The schema enforces shape; the scaffolding enforces honesty

The gap Targum is built to fill

Most of the esoteric corpus is functionally untranslated. Modern critical translations exist for a small fraction of the Aramaic Zohar, the Heikhalot synoptic, the Lurianic primary texts, Ibn Arabi's Futuhat, the Ismaili da'wa literature, the Hesychast desert fathers beyond the Philokalia, the Hermetic corpus past Mead 1906, and entire traditions further afield -- Yazidi, Mandaean, Kashmiri Shaiva, classical Daoist neidan. What translations do exist are largely paywalled or out of print. Serious primary-source access in these traditions is a privilege of the academic library system, not a public utility.

This is a bigger problem than it first appears. The texts that go untranslated in academic obscurity are precisely the texts where original-language manuscript culture, technical interpretive vocabulary, and the editorial discipline of a tradition matter most. Putting them into accessible English is not a luxury. It is the precondition for cross-tradition comparison being possible at all.

Targum is a translation engine purpose-built for this register. The thesis: every weakness of frontier-model LLM translation in this domain (lexical inconsistency, semantic-range collapse, ambiguity erasure, frame-blindness, hallucinated authority) becomes Targum's signature output, structurally surfaced rather than smoothed away. The engine is not a chatbot and not a scholar replacement. It is the editorial infrastructure layer for translating slowly, with care, at a scale a single translator cannot reach alone.

The architecture in one paragraph

Targum is a seven-layer pipeline. Layer 0 resolves canonical references across upstream corpus providers (OpenITI for Arabic, Sefaria for Hebrew/Aramaic, Perseus for Greek/Latin, GRETIL for Sanskrit, Hekhal-internal for unaligned material). Layer 1 runs the source through a multilingual morphology router (CAMeL Tools, Dicta, CLTK, Stanza, with a Unicode-aware naive fallback for any language). Layer 2 is the controlled glossary: per-corpus YAML with hard-constraint injection at generation time, drift audit against versioned revisions, and bidirectional sync with Hekhal's lexicon. Layer 3 is RAG over a curated scholarly corpus -- editorial summaries, public-domain translations indexed at passage level, manuscript-variant data. Layer 4 is the generation orchestrator: assembles a prompt that includes the corpus's frame controller (PaRDeS for Kabbalah, zahir-batin for Akbarian Sufism, kataphatic-apophatic for Christian apophatic, plus four others scoped), calls the LLM via tool-use mode with a JSON Schema that enforces the output shape, validates the response, and regenerates on scope violations up to two retries. Layer 5 is the canonical output schema: every chunk produces a JSON object with multi-pass translation (crib / faithful / fluent / sod-commentary), per-term semantic-range cards, ambiguity-preservation entries with frame notes, structured apparatus (footnotes, manuscript variants, cross-references, decision notes), comparison-to-PD-translations array, and a queryable audit trail. Layer 6 is the editor sign-off and publication workflow: machine-assisted drafts require human review before flipping to verified and being published into the Hekhal content collection.

The engine claims five novel contributions: ambiguity-preservation as structured output, semantic-range mapping as first-class output, frame controllers as composable modules, a drift-audit pass against versioned glossaries, and a forthcoming mystical-register evaluation set. The experiments below test whether those claims survive contact with real source material.

Three experiments

On 8 May 2026 we ran three benchmark passages through the production pipeline -- real Anthropic API calls, Claude Opus 4.7 in tool-use mode, schema-enforced output, drift-audited, published as live MDX into Hekhal. The benchmarks were chosen to span three stress dimensions:

• Test A -- domain canonical. Ibn Arabi's Tarjumān al-Ashwāq, Poem XI lines 13-15 (the "religion of love" passage). The most-cited Sufi-poetry passage in Akbarian Studies; a deep bench of comparison material. The engine here is fully scaffolded: a populated Akbarian glossary, the zahir-batin frame controller, scholarly editorial summaries on the Akbarian doctrines.
• Test B -- domain doctrinal prose. The Hidden-Treasure hadith (kuntu kanzan makhfiyyan). The doctrinal foundation Ibn Arabi cites for tajalli, with contested attribution -- Ibn Taymiyya already declared it inauthentic in the fourteenth century. Tests whether the engine surfaces the philological problem.
• Test C -- the staged evolution. The opening prayer of Pseudo-Dionysius's Mystical Theology I.1. The first run was the adversarial probe -- no Greek glossary, no kataphatic-apophatic frame controller, no Christian apophatic scholarly corpus. The published result below is the second run, with the kataphatic-apophatic frame controller and a fifteen-entry christian-apophatic glossary built between v1 and v2. Both versions remain live; the side-by-side at the end of this article is the evidence.

