket-000055-windows-quickstart-no-make.md
@ -0,0 +1,113 @@
+# Ticket #000055 — Windows quickstart without `make` (`tasks.py` + `make.bat`)
+
+**Status:** in progress — opened 2026-05-16, implementation in same commit.
+**Opened:** 2026-05-16
+**Asked by:** fox ("we need bat files or some shit so we could avoid
+needing makefile for windows (but keep it as an option) windows uses
+py py3 for python3 etc we will test the quickstart on windows so fix
+it up").
+**Scope:** One new pure-stdlib runner `tasks.py` at repo root + one
+small `make.bat` Windows shim + README Windows/Quickstart rewrite +
+one smoke/drift test. **No** change to `arborist/`, the schema, the
+Makefile recipes, or any audit/proof path. The Makefile stays the
+canonical Unix entry point and is left byte-for-byte unchanged ("keep
+it as an option").
+**Audience:** fox + anyone running the quickstart on native Windows +
+maintainers who touch the quickstart and must keep two thin entry
+points in sync.
+
+---
+
+## 1. The gap, from the field
+
+`README.md` Setup §Windows currently says native Windows is *not*
+supported — WSL2 only — because the Makefile uses bash idioms
+(`for i in $(seq …); do … & done; wait`, `sed`, backgrounding) and
+GNU-make conditionals, and the docs assume `make`, `curl`, `bzip2`.
+
+fox wants the **quickstart** runnable on native Windows (cmd /
+PowerShell), no `make`, no WSL. Windows ships the `py` launcher
+(`py -3`), not `python3`.
+
+## 2. What actually blocks Windows (and what does not)
+
+Audited the quickstart targets against the artifact, not the docs:
+
+- **Decompression:** the 2003 dump is opened with Python's stdlib
+  `bz2` (`arborist/sources/wikipedia.py:288`). **No external `bzip2`
+  needed** — the README's "needs … `bzip2`" line is stale for the
+  quickstart path.
+- **Download:** the only `curl` use in the quickstart is `fetch-cur`.
+  Replaceable with stdlib `urllib.request` → no `curl` dependency.
+- **Sharding/distill loops:** bash `for … & wait`. Replaceable with
+  `subprocess.Popen` fan-out + wait.
+- **CLI:** `arborist` is a console script (`arborist.cli:main`,
+  `pyproject.toml:120`). On a Windows venv it lands at
+  `.venv\Scripts\arborist.exe`; on POSIX `.venv/bin/arborist`.
+- **Net:** after this, the quickstart needs only **Python 3.10+ and
+  sqlite3** (sqlite3 ships with CPython) — exactly the repo's
+  "fresh checkout needs only python3.12 + venv + sqlite3" ethos.
+  No `make`, no `bzip2`, no `curl`, no `bash`.
+
+## 3. Design — why one runner, not N `.bat` files
+
+Five-step algorithm, step 2 (delete the part): hand-porting ~20 bash
+recipes into brittle per-target `.bat` scripts would (a) duplicate
+logic 1:1 with the Makefile in a second fragile dialect, (b)
+proliferate files, (c) drift fast (DRY-in-context). Instead:
+
+- **`tasks.py`** — one pure-stdlib runner implementing the quickstart
+  *subset only*. Cross-platform (also runs on POSIX as
+  `python3 tasks.py …`, but the Makefile remains the documented Unix
+  path). Detects venv layout (`Scripts` vs `bin`) and the Python
+  launcher (`py -3` on Windows, `python3` on POSIX; `ARBORIST_PYTHON`
+  overrides). Accepts the **same `KEY=VALUE` token style as `make`**
+  (`query Q="…" JSON=1`, `inspect KEY=… JSON=1`, `crawl-ingest
+  URL=… DEPTH=2`) so the documented commands translate 1:1 — one doc
+  form, both platforms.
+- **`make.bat`** — a ~10-line dispatcher so `make <target>` keeps
+  working in Windows `cmd` (cmd searches the current dir for
+  `make.bat`) and `.\make.bat <target>` in PowerShell. It only
+  resolves a Python (`py -3`, fallback `python`) and forwards argv to
+  `tasks.py`. No logic in the `.bat`.
+
+This keeps the Windows surface to **two files, one of them Python**,
+mirrors the existing "Makefile is build glue, logic is Python"
+precedent, and gives Windows the exact `make`-style UX.
