Subtracting the Typological Baseline: Discourse Organisation and the Assessment of Fidelity in a Tibetan–French Translation
Nov 17, 2026 · @Drupchen
Abstract
Assessments of translation from Tibetan into Romance languages routinely record as semantic loss what is in fact obligatory typological re-encoding. This paper demonstrates the error, and the procedure that corrects it, on a single short passage: an oral meditation instruction in spoken Tibetan and its published French rendering. The two texts were first analysed independently and in isolation from one another — segmentation, connective inventory, predicate hierarchy, macro-architecture, agency distribution, modality, lexical texture — with neither description permitted to borrow a category from the other. A first comparison was then conducted bottom-up, ascending from predicate to whole text. That comparison returned a verdict of architectural failure: two of five identified source devices reversed, the central argument of the passage not carried. A second pass, after partitioning the Tibetan features into typological defaults, genre conventions and genuinely marked choices, withdrew almost all of it. Head-final clause chaining, verb-final position, repetition as the unmarked device for iteration and intensification, and lowabstraction figuration are properties of the language, not assertions by the speaker; the French renderings that the first pass scored as losses — a scalar adverbial for iconic repetition, a cleft for positional focus, anaphora for flat repetition, a redistributed ideophone compound — are the correct Romance encodings of the same functions. A strict meaning-unit audit over twenty-six units leaves one loss that touches propositional content: the substitution of an addressee-directed simple future for Tibetan genericpredictive morphology, a substitution for which French possessed an unused idiomatic alternative. The paper proposes a four-step assessment procedure in which a source choice counts as lost only if it was marked in the source and the target language demonstrably held an equivalent resource, and notes the consequence for corpus-based extraction of translation norms: an exemplar corpus weighted toward philologically literal work will encode the very bias the procedure is designed to remove. A second study then tests these findings against a larger corpus of the same genre — roughly twenty minutes of transcribed oral teaching with its French rendering. The structural findings are corroborated and two are strengthened: the definition pattern recurs at least ten times, with a rhetorical question-answer variant more frequent still, and the finite-to-non-finite ratio proves to vary so widely within one speaker that it cannot encode prominence at all. The encoding findings require recalibration, and the single propositional loss survives with its explanation replaced — the addressee-directed future is not a translator’s choice but the terminus of a four-step cascade beginning in French’s resistance to stacked nominalisation. The paper’s own account of self-concealing assessment was corroborated unintentionally, by the author committing the error again, on the new material, with the corrective procedure already written.
1. Introduction
The recurrent failure in assessments of Tibetan-to-Romance translation is not a failure of philology. It is a failure to establish a baseline before measuring.
A translation assessment proceeds, in practice, by laying source against target and recording divergences. Where the source has a finite verb and the target has none, a predicate is recorded as lost. Where the source repeats a clause and the target states a scalar, iteration is recorded as flattened. Where the source predicates a concrete trait of an abstract process and the target states the process analytically, a figure is recorded as effaced. Each of these records is accurate as a description of surface divergence. None of them is evidence of semantic loss, because in each case the source form is what the source language produces by default when nothing in particular is being asserted.
The consequence is systematic and directional. Classical Tibetan and its spoken registers are rigidly head-final: subordinate clauses precede main clauses, modifiers precede heads, and the finite verb closes its period. A period therefore contains one finite verb and a chain of converbs and nominalisations — not because the speaker has ranked the chained material as subordinate in importance, but because that is the only shape a Tibetan period has. French has no such rank. It encodes prominence through mood, sentence boundary, connective selection and cleft. A translator moving between them must reassign prominence, and an assessor who treats the source’s distribution as the norm will find the reassignment everywhere and will call it damage.
The error has a second property that makes it durable: it is self-confirming. A literal translation shares surface structure with its source, so laying the two side by side produces an impression of confirmation. A discourse-level translation diverges in surface structure, so the same operation produces an impression of detected error. The method that cannot see the problem is the method that feels more rigorous, and every finding it yields feels well evidenced.
This paper demonstrates both the error and its correction on a single passage, chosen because it is short enough to analyse exhaustively and because it concentrates in nine lines most of the features at issue: a converb chain with a repeated-verb envelope, two consecutive clause-chaining conditionals, an ideophonic compound recurring across two outcomes, an iconically repeated formulaic couplet, generic-predictive verbal morphology, and a terminal identificational clause carrying the technical term the whole passage exists to define.