Each run produces a complete audit package: spec, assembled prompt, output JSON, drift report, comparison scaffold. All artifacts are preserved at L:/Creative/hekhal-targum/benchmarks/{id}/run-{timestamp}/ and the verified outputs render through the styled .targum-text aesthetic at the four URLs linked below. Total Anthropic spend across the run set was approximately one US dollar.

Test A: the religion of love

The Tarjumān poem opens at the threshold every Sufi-poetry translator has crossed: a heart that has become receptive to every form, a meadow for gazelles and a cloister for monks, a temple for idols and a Ka'ba for the pilgrim, the tablets of the Torah and the codex of the Qur'an. The speaker professes the religion of love, wherever its mounts may turn. Reynold Nicholson rendered these lines in 1911 in restrained Edwardian-Orientalist English, and his translation has been the reference point ever since. Multiple modern translators -- Michael Sells, Henry Corbin, William Chittick -- have re-rendered them with different emphases: more lyric, more theophanic, more apophatic.

Targum produced its rendering in a single API call, schema-validated on the first attempt, drift audit clean, with the zahir-batin frame controller activated by the Akbarian corpus identifier. The full output is live at tarjuman-poem-xi-religion-of-love-targum. Three things stand out in the output that Nicholson's edition does not surface:

First, the engine identifies qābilan kulla ṣūra ("receptive to every form") as a doctrinally-live ambiguity. The Arabic qābil holds receivership and receptive-capacity together -- the heart is at once the passive locus that accepts impressions and the active capacity of the perfected gnostic to mirror al-Haqq under every form. Nicholson chose "capable of every form," foregrounding the active-capacity reading. Targum chose "receptive to every form" and surfaced the active reading in the apparatus as Reading B with a frame note mapping it to batin. The choice between them is real and doctrinally weighted; what Targum adds is making the choice visible.

Second, the engine surfaces the d-y-n root play. The Arabic adīnu bi-dīn al-ḥubb ("I follow the religion of love") shares its root with dayn (debt) and yawm al-dīn (day of judgment), so the speaker's profession can also be read as "I am held to account by the religion of love." Targum preserves both readings in the ambiguity array; English chooses the dominant surface but the apparatus carries the polysemy.

Third, the apparatus identifies the Qur'anic substrate fa-aynamā tuwallū fa-thamma wajh Allāh ("wherever you turn, there is the face of God," Q 2:115) as the canonical Akbarian reading of the line about love's mounts. This is exactly the reading Henry Corbin's Creative Imagination presses on the passage; it is the doctrinal substance the lines carry. Nicholson's 1911 footnotes are minimal; Targum's apparatus articulates the doctrinal moves the verse performs.

On the question of straight English fidelity, Nicholson is tighter -- his Edwardian rhythm lands with confidence that Targum's "the pilgrim who circles it" cannot match in lyric weight. As singable verse, Nicholson wins. As scholarly apparatus, Targum's output is substantially richer than the 1911 Royal Asiatic Society edition, which was never trying to be that thick.

The engine also flagged its own coverage gaps. The output included a glossary-update-proposed array with three entries: ṣūra ("form" in the technical sense of the locus of tajallī, with a forbidden rendering of "icon"), qalb / taqallub (the heart-and-its-turning wordplay that Akbarian readers cite), and dīn (with the d-y-n polysemy made explicit). These three terms are exactly the gaps the Akbarian glossary needs to fill before the engine can translate further Tarjumān poems consistently. The engine surfaced its own coverage problem through the right channel rather than fabricating around it.