+
+`tasks.py` is the quickstart subset, not the full 1170-line Makefile.
+Bench/π*/textbook/docs targets stay Makefile-only (Unix/WSL2). The
+runner's `help` enumerates exactly the supported targets so the
+boundary is self-documenting.
+
+## 4. Targets covered
+
+`bootstrap fetch-cur ingest-cur-attached distill-shards-parallel
+distill-shards-tfidf-parallel query query-dry inspect falsify burn
+burn-kindergarten bootstrap-crawler crawl-ingest recrawl-check stats
+stats-shards verify search test help clean clean-db clean-data` —
+the full README Quickstart (Wikipedia + crawl paths), "After the
+answer", and Setup blocks.
+
+## 5. Discrepancy found (reported to fox)
+
+`README.md` Setup §Bootstrap claims bootstrap installs the
+`[dev,html]` extras; `Makefile:52` installs `.[dev]` only. Artifact
+wins (CLAUDE.md "artifact over instruction"): `tasks.py bootstrap`
+installs `.[dev]` to match the Makefile. README corrected in the same
+commit.
+
+## 6. Drift control
+
+`tests/test_tasks_runner.py`: importable on the host platform;
+`help` lists exactly the documented quickstart targets (a removed or
+renamed target fails the test); venv-layout + launcher detection
+verified for both `os.name == "nt"` and `"posix"` via monkeypatch.
+Two thin entry points are a known DRY cost — the test pins the target
+set so divergence from the documented quickstart is loud.
+
+## 7. Out of scope
+
+Full Makefile parity on Windows; bench/π*/NLI/textbook/docs targets;
+changing the Unix workflow (Makefile untouched, still canonical on
+POSIX).
--- a/docs/tickets/ticket-000056-operation-sandwich-crosslang-grounding.md
+++ b/docs/tickets/ticket-000056-operation-sandwich-crosslang-grounding.md
@ -0,0 +1,337 @@
+# Ticket #000056 — Operation Sandwich (cross-language grounding via query+display MT)
+
+**Status:** implemented & landed 2026-05-17 — mechanism + bright
+line verified live end-to-end (§8); 6 tests + full suite 2477
+passed. **Default OFF** (`crosslang_translate_enabled`, gated under
+Phase-0 `crosslang_guard_enabled`). NOT "cross-language solved":
+the live flagship run is still `UNGROUNDED` (opus-mt
+output↔title-normalisation recall gap, §8 limit 1) — recall tuning +
+JSON-display polish (§8 limit 2) + SHA-pinning the manifest are open
+follow-ups behind the experiment flag; a bench-gated default-flip is
+a fox decision. Phase 1 of the #000001 §7 family.
+**Opened:** 2026-05-17
+**Asked by:** fox ("yes call it operation sandwich create a new
+ticket and finish it") — closing the chain that started with the
+2026-05-16 `¿Qué es el anarcocapitalismo?` field case.
+
+**Why a new ticket (clears the don't-proliferate bar).** CLAUDE.md
+default is *extend*, split only when the piece needs a **distinct
+Dav1d-reviewable audience** or is an **architectural inflection**.
+Operation Sandwich is both: it introduces a **model dependency**
+(`[mt]`) and a **presentation-translation layer** — a self-contained
+design decision a de-novo review must read independently of #000001's
+retrieval-keywords provenance story. #000001 §7 explicitly anticipated
+this split ("If the MT model choice becomes large, split the
+model-distribution work into a later `[mt]` ticket, similar to
+`[nli]` and vecpack"). #000001 §7 stays the Phase-0 home; this is
+Phase 1. Cross-linked both ways.
+
+**Audience:** fox + a Dav1d de-novo review of the
+translation-in-the-pipeline boundary + maintainers of the verifier
+proof path.
+
+---
+
+## 1. The sandwich
+
+```
+Spanish query ─▶ [MT es→en] ─▶ English retrieval + English answer + EXISTING verifier ─▶ [MT en→es] ─▶ Spanish display
+                 translator IN        ←—— grounded core, byte-for-byte untouched ——→        translator OUT
+                 (retrieval-side)      (English-vs-English — the proof)                    (presentation-only)
+```
+
+Translation lives on the **two edges**, never the middle. The
+verifier, `audit_mode`, `cache_key`, `question_hash`,
+`governance_policy_hash` all see only the English core. The Spanish
+the user reads is a labelled rendering carrying **zero** grounding.