The argument proceeds in three movements. Sections 4 and 5 present independent analyses of the two texts, each conducted in isolation from the other, so that neither description is contaminated by the assumption of equivalence. Section 6 presents a first comparison conducted in the ordinary way, against source form, and the verdict it produced; this is retained in full and presented as data rather than as result, because the paper’s claim is partly about what that method reliably yields. Sections 7 to 10 re-sort the Tibetan features by markedness, re-run the comparison by function delivered, and audit the result against a strict textual-linguistic criterion. Section 11 draws out the methodological and practical consequences, including for the machine-assisted assessment of translation, where the bias is reproduced from a training corpus rather than from a theory and is correspondingly harder to detect.
2. Theoretical framework
Five distinctions carry the analysis. None is novel; the contribution lies in insisting that all five be applied before any divergence is recorded as loss.
2.1 Clause chaining and the finiteness constant
Tibetan builds extended discourse by chaining: a sequence of non-finite clauses, each marked by a converb or a nominaliser, closed by one finite verb. The converb inventory is functionally differentiated — ནས་ for sequence and means, སྟེ་/ཏེ་ for non-sequential elaboration, ན་ for consecution, པར་/བར་ for manner including negated manner — and the selection among them is informative. The ratio of finite to non-finite clauses is not. It is fixed by the grammar.
This matters because the ratio is the first thing a comparison notices. A Tibetan period with one finite verb and five converbs, set beside a French sentence with five finite verbs, invites the reading that the source subordinated four predications which the target has promoted. The reading is an artefact. What is informative is which element the source chose to place under a particular subordinator when coordination was available.
2.2 Prominence: morphological rank against positional and modal encoding
In a head-final clause-chaining language, final position and finite marking coincide and are obligatory. Neither therefore signals prominence on its own. Tibetan does have prominence devices — topic particles, the repeated-verb envelope in which a nominalised verb opens a period and its finite form closes it, selection of an unexpected subordinator — but these are deviations from the chaining default. French has no morphological rank to assign. Its prominence resources are mood (imperative against indicative against future), sentence boundary placement, connective selection, and the cleft. A French translator reassigning prominence is therefore not choosing between preserving and discarding a source ranking; the ranking is not transferable. The translator is choosing which French device will carry the function the source carried positionally.
2.3 Repetition: marked in one language, unmarked in the other
Tibetan uses repetition as its ordinary device for durativity (verbal reduplication), iteration (clausal repetition) and intensification (ideophonic reduplication). It carries no stylistic charge. French repetition is marked, and what it marks is either emphasis or poverty of style. The asymmetry has a strong consequence: preserving source repetition in French adds meaning that the source did not carry. A translator who renders an iconically repeated Tibetan clause as a repeated French clause has introduced a stylistic signal — and, in a pedagogical text, a signal of incompetence — that is absent from the original. The faithful French rendering of iteration is a scalar or iterative adverbial. The corpus of §10.4 shows two such devices rather than one, selected by whether the reduplication is simple or correlative: ཉུང་དུ་ཉུང་དུ་ and མང་དུ་མང་དུ་ take de plus en plus , while the paired ཞེ་སྡང་མང་རུ་མང་རུ་ … དགྲ་ཆེ་རུ་ཆེ་རུ་ takes the correlative plus … plus … . Lexicalised iteratives are a third case and keep their French equivalent: ཡང་ནས་ཡང་དུ་ remains encore et encore . Whether the markedness of repetition that licenses all three holds at the same strength across the other Romance languages is untested here and is not to be assumed (§11.6).
2.4 The abstraction differential
Tibetan, and spoken Tibetan in particular, predicates concrete traits of abstract processes as a matter of course: a thought grows old, a mind goes smooth, thoughts go away in a direction. These are not figures chosen for effect; they are the available vocabulary. French explanatory prose of comparable authority operates at a higher level of abstraction, and a French text that predicates senescence of a cognition reads as either poetic or juvenile — neither of which the source is.
It follows that a faithful French rendering must often raise the abstraction level to hold the register constant. The concrete figure is extracted for its propositional content and restated analytically. Assessed against source form this registers as effacement of imagery. Assessed against function it is the operation that keeps the text sounding like what it is.