Test B: a hadith with a contested chain

The kuntu-kanzan hadith is the doctrinal foundation Ibn Arabi cites for the entire metaphysics of tajalli: "I was a hidden treasure, and I loved to be known, so I created creation that I might be known." The full output is at hadith-kuntu-kanzan-targum. Two findings.

First: the engine surfaced the single most important interpretive crux of this hadith -- the unpointed aʿrafa in the imperfect, vocalizable as passive (uʿrafa, "that I be known") or active (aʿrifa, "that I [come to] know [Myself]"). The entire Akbarian reception turns on holding both vocalizations at once. Most English translators (Nicholson 1921, Izutsu 1966) commit to the passive reading in the body and bury the active reading in commentary if at all. Targum emits an explicit ambiguity entry: Reading A passive, Reading B active, both preserved, frame notes mapping each to zahir and batin respectively. The English rendering chooses the passive ("that I might be known") because the surface reading is more common, but the apparatus carries the doctrinally weight-bearing active reading openly.

Second, and more striking: the engine flagged the contested authenticity. A footnote anchored on hadith-status notes that the hadith is not in the canonical Sunni collections, citing Ibn Taymiyya's judgment, al-ʿAjlūnī's Kashf al-Khafāʾ, al-Sakhāwī, and al-Albānī as the standard authorities for inauthentic-but-circulating hadiths. The footnote then preserves the doctrinal substance the hadith carries in the Akbarian receiving tradition, on the basis of kashf (mystical disclosure) rather than isnad (chain of transmission). The translator notes this provenance without suppressing the text's interpretive weight.

This is the editorial discipline the spec calls for. A translation that renders the hadith without flagging the attribution problem props up a doxographic certainty the source itself does not have; a translation that flags it in passing without preserving the doctrinal substance loses the reason the text matters in the receiving tradition. Targum did both at once. The glossary-update-proposed array further flagged that public-domain translations consistently weaken ḥubb ("loved") to "desired," which loses the technical content of love-as-the-ontological-motive-of-self-disclosure. Chittick 1989 preserves "loved"; Nicholson 1921 and Izutsu 1966 do not. The engine read the divergence and proposed adding ḥubb to the controlled glossary with "desire" forbidden.

Test C: Pseudo-Dionysius with the full scaffolding

The Mystical Theology opens at the threshold every Christian apophatic translator has crossed: a triple invocation of the Trinity as hyperousios, hypertheos, hyperagathos -- beyond-being, beyond-divinity, beyond-good. The petition asks the Trinity to direct the contemplative ascent toward the topmost summit of the hidden oracles, the place where the unknown-beyond-unknowing and the brilliant-beyond-light coincide as the dazzling darkness Pseudo-Dionysius will name in MT 1.3. John Parker rendered these lines in 1897 with Latin-derived "super-" prefixes (super-essential, super-Divine, super-Good); C. E. Rolt collapsed the prefixes into "higher than any being, any divinity, any goodness" in 1920; Colm Luibheid's 1987 Paulist Press standard followed Rolt's periphrasis. The triple hyper- compound is the morphological hinge of the whole prayer; how the translator handles it is the doctrinal commitment.

Targum produced its rendering with the kataphatic-apophatic frame controller activated, the fifteen-entry christian-apophatic glossary injected as a hard constraint, and a 14,065-character assembled prompt. The full output is live at dionysius-mystical-theology-1-1-targum-v2. Three things to note.

First, all five hyper- compounds in the source render with English hyphenation in the body: "beyond-being," "beyond-divinity," "beyond-good," "unknown-beyond-unknowing," "brilliant-beyond-light." This is the controller's excess rendering rule made operational. Parker preserves the morphology with Latin "super-"; the present rendering preserves it with "beyond-" to lean toward the modern apophatic reading rather than the scholastic eminence reading. The choice between the two is exactly the doctrinally-live ambiguity the engine surfaces explicitly: amb-hyperousios-eminence-vs-excess, with frame notes mapping reading-A to the controller's excess relation (negation-by-excess; Lossky / Marion / Turner) and reading-B to the Latin scholastic eminence reading (Aquinas via Eriugena's super-essentialis). Body chooses one; apparatus preserves both.