+
+## 2. Hard constraints (the bright line, as code invariants)
+
+1. **`answer_text` stays the grounded English string.** The verifier
+   runs English-answer-vs-English-source, unchanged byte-for-byte.
+2. **The Spanish rendering goes in a NEW field** (`display_answer`,
+   `display_lang`, `display_translated=True`,
+   `display_unverified_banner`). It is never the verifier's input,
+   never what `audit_mode`/STRICT attaches to, never hashed into the
+   proof. Same discipline as `cli._render_audit_label` (render
+   projection ≠ proof).
+3. **Query MT is retrieval-side, == `--retrieval-keywords`.** It
+   changes which sources rank (→ `context_root` → `cache_key`
+   indirectly, exactly the #000001 §5/§6 story), **not**
+   `question_hash`. The MT engine identity binds into the **run-DAG
+   retrieval stage** via the existing `retrieval_plan_hash` (the
+   §7.3-deferred `RetrievalPlan` extension lands here).
+4. **No translated answer ever re-enters the verifier** (the
+   forbidden path / #000049 model-in-proof-path cage). Display MT is
+   output-only and one-directional.
+5. **Default OFF**, gated under `crosslang_guard_enabled`. The MT
+   engine is the optional `[mt]` extra; absent → degrade to
+   `available=False` → the sandwich silently no-ops (query proceeds
+   as Phase-0). English path byte-identical by construction.
+6. **Hash invariants (corrected 2026-05-17 — artifact over my own
+   prose).** `governance_policy_hash` is `sha256` of the *whole*
+   policy dict (`arborist/qa/keys.py:182`), so — like **every**
+   policy flag (quantifier, metacognition, soft-preflight) — the
+   crosslang flags DO move it. An earlier draft of this section
+   wrongly said "governance untouched"; that was false and is
+   corrected here. The behaviour is *correct*: a sandwich-on answer
+   must not be served to a sandwich-off lookup — the cache partitions
+   by config, it does not leak. The load-bearing invariants are the
+   two that ARE untouched: **`question_hash`** (the user's Spanish
+   question is preserved verbatim — MT never rewrites what was asked)
+   and **`verifier_policy_hash`** (the verifier is byte-identical;
+   the crosslang flags are absent from `_VERIFIER_POLICY_FIELDS`).
+   The MT *engine identity* binds into the run-DAG retrieval plan
+   (`RetrievalPlan.mt_*`), which is not a policy hash. Test:
+   `test_sandwich_hash_invariants`.
+
+## 3. Engine — local `[mt]` extra (fox-decided 2026-05-17)
+
+`opus-mt-es-en` (query in) + `opus-mt-en-es` (display out),
+Helsinki-NLP, Apache-2.0, hash-pinned by HF revision, weights
+off-repo under `~/.arborist/models/mt/`, optional dependency. **Not
+Hermes-3-8B** (blackops rule: 8B unfit for human-language
+translation). **Not an external API**: reproducibility (a pinned
+checkpoint replays byte-identically — provenance-critical for a
+Merkle-replay system) + zero egress (Operation Voyeur / sovereignty)
+ es↔en is the best-resourced MT pair so the local-vs-frontier
+quality gap is smallest exactly here. Grok stays the documented
+future bench-comparison candidate (blackops: "evaluate before
+committing spend"), not the v1 substrate. Mirrors the `[nli]` /
+`ShadowNLI` pattern (#000049) and the vecpack off-box pattern
+(#000051) exactly — zero new architectural surface.
+
+## 4. Build
+
+- `arborist/qa/mt/` — `manifest.json` (pinned es-en + en-es repos +
+  revisions), `Translator` protocol, `OpusMTTranslator` (lazy
+  `transformers` import, `available` flag, graceful degrade — the
+  `ShadowNLI` template verbatim), `StubTranslator` (deterministic;
+  the test substrate **and** the safe fallback when the extra is
+  absent — identity passthrough, so the sandwich degrades to Phase-0
+  behaviour, never raises).
+- `[mt]` extra in `pyproject.toml`.