The corpus narrows the claim. The differential applies to figurative predication — a concrete verb taking an abstract subject, as when a thought grows old (རྒས་) or a mind bolts like a horse (སྐྱག་ལང་), both of which the corpus translator de-metaphorises consistently. It does not apply to explicit comparison. The corpus’s extended similes — the monkey at five windows, bubbles rising in water, the elephant and its driver, the king and his subjects — transfer intact into French and are sometimes expanded for clarity, the monkey acquiring a ne restant à chacune d’elles qu’un bref instant that renders ལྟ་ཙམ་ལྟ་ཙམ་ with material the Tibetan does not contain. So the rule is not that French requires more abstraction in general, but that it requires it where Tibetan predicates concretely of an abstraction.
2.5 Textual linguistics and the unit of fidelity
The criterion applied here is that of text linguistics rather than of equivalence at any lower rank. Units below the text — morphemes, words, phrases, clause types, syntactic configurations, sentence boundaries, even the count of periods — may be freely lost, substituted or redistributed. What may not be lost is a meaning unit : an idea present in the source text and absent from the target. Epistemic and deontic modality fall under this criterion, being propositional rather than stylistic; a statement of how things generally behave and a prediction addressed to a hearer are different claims about the world.
This criterion yields a decisive test, which Section 10 applies. A divergence counts as loss only when two conditions hold jointly: the source choice was marked rather than default, and the target language demonstrably possessed a resource for the same function which the translator did not use. Where the first condition fails, the assessor is measuring typology. Where the second fails, the assessor is recording an untranslatability, which is a fact about the language pair and not a defect in the translation.
3. Materials and method
3.1 The texts
The source is a nine-line passage of transcribed spoken Tibetan, an oral meditation instruction belonging to a shedra-style expository register. It instructs a practice, predicts its outcome in two stages, and closes by naming the resulting mental state with a technical term. The target is a French rendering of continuous prose, five sentences long — a working draft, still in progress, by the same translator whose finished work forms the corroboration corpus of §10.4. Both are given in full, segmented, in Appendices A and B.
The passage is not self-contained at its left edge: it opens with the anaphoric locative དེར་, which points to a referent established in prior discourse not available to the analyst. The consequence is recorded in Section 12 as the single undecidable question in the audit.
3.2 The sealed-analysis protocol
Each text was described independently and in isolation. The Tibetan was analysed without reference to the French; the French was then analysed without reference to the Tibetan and without the fact of its being a translation entering the description. Neither analysis was permitted to import a category from the other, and neither used the vocabulary of loss, addition or correspondence.
The protocol matters more than it may appear. The ordinary assessment procedure — reading source and target together, clause by clause — makes it impossible to see the target as a text, because every feature is immediately interpreted as a response to a source feature. The French nominal absolute examined in Section 5 is a case in point: read against the Tibetan it presents as a missing predicate, and read as French it presents as a circumstantial construction doing exactly the work a Tibetan subordinate clause does. Which of those two descriptions the analyst produces is determined by the order of operations, not by the evidence.
The protocol also localises the bias. In the present case, the independent descriptions required no retraction; every finding withdrawn in Section 8 was generated after comparison was licensed. The bias is therefore a property of the comparative operation , not of descriptive capacity, which has a direct practical consequence for how assessment work should be divided (Section 11.5).
3.3 The two-pass design
The comparison was then run twice, deliberately.
The first pass (Section 6) ascended from the smallest units upward — predicate, argument structure, clause rank, junction, period, macro-architecture, modality, whole text — reusing each level’s findings at the next. It took source form as the norm, which is the ordinary procedure. It is reported in full, including its verdict, because the paper’s claim is in part a claim about what that procedure yields on a well-made translation.
The second pass (Sections 7 to 9) inserts one step before comparing: the source features are partitioned by markedness, and only marked features are carried forward. The same ascent is then repeated, scoring by function delivered rather than by form matched, and the result is audited against the meaning-unit criterion of §2.5.
The difference between the two verdicts is the measurement of the bias.
3.4 Production of the analysis
The descriptive passes, both comparisons, the corroboration study and the text of this paper were produced by a large language model (Claude Opus 5, Anthropic) in extended dialogue with the author. The division of contributions is set out at the end of the paper. This bears on how the paper is written. Where the model’s analysis was corrected by the author, the original and the correction are both retained in the text rather than the first being silently revised away. §10.8 reports one such episode at length, because the episode is evidence for §11.2 — a paper arguing that source-form assessment is selfconfirming is in a position to demonstrate it, and did.
4. The Tibetan text analysed independently
4.1 Segmentation
The passage falls into four ཤད-delimited periods, with a cluster of short clausal ཤད inside the third. Finite verbs are marked in bold.