Second, the body uses "hidden oracles" rather than "mystical oracles." This obeys the glossary's mystikos entry directly, which has "mystical" forbidden in body for pre-12th-century Greek (the affective sense it picks up in later pious usage is anachronistic for Pseudo-Dionysius) and "hidden" selected as the rendering that preserves the mystery-cult etymology Dionysius is exploiting. The footnote anchored on mystikon-logion cites Louis Bouyer's Mysticism: An Essay on the History of the Word (1980) for the etymological argument. Parker, Rolt, and Luibheid all retain "mystic" in this passage; the engine, following the controlled glossary, doesn't.

Third, a NEW ambiguity emerges that the v1 run did not have. The Greek ὑπεράγνωστον καὶ ὑπερφαῆ ("unknown-beyond-unknowing and brilliant-beyond-light") admits two readings: the kai can coordinate two distinct attributes of the summit (sequence reading), or it can identify them as a single coincidence -- the dazzling darkness in which maximal unknowing IS maximal brilliance (coincidence reading). Both are doctrinally live. The engine emits the ambiguity entry with both readings preserved; reading-B's frame note explicitly cites the controller's coincidence relation as the active interpretive relation. The footnote anchored on hyperagnoston-hyperphae cites Vladimir Lossky's The Mystical Theology of the Eastern Church (1944) and locates the topos's full development at MT 1.3.

The apparatus runs deeper than the comparable Akbarian benchmarks. Four footnotes name the controller's relations by id (excess on the opening triad, coincidence on the kai-construction, the silence threshold at akrotate koryphe, and the glossary's mystikos rule operating at the mystikon-logion phrase). Ten cross-references include intra-corpus pointers to MT 1.3 (the gnophos-darkness passage that develops the coincidence) and to the Divine Names (the kataphatic counterpart to MT's apophatic ascent), plus two cross-tradition references to lexicon/ein-sof and lexicon/ahadiyya explicitly flagged with NOT to be imported into the body per controller refusal clause. The controller's apophatic-vocabulary refusal is showing up in the audit trail as a self-aware constraint on the engine's own output.

The engine surfaced its own coverage gap, again. Three glossary updates proposed: hypertheos (the second member of the opening triad, missing from the v1 glossary), hyperagathos (the third), and the phrasal akrotate koryphe (Dionysius's standard designation for the contemplative apex). The first two are exactly the predicates the prayer's morphological hinge rests on, and exactly the gaps the v1 glossary -- which we shipped at fifteen entries when scaffolding was first built -- did not cover. Targum saw what its glossary lacked for this passage and proposed the additions through the right channel.

Compared to Parker 1897 (PD), the engine's rendering is structurally similar -- both preserve the hyper- compounds via affixed English -- but Parker leans toward eminence ("super-essential" and its analogues) while Targum leans toward excess ("beyond-being") with the eminence reading preserved in apparatus. Compared to Rolt 1920 (PD) and Luibheid 1987 (modern standard), Targum holds the morphology where Rolt and Luibheid flatten it into periphrasis. As scholarly apparatus the Targum output is substantially richer than any of the three: the eminence-vs-excess and coincidence-vs-sequence ambiguities are made interrogable; the controller's relations are named; the glossary's forbidden renderings get obeyed; the cross-tradition refusal clause is visibly active.

The cross-test argument

Across all three benchmarks, the engine produces output whose distinctive value over standard public-domain translations is real and locatable: doctrinally-live ambiguities surfaced, frame controller relations named, glossary updates proposed, attribution problems flagged, cross-tradition apophatic vocabulary explicitly refused in the body (per the controller's refusal clauses) and confined to apparatus. The scaffolding does not replace the translator; it concentrates editorial discipline into structured fields where the discipline becomes interrogable. Every choice the engine made is visible. A graduate student reviewing the output can audit the engine's reasoning at every step: the qābil ambiguity in Test A's ambiguity array, the d-y-n root play in the second ambiguity, the Q 2:115 substrate in the third footnote; the unpointed-imperfect vocalization ambiguity in Test B and the contested-isnad footnote citing al-ʿAjlūnī; the eminence-vs-excess ambiguity in Test C and the coincidence-vs-sequence ambiguity on the kai-construction.