+- `arborist/qa/query.py` — under
+  `crosslang_guard_enabled AND crosslang_translate_enabled` AND the
+  Phase-0 non-English signal fired AND translator `available`:
+  es→en the query → `retrieval_query` **and** the LLM prompt
+  (`llm_question`, so the model answers in English; `question_hash`
+  still uses the original Spanish); after the verified English
+  result, en→es the model **prose** (`raw_answer`, never the rendered
+  evidence scaffold — pinned verbatim English spans must not be
+  MT-ed) → `display_answer` (+ banner). Extend
+  `RetrievalPlan` with `mt_engine` / `mt_manifest_hash` /
+  `source_lang` (omitted-when-empty so existing `retrieval_plan_hash`
+  values do not churn — the §5 zero-churn discipline).
+- CLI `--crosslang-translate`; `make query XLANG_MT=1`;
+  `tasks.py query XLANG_MT=1`.
+
+## 5. Acceptance criteria
+
+1. Sandwich path (StubTranslator): es query → en `retrieval_query`
+   (asserted), English `answer_text` verified by the **unmodified**
+   verifier, `display_answer` is the es rendering, banner present.
+2. `question_hash` and `verifier_policy_hash` **identical** with the
+   flag on vs off (the load-bearing invariant: user-question identity
+   preserved, verifier byte-identical). `governance_policy_hash`
+   **differs** (whole-policy hash — like every policy flag; correct
+   cache partitioning, not a leak). Test:
+   `test_sandwich_hash_invariants`.
+3. `RetrievalPlan` MT fields bind into `run_dag` (different engine →
+   different `retrieval_plan_hash`); empty MT fields → **unchanged**
+   hash (no churn on non-MT runs).
+4. The translated Spanish text is **never** the verifier's input —
+   `answer_text` (English) is finalised and verified before display
+   MT runs; the rendered evidence scaffold (pinned verbatim English
+   spans) is never fed to MT (display translates `raw_answer` prose).
+5. Default OFF: `crosslang_translate_enabled=False` → no `display_*`,
+   pure Phase-0 behaviour. English query → byte-identical (signal
+   None → translator never consulted).
+6. `[mt]` absent → `OpusMTTranslator.available is False` → sandwich
+   no-ops to Phase-0, never raises (graceful-degrade test).
+7. The translated answer is **never** passed to `verify_*`
+   (asserted structurally — the verifier input is `answer_text`,
+   set before display MT runs).
+
+## 6. Out of scope
+
+Languages beyond es-v1; the Grok bench-comparison; an
+`audit_mode`-affecting strict mode that folds MT into
+`governance_policy_hash` (the #000001 §6 optional-strict analogue);
+auto language *detection* beyond the Phase-0 deterministic signal
+(accent-free Spanish still slips — a detector is future work).
+
+## 7. Status
+
+Implementation landed 2026-05-17 — see §8.
+
+## 8. Landing — what works, and the honest limits
+
+**Mechanism: complete and verified live end-to-end.** A real run
+(`arborist query --crosslang-translate --json "¿Qué es el
+anarcocapitalismo?"` against live Hermes + the real `opus-mt`
+engine, which loads in this env via the `[nli]`-shared
+`transformers`/`torch`):
+
+- query translated es→en, LLM prompted in English, **`answer_text`
+  is the English answer** ("Anarchocapitalism is a political
+  philosophy and economic theory that advocates the elimination of
+  the state…"); the verifier ran on that English core;
+- `display_answer` is the opus-mt es rendering of the verified English
+  prose ("El anarcocapitalismo es una filosofía política y una teoría
+  económica que aboga por la eliminación del Estado…"),
+  `display_translated=True`, `display_lang=es`, `engine=opus-mt-v1`,
+  banner present;
+- the bright line held: the Spanish text is additive/display-only;
+  `question_hash` + `verifier_policy_hash` invariant (test
+  `test_sandwich_hash_invariants`); `RetrievalPlan` MT binding is
+  zero-churn on non-MT runs (`test_retrieval_plan_mt_fields_*`).
+- 6 sandwich tests + full suite **2477 passed**, 0 regressions;
+  `make`/`tasks.py`/CLI surface (`--crosslang-translate` /
+  `XLANG_MT=1`) wired; default OFF.