Period A. དེར་རྩེ་གཅིག་ཏུ་བལྟས་པ་ · དཔེར་ན་ [mantra] ཟེར་མཁན་གྱི་སྔགས་དེ་ · སྟོང་ཕྲག་གཅིག་ཙམ་བཟླ་བའི་ཚེ་ན་ · སངས་རྒྱས་ཀྱི་སྐུ་ལ་དམིགས་པ་དེ་འདྲ་བྱས་ནས་ · བལྟས། Period B. དེ་ནས་ · མཚམས་བཞག་ནས་ · ག་ལེར་རང་གི་སེམས་མ་བཅོས་པར་བཞག་བཞག་ ན་ · སེམས་ལ་རྣམ་རྟོག་གང་ཡང་ མི་འཕྲོ་བར་ · ཆགས་སྡང་གི་རྟོག་པ་གཏན་ནས་མེད་པ་ · འཇམ་ཐིང་ཐིང་ལྷོད་དེ་བ་བྱས་ · སེམས་གནས་སྡོད་མཁན་གཅིག་ ཡོང་གི་ རེད། Period C. དེ་གནས་བསྡད་པ་དེའི་ཐོག་ལ་ · གནས་སུ་བཅུག་སྟེ་ · སེམས་མ་ཡེངས་པ་ཙམ་དུ་སྐྱོང་བསྡད་ ན་ · དེ་ནས་དེའི་རྣམ་ རྟོག་ སྔ་མ་འོང། ཕྱི་མ་རྒས། ཕྱི་མ་རྒས། · སྔ་མ་འོངས་ཏེ་ཕར་མེད་པ་ཆགས་ནས་ཐིང་ཐིང་ འགྲོ་ཡི་རེད། Period D. དུས་དེའི་ནང་ན་ · སེམས་མི་རྟོག་པ་ཟེར་མཁན་ འདི་རེད།
Seven finite verbs across twenty-six predicate-bearing forms.
4.2 Connective inventory
| Marker | Type | Function in this passage |
|---|---|---|
| དེར་ | anaphoric locative | opens by pointing outside the passage |
| དཔེར་ན་…ཟེར་མཁན་གྱི་ | exemplification bracket | encloses the mantra citation |
| -བའི་ཚེ་ན་ | temporal subordinator | builds a frame, backgrounds its clause |
| -བྱས་ནས་ | manner / means converb | subordinates the pro-verbདེ་འདྲ་བྱས་ |
| དེ་ནས་(×2) | sequential stage marker | onset of B; onset of C’s apodosis |
| -ནས་ | sequential converb | མཚམས་བཞག་ནས་,ཆགས་ནས་ |
| -ན་(×2) | consecutive / conditional | the two load-bearing joints |
| -པར་/ -བར་ | negative manner converb | མ་བཅོས་པར་,མི་འཕྲོ་བར་ |
| -སྟེ་/ -ཏེ་ | elaborative converb | བཅུག་སྟེ་,འོངས་ཏེ་ |
| Marker | Type | Function in this passage |
| -ཐོག་ལ་ | basis anchor | “building on that” |
| -ཙམ་/ཙམ་དུ་ | approximative / limitative | on the count; on the sustaining |
| དུས་དེའི་ནང་ན་ | temporal anchor | bridges to the naming clause |
| ཟེར་མཁན་…འདི་རེད་ | identificational closure | attaches the technical label |
4.3 Predicate hierarchy, period by period
Period A carries one finite predicate, བལྟས་. Four elements are subordinated to it: the period opens with the nominalised བལྟས་པ་ and closes with the finite བལྟས།, forming a repeated-verb envelope; the recitation is placed under ཚེ་ན་ as a temporal frame; the visualisation is placed under བྱས་ནས་ as manner, by way of the pro-verb དེ་འདྲ་བྱས་; and the mantra is cited inside a དཔེར་ན་ … ཟེར་མཁན་གྱི་སྔགས་དེ་ bracket whose exemplified item is the mantra itself, not the count and not the practice.
The period keeps two faculties lexically distinct throughout: བལྟ་, directing the eye, and དམིགས་, taking as mental object. One is the matrix verb; the other is its manner adjunct.