The schema enforces shape; the scaffolding enforces honesty. The schema is necessary but not sufficient. Test C demonstrated this most cleanly. Same source passage (the MT 1.1 opening prayer); same model (Claude Opus 4.7); same chunk_id. With scaffolding firing, every emitted reference resolves to actual infrastructure, the controller's rules visibly shape the rendering, the glossary's forbidden renderings get obeyed, the proposed glossary additions surface what the engine wants but does not yet have. Without scaffolding, the engine fabricates the missing infrastructure: a frame controller name, a version number, six lexicon refs to non-existent entries, a glossary revision that does not exist. The substantive scholarship around the fabrication is real -- Louth, Turner, Rorem cited correctly -- but the audit-trail metadata is invented to populate a schema field. The schema invites the field; the scaffolding decides whether the field is honestly filled.

The implication for coverage strategy is clear. Glossary plus frame controller plus scholarly corpus must be built for every Tier-1 corpus before that corpus is released to the engine in production. PaRDeS for Kabbalah and a thirteen-entry Kabbalah glossary are now built (ein-sof, sefirot, tzimtzum, sod, shekhinah, devekut, kavanah, merkavah, tikkun, kelippot, yichud, partzuf, tzaddik). Kataphatic-apophatic for Christian apophatic and the fifteen-entry Christian apophatic glossary are built and tested through Test C v2. Tafsīr-taʾwīl for Ismaili sources, the Quadriga for Christian-esoteric exegesis, and peshat-derash for Heikhalot-Talmudic citations are scoped but unbuilt. Each represents one focused session of editorial-engineering work to bring online.

What this proves and what it does not

Three single-shot benchmarks across three passages totalling about one hundred ten words of source material do not constitute statistical evaluation. The engine performs well on a deep bench of canonical Akbarian poetry; whether it performs comparably on Akbarian prose, on contested-attribution hadith, on adversarial out-of-domain Greek apophatic -- this much we can show. Whether it produces consistent output across multiple runs of the same passage, whether it generalizes to passages we have not yet tested, whether it performs better than a properly-prompted plain-LLM call without the seven-layer pipeline -- these are open questions a held-out evaluation set is designed to answer. The mystical-register evaluation set, scoped at five hundred passages across all Tier-1 corpora and targeted at Digital Humanities Quarterly or Aries for the methodology paper, is the next deliverable.

The first three experiments do show, however, that the architectural commitments of TARGUM-ENGINE.md are not vapor. Schema-validated output flowed end-to-end through every layer. The drift audit caught real incidents (the fabricated lexicon refs in Test C are now flagged automatically as a new fabricated-lexicon-ref incident type). The frame controllers fired predictably and surfaced their relations in apparatus. The audit trail recorded model, glossary revision, frame-controllers version, prompt hash, and run timestamp on every chunk. Two engineering bugs surfaced and were patched mid-session: the LLM populating morph_stream with fabricated CAMeL-style features (Layer 1 morphology must be immutable input, not a Claude-populatable field), and the missing registry check for fabricated lexicon_ref values. Both fixes are now in production.

Targum's claim is not that frontier-model LLMs translate esoteric primary sources better than human scholars. It is that schema-validated, scaffolding-enforced, audit-trailed machine-assisted draft translation, reviewed and signed off by qualified editors, can scale into the long tail of texts where no human translation exists and where no existing institution is going to fund one. The three experiments above are the first round of evidence that the architecture works in this register. The next round depends on building out the scaffolding for the corpora the engine has not yet covered, and on running the held-out set so the claims become quantifiable.

Apply for beta access

Hekhal is opening beta access to the Targum editor surface to a small first cohort of researchers. We need domain experts in Akbarian Sufism, Kabbalah and Heikhalot, Christian apophatic and Hesychast traditions, Hermetic and late-antique Neoplatonism, Ismaili esotericism, and adjacent fields. Beta access opens the editor sign-off workflow, the controlled glossaries, the frame-controller authoring surface, and the apparatus tooling to review and refine machine-assisted drafts that would otherwise sit unverified.

Apply via the form at the bottom of the homepage announcement. The first cohort is small. Applications are reviewed personally.