+
+**Two honest limitations the live run exposed (NOT solved here):**
+
+1. **Recall is not yet good.** `opus-mt-es-en` rendered
+   `anarcocapitalismo` → "Anarchocapitalism" (one token, no hyphen),
+   which still did **not** rank the canonical *Anarcho-capitalism*
+   article primary — the live run was `UNGROUNDED` with off-target
+   sources (*Consensus decision-making*, *Karl Hess*, …). The
+   architecture is correct; the MT-output↔title normalization /
+   retrieval interaction needs tuning. This is exactly the experiment
+   the **default-OFF flag + bench gate** exist for, and why the
+   manifest revisions are placeholder (`main`, not SHA-pinned). "The
+   sandwich works" means the *mechanism*; it does **not** mean
+   cross-language grounding is solved.
+2. **JSON/lattice display is rough.** In `claim_lattice` (JSON) mode
+   `raw_answer` is JSON, so display-MT translates the envelope
+   (`"declaraciones"`). Prose/quote mode renders clean. JSON-mode
+   display polish (translate field *values* only, or render from the
+   parsed claim list) is future work — banner-labelled and default-OFF,
+   so acceptable as an experimental edge.
+
+**Doc honesty correction (2026-05-17).** Earlier drafts of §2 #6 /
+#000001 §7 said "governance untouched" — false: `governance_policy_hash`
+is sha256 of the *whole* policy (`keys.py:182`), so the crosslang
+flags move it like every policy flag (correct cache partitioning).
+Corrected in §2 #6 (authoritative), #000001 §7, and both index rows;
+artifact-over-instruction.
+
+## 9. Fan-out benchmark (2026-05-17) — measured, with the dead ends named
+
+**Repro (deterministic, no egress, no Hermes-for-translation).**
+`bench/make_lang_questions.py <lang>` regenerates the per-language
+question set from `bench/qa_questions.txt` via `opus-mt-en-<lang>`
+(the sandwich's own engine); `bench/qa_sweep.py … --policy
+crosslang_translate_enabled=true [--policy crosslang_entity_mask=…]
+[--policy crosslang_source_lang=…]` runs the live sweep;
+`bench/es_delta.py <en_baseline.jsonl> <xx_sandwich.jsonl>` is the
+authoritative metric (per-question EN→xx transition).
+`es_roundtrip_analysis.py` / `es_join_patterns.py` produced the
+*refuted* metric below — kept only as the cautionary record.
+
+**The real measurement (n=1, claim_lattice, same 75 questions).**
+EN baseline **85 %** grounded (64/75) → es+sandwich (no mask)
+**71 %** (53/75): the sandwich costs **−14 pp**. Transitions:
+PRESERVED 31 · DOWNGRADE S→H 16 · **LOST 17** · N/A 11 (6 of which
+*gained*).
+
+**Dead end (named, not buried).** A deterministic es→en round-trip
+content-token-recall metric (CLEAN 77 % / DRIFT 20 % / COLLAPSE 3 %)
+looked like a sharp predictor of grounding loss. It is **not**:
+against the live delta, CLEAN grounds 71 % and DRIFT 73 % (no
+separation), and **12 of the 17 LOST questions round-tripped CLEAN**.
+The metric is abandoned. This is another instance of the
+already-codified CLAUDE.md bench-maxing standing lesson (a clean
+synthetic metric ≠ real bench-qa behaviour — #000049 §7); **not**
+re-added to CLAUDE.md (DRY — it is already there).
+
+**The LOST 17, from the artifact (not a hypothesis):** ≈7 *entity
+translated/garbled* — `New London`→"nuevo Londres", `Tarsus`→"Tarso",
+**`Boltzmann`→"perntzmann"** (opus-mt hallucinated the name),
+`World War 2`→"Segunda Guerra Mundial"→"Second World War" (≠ corpus
+"World War II"); ≈3 *broad-enumeration* shapes ("name all members of
+the beatles", "list all planets") — a quantifier-guard×translation
+interaction, separate lever; ≈several *n=1 stochastic* (`founder of
+microsoft?`, `what continent is egypt on?` round-tripped clean,
+entities intact, yet LOST — needs n=3 to separate).
+
+**Lever built + validated (the entity slice).**
+`arborist/qa/mt/entity_mask.py` — deterministic mask/restore (quoted
+spans + Capitalised runs; corpus-title anchoring deferred to v2),
+`MaskedTranslator` wrapper, default-ON within the (default-OFF)
+sandwich, `crosslang_entity_mask` toggles it for the A/B. Measured
+through **real opus-mt** round-trip: `who is Paul of Tarsus?` base →
+"Pablo de Tarso" (the exact LOST case) → **+mask → "Paul of
+Tarsus"**; `Boltzmann` is now garble-proof (never reaches MT). 5
+tests.