Period B is a consecutive construction. Its protasis is མཚམས་བཞག་ནས་ … བཞག་བཞག་ན་; its apodosis is ཡོང་གི་རེད་. The subject of the apodosis, སེམས་གནས་སྡོད་མཁན་གཅིག་, is premodified by three attributive blocks between which there is no connective whatever:
-
སེམས་ལ་རྣམ་རྟོག་གང་ཡང་མི་འཕྲོ་བར་
-
ཆགས་སྡང་གི་རྟོག་པ་གཏན་ནས་མེད་པ་
-
འཇམ་ཐིང་ཐིང་ལྷོད་དེ་བ་བྱས་
The absence of connectives is itself the signal: these are not stages but three simultaneous characterisations of one emergent condition. The indefinite གཅིག་ keeps that condition non-specific — it is being introduced, not identified. བཞག་བཞག་ is reduplicated, an oral durative.
Period C opens a second consecutive construction parallel to B’s, with ཐོག་ལ་ explicitly building it on B’s output. Three features of its protasis are notable: བཅུག་ is permissive rather than coercive, letting the mind be placed rather than placing it; བསྡད་ auxiliarises སྐྱོང་ to give continuative aspect; and than non-distraction. ཙམ་དུ་ specifies the measure of the sustaining as no more
C’s apodosis contains the only stylised stretch in the passage: སྔ་མ་འོང། ཕྱི་མ་རྒས། ཕྱི་མ་རྒས། སྔ་མ་ འོངས་ཏེ་ …. The pattern is [A B] [B] [B] [A → consequence]: a nascent predicate (འོང་, come) set against a senescent one (རྒས་, grow old) across the ordinal pair སྔ་མ་ / ཕྱི་མ་; the second member isolated and repeated; then the first returning in past stem with ཏེ་ and carrying into ཕར་མེད་པ་ཆགས་ནས་, becoming absent in a direction away from the practitioner. The repetition is iconic: the iteration of the clause enacts the succession it describes, and each short clause takes its own ཤད, giving the stretch a beat.
Period D is the naming move, and the technical term སེམས་མི་རྟོག་པ་ is withheld until this final clause, so that A to C function retroactively as its definiens.
4.4 Macro-architecture
[anaphoric opening — referent outside the passage] A PROCEDURE one finite verb; recitation = frame, visualisation = manner B JOINT 1 ན་ → ཡོང་གི་རེད་ (3 asyndetic descriptors of an arising state) C JOINT 2 ན་ → འགྲོ་ཡི་རེད་ (built on B via ཐོག་ལ་ ; stylised couplet) D NAMING འདི་རེད་
Two consecutive joints, chained; one procedural stage feeding them; a terminal label. Structurally the passage is a definition by ostension rather than a set of instructions with a comment appended.
4.5 Agency and modality
The distribution of subjects is clean. In Period A and in both protases the practitioner is subject: gazes, recites, stops, leaves, sustains. In both apodoses the practitioner vanishes. སེམས་གནས་སྡོད་མཁན་གཅིག་ཡོང་གི་རེད་ has the state coming rather than being produced; སྔ་མ་འོང་ and ཕྱི་མ་རྒས་ make the thoughts their own subjects; ཐིང་ཐིང་འགྲོ་ཡི་རེད་ is impersonal. Modality tracks the same division. Both apodoses carry V-གི་རེད་, which states how things behave in general rather than promising an outcome to this hearer; D is the flat equative རེད་. Three speech acts, each morphologically marked: do, this is what happens, this is what it is called.
4.6 Lexical threads
ཐིང་ཐིང་ occurs in B inside the compound འཇམ་ཐིང་ཐིང་ and again as the final predicate of C. The ideophone binds the two outcomes, presenting the second as the first intensified rather than as a further result. མཁན་ nominalises three times (ཟེར་མཁན་ in A, སྡོད་མཁན་ in B, ཟེར་མཁན་ in D), the first and last bracketing the passage. སྡོད་/བསྡད་ appears as main verb, as past nominal and as continuative auxiliary. ཙམ་ appears twice in different functions.
4.7 Indeterminacies
Four points require a reading to be chosen rather than read off.
-
བལྟས་ is a bare past stem, admitting narrative-past and procedural-directive readings; the ambient V-གི་རེད་ forms favour the procedural, but the text does not settle it.
-
དེར་ is anaphoric to a referent outside the excerpt, almost certainly physical given སངས་ རྒྱས་ཀྱི་སྐུ་.
-
ཕྱི་མ་རྒས་ predicates senescence of a cognition without gloss; the doubled clause may be emphatic iteration or a transcription artefact of hesitation. The two readings differ in rhythm, not in vector.
-
འགྲོ་ཡི་རེད་དུས་ admits two boundary placements, both yielding a temporal bridge into D.