+
+**Honest caveat — the bench under-states the lever.** The bench
+questions are *lowercased* (`a bridge between new london…`), so the
+Capitalisation detector cannot fire on the en side; on the xx→en
+edge it still catches `Egipto`/`London` etc. So any lift on *this*
+bench is a **lower bound** — real cased user input does strictly
+better. The detector that catches lowercase entities is
+**corpus-title anchoring** (protect any query span matching a known
+article title), deferred to v2.
+
+**Also landed:** the #000056 fan-out caught a real defect —
+`get_translator()` built a fresh ~300 MB model per call → at
+concurrency 4 the first sweep was **88 % engine-ERROR**
+(`meta tensors` race). Fixed: process-level memoised singleton +
+lazy *per-pair* load (6 manifest pairs never load eagerly) + load
+lock; 6 concurrent threads → 0 errors. French (`opus-mt-fr-en/-en-fr`)
+and Russian (`ru`) added as second/third bread (manifest +
+`crosslang_source_lang`); fr question set generated (`GNU Linux`
+survives). [Tasks #14–#17]
+
+**Lift result (2026-05-17, measured — and it refutes the lever).**
+n=1, claim_lattice, 75 q, vs EN baseline 64/75 (85 %):
+
+| config        | grounded   | Δ vs EN | LOST |
+|---------------|------------|---------|------|
+| es, no mask   | 53 (71 %)  | −14 pp  | 17   |
+| **es, +mask** | **49 (65 %)** | **−20 pp** | **22** |
+| **fr, +mask** | **35 (47 %)** | **−38 pp** | **33** |
+
+**Entity-mask did not help — it is neutral-to-negative.** The
+isolated `Paul of Tarsus` win did **not** replicate at bench scale:
+es went 53→49 (−4 q ≈ the 5 pp n=1 noise floor — *ambiguous*, not a
+confirmed regression but certainly no lift). fr is unambiguous and
+bad (−38 pp, well past noise). My earlier "lever validated" claim
+was premature; the bench refutes it. **Third instance this thread of
+the CLAUDE.md bench-maxing lesson: an isolated/synthetic win ≠
+bench-qa behaviour.** Likely causes (untested): the lowercased bench
+gives the mask almost nothing to grab on the en side, so on the
+xx→en edge it fires on *Spanish/French* capitalisation (which is not
+English entity structure) and the sentinels perturb opus-mt's
+translation of the rest of the sentence — net harm, not help.
+
+**Honest standing verdict (comparator corrected 2026-05-17, fox).**
+The right baseline is **not** native English — it is what
+cross-language did *before this ticket*: the 2026-05-16
+`¿Qué es el anarcocapitalismo?` run retrieved Spanish-song-title
+noise, burned 10.4 s, returned UNGROUNDED — **≈0 % grounded**.
+Against that true baseline:
+
+- **The sandwich is a large net win: ~0 % → 71 % grounded (es)**, with
+  the proof core never corrupted. The −14 pp vs *English* is the
+  *cost of a capability that previously did not exist at all*, not a
+  regression. Calling that a "fail" was a comparator error — doom-
+  framing is as dishonest as hype.
+- **The entity-mask *lever* is the genuine negative:** net-negative
+  at scale (es 71→65, fr −38 pp), so it is now **default-OFF**
+  (`crosslang_entity_mask=False`); the no-mask sandwich is the
+  better config and the one worth keeping.
+- French is materially weaker than Spanish (fr+mask 47 %; fr no-mask
+  unmeasured — likely higher since mask hurt es). Still ≫ the ≈0 %
+  baseline.
+
+So: **keep the sandwich (clear win over no support); drop the mask
+default (failed lever); the −14 pp-vs-English is a discount on a new
+capability, not a defect.** Remaining: n=3 to firm es/fr numbers,
+fr no-mask measurement, corpus-title anchoring (the one untried
+detector that works on lowercase). Default-OFF as an *experiment
+flag*, but the experiment is now **promising, not failing**. fox
+decides next spend; the honest recommendation flipped — this is
+worth continuing, just not with the mask.