4.8 Preliminary macro-reading
Read on these elements alone, the passage answers an implicit question — what is སེམས་མི་ རྟོག་པ་, and how is it recognised — by ostension rather than by stipulation, and traces a descending curve of effort: structured activity, cessation, spontaneous arising, bounded re-engagement, spontaneous deepening. Agency and modality both encode the proposition that the practitioner supplies the conditions and not the result; the withheld name prevents the outcome from becoming a target during the procedure.
This reading is revised in Section 7. It attributes to the speaker a set of structural devices, three of which are subsequently shown to be properties of the language rather than choices. The reading is retained here as it was produced, because the revision is part of the paper’s evidence.
5. The French text analysed independently
This section describes the French as a French text. Nothing in it presupposes that the text is a translation.
5.1 Segmentation
Five sentences. Finite verbs in bold.
S1. Maintenant votre regard intérieur en un seul point sur le corps du Bouddha, récitez par exemple un millier de fois le mantra TADYATHA …
S2. Puis vous vous arrêtez , et vous laissez votre esprit complètement détendu, sans aucune altération.
S3. Arrivera un moment où plus aucune pensée ne s’élèvera , l’esprit sera complètement paisible et détendu, sans aucune pensée d’attachement ou d’aversion.
S4. Alors, si vous posez votre esprit dans ce calme, et si vous le maintenez dans cette absence de distraction, eh bien les pensées qui arrivent les unes à la suite des autres deviendront de plus en plus faibles et l’esprit deviendra limpide.
S5. C’est là que se trouve ce qu’on appelle « l’esprit sans pensées ».
Thirteen finite verbs; one non-finite form ( détendu , predicative).
5.2 Connective inventory
| Marker | Type | Function in this text |
|---|---|---|
| Maintenant | temporal stage opener | opens, pointing forward |
| par exemple | exemplifier, post-verbal | scope undetermined (§5.4) |
| Puis | sequential | onset of S2 |
| et (×3) | coordination | equal-rank joins |
| — | asyndeton | S2 to S3: no connective |
| où | relative | scope leaks (§5.4) |
| sans / plus + aucun(e) (×3) | negative framing | three times across two sentences |
| Alors, | consequential | onset of S4 |
| si … et si … | doubled protasis | two conditions, one apodosis |
| eh bien | oral apodosis particle | marks the turn to the payoff |
| C’est là que | identificational cleft | closure |
5.3 Predicate hierarchy, sentence by sentence
S1 carries one finite verb, the imperative récitez . The sentence’s instruction is therefore to recite.
The opening, Maintenant votre regard intérieur en un seul point sur le corps du Bouddha, , is verbless. Two parses are available. Under the first it is elliptical for an imperative with a dropped verb ( [portez] votre regard … ), which leaves the construction incomplete. Under the second it is a nominal absolute — the ablative-absolute pattern, available in literary French — setting a circumstance within which the main predicate holds: “your inner gaze being one-pointed on the Buddha’s body, recite …”. The second parse is grammatical and is the one adopted here, since French admits such absolutes and the comma placement supports it. Under either parse the gaze receives no predicate of its own, and therefore no illocutionary force: it is posited as circumstance, not commanded.
S2 shifts to the present indicative. French licenses this as the présent de procédure of recipes and protocols, so the sentence still reads as instruction; but the mood has changed from S1 without any change in discourse function to motivate it. Two co-ranked actions joined by et . complètement détendu is a resultative predicative on votre esprit ; sans aucune altération attaches loosely, modifying either the manner of leaving or the resulting state.
S3 opens with no connective at all. The sentence is marked in register: Arrivera un moment is a verb-initial presentative inversion, a literary construction whose function is to bring a new entity onto the scene with emphasis. Three futures make the sentence predictive. It stacks three elements: plus aucune pensée ne s’élèvera , inside the relative où ; l’esprit sera complètement paisible et détendu , separated only by a comma; and sans aucune pensée d’attachement ou d’aversion , a prepositional phrase on the second.
What arrives is un moment — a temporal window, with the mental properties predicated inside it. The outcome is thereby framed as episodic.
S4 carries the text’s only conditional, and it is doubled: si … et si … , two coordinated protases discharging into one apodosis. Both protases are anchored anaphorically to S3 by ce calme and cette absence de distraction , so S4 is tightly chained backwards. Alors, opens the sentence; eh bien introduces the apodosis and is a frankly spoken-register particle, standing two sentences after the literary inversion. The apodosis is two coordinated futures: deviendront de plus en plus faibles , a scalar and gradual attenuation of a stream characterised by a relative clause as arriving in succession; and l’esprit deviendra limpide . Both are devenir : becomings in the entities themselves.
Nothing in the protases bounds the activity. Si vous le maintenez dans cette absence de distraction places no ceiling on the maintaining.
S5 is a cleft focusing là . In this construction là is not primarily spatial: c’est là que is the standard French presentative for “this is the point at which”. The cleft is French’s syntactic device for terminal focus. Ce qu’on appelle , with guillemets, presents the term as received vocabulary.
5.4 Scope indeterminacies
Three, mutually independent.
-
par exemple , placed between récitez and un millier de fois , admits three scopes: the quantity, the mantra, or the whole act. Post-verbal placement favours the nearest quantified constituent, i.e. the count. Nothing disambiguates.
-
où in S3: French would normally repeat où or insert et to keep l’esprit sera … inside the relative. With only a comma, the clause may be read as a second property of un moment or as a juxtaposed independent clause.
-
sans aucune altération in S2, as above.
5.5 Macro-architecture
[Maintenant — anchored forward]
S1 IMPERATIVE one finite verb (récitez); verbless absolute preceding
S2 PRES-PROCEDURAL two coordinated actions
— no connective
S3 PREDICTION literary inversion + 3 futures; "un moment" arrives;
three stacked properties, the second unmoored
S4 Alors + si … et si … → eh bien + 2 futures; anaphoric to S3
S5 CLEFT IDENTIFICATION
One conditional joint, in S4. S3 is an unconditioned prediction standing between two procedural sentences with no junction to either.
5.6 Agency and modality
Vous holds subject position in S1, S2 and both protases of S4. It vacates in S3 and in S4’s apodosis, where the subjects become un moment , aucune pensée , l’esprit , les pensées , l’esprit . The text therefore relocates agency away from the addressee across its predictive stretches.
The modal sequence is imperative, present-procedural, future-predictive, conditional plus future-predictive, present-identificational. Five settings across five sentences. The S1-toS2 shift is the one unmotivated by any change in function: both sentences tell the reader what to do.
5.7 Lexical texture
Détendu occurs in S2 and again in S3, for the state produced and the state that arrives. The negative frame sans / plus + aucun(e) occurs three times within two consecutive sentences, and the three instances are not structurally parallel, so the density reads as a tic rather than as patterning. The quality vocabulary — paisible , détendu , limpide , faibles — is uniformly of calm and attenuation, but the two outcome states share no term: nothing links S4’s limpide to S3’s paisible . Register runs high ( Arrivera un moment , limpide ) and low ( eh bien , vous vous arrêtez ) without settling.
5.8 Macro-reading
The text gives a reader one concrete and countable thing to do; a release, narrated rather than commanded; a detached, high-register prediction that an experience will come; a familiar delivery of a further improvement conditional on settling into it; and a label located on the route. Its climactic sentence, marked as such by register, is a promise, and it stands before any condition is stated. The resulting shape is a reassurance structure: an instruction, then a payoff, then a refinement, then a name. The register instability across that route gives the text the character of speech partly smoothed into prose: some sentences carry editorial polish, others conversational residue, and the voice does not settle on whether it is a manual or a talk.
6. First comparison, conducted against source form
This section reports the comparison as it was first produced, ascending from the smallest units and reusing each level’s findings at the next. It takes source form as the norm. Its findings are reported here as data; Section 8 withdraws most of them.
6.1 Level 1 — predicates
The gross counts: Tibetan, seven finite to nineteen non-finite; French, thirteen finite to one non-finite. The correspondences then show a consistent pattern of reassignment (full table at Appendix C):
-
The matrix verb of Period A, བལྟས་, receives no French predicate at all; the repeatedverb envelope collapses to a verbless phrase. The most deeply subordinated Tibetan predicate, བཟླ་བ་ under ཚེ་ན་, becomes the only French imperative. Rank inversion.
-
བཞག་བཞག་ན་, a conditional protasis, becomes a finite declarative.
-
Three attributive descriptors become two finite futures and one prepositional phrase. བཅུག་, permissive, becomes poser , agentive. ཙམ་དུ་, limitative, is deleted.
-
Three finite clauses of the couplet become one relative clause. Rank inversion. ཕར་མེད་པ་ཆགས་, becoming absent, becomes de plus en plus faibles , scalar fading. Change of event type.
Two regularities were extracted. First, promotion appeared selective: what becomes finite in French is consistently the countable and performable element, while what is deleted or demoted is consistently the stative, hedged or iconic. Second, hedges appeared to have no landing site: ཙམ་ twice, the permissive, the generic auxiliaries — none survives as a hedge.
6.2 Level 2 — argument structure and subject selection
Both texts evacuate the addressee from subject position in exactly the predictive zones. Recorded as the one clean convergence.
One divergence: the Tibetan apodosis installs སེམས་གནས་སྡོད་མཁན་གཅིག་, an entity; the French installs un moment , a temporal window. A state can be built upon; a moment cannot. Carried forward to Level 4.
6.3 Level 3 — clause rank
Period A and S1 carry one finite verb each, but a different one, and the demoted material is handled by different mechanisms. Period B’s single predication with three co-ranked attributes becomes five French assertions, so that what were criteria for recognising one thing become three things that will happen; and the middle one is unmoored from its relative. Period C’s three foregrounded clauses are demoted to a modifier. Recorded verdict: the prominence hierarchies are near-inverted at the two joints that carry the argument.
6.4 Level 4 — junctions
Three findings. Tibetan has two ན་ joints, French one si ; the lost one is the first, so the arrival of the state becomes unconditioned, with the doubled si in S4 concentrating into one late joint the force the source distributed across two. The basis anchor ཐོག་ལ་, which attaches to an entity, is replaced by anaphora to properties, a degradation following directly from Level 2. And the opening deixis flips from anaphoric (དེར་, pointing outside) to cataphoric ( Maintenant , pointing forward), erasing the external support that the source’s sequence requires as a starting point.
6.5 Level 5 — period architecture
Four Tibetan periods, five French sentences. The placement of the extra boundary is the point: the French splits the first consecutive construction at the protasis-apodosis seam. The Level-4 asyndeton and the Level-5 boundary are one event seen from two heights. Convergence at the terminal position: Period D and S5 are coextensive and both final.
6.6 Level 6 — macro-architecture
|TIBETAN|FRENCH|
|—|—|
|[དེར་— anchored outside]|[Maintenant — anchored forward]|
|A PROCEDURE|S1 IMPERATIVE: recite|
|gaze = act, recitation = frame|+ verbless phrase (gaze)|
||S2 PROCEDURE, narrated|
|B JOINT 1ན་→ཡོང་གི་རེད་|— no junction|
|a STATE arrives|S3 UNCONDITIONED PROMISE|
|3 co-ranked criteria|a MOMENT arrives, 3 assertions|
|C JOINT 2ན་→འགྲོ་ཡི་རེད་ |S4 JOINT (doubled si) → 2 futures|
|built on B viaཐོག་ལ་ |anchored to S3's properties|
|effort capped byཙམ་དུ་ |effort uncapped|
|D NAMING:འདི་རེད་(equative)|S5 NAMING: cleft (locative)|
Recorded as: Tibetan builds two chained joints and one state maturing through two depths, with the name withheld; French builds one promise plus one joint, two separate outcomes in sequence, and a name located. Four mechanisms were identified as jointly producing the second shape — the broken ཐིང་ཐིང་ thread, the installation of un moment , the splitting of the single predication into five, and the property-to-property anchor.
6.7 Level 7 — modal spine
V-གི་རེད་ states how things behave; a simple future addressed to vous states what will happen to this reader. Compounded with Level 4: an unconditioned future addressed to vous is maximally promissory, where a future standing as the consequent of a surviving si would have been damped. Two independent mechanisms acting on the same clause in the same direction.
6.8 Level 8 — the verdict reached
Five devices were attributed to the source and scored:
| Source device | Fate recorded |
|---|---|
| Subordinate the technique | Reversed — it becomes the only imperative |
| Pivot on cessation | Severed — the stop survives, its causal link removed |
| Outcomes non-agentive | Kept |
| Cap the sustaining | Reversed — cap deleted, permissive made agentive |
| Withhold the name | Kept |
Two of five kept, one severed, two reversed — and the two reversed were those carrying the passage’s argument. The verdict: a text locally defensible nearly everywhere whose architecture nonetheless asserts perform this practice diligently, these results will come to you, then work to maintain them and they will improve , which is approximately the proposition the source was built to prevent.
This verdict is substantially wrong. Section 7 shows why.
7. Typological re-sorting
Every feature catalogued in Section 4 falls into one of three bins. Only the third can carry message, because only the third represents a choice the speaker made.
|Feature|Bin|
|—|—|
|One finite verb per period, the rest converb
or nominalised|Typological. Head-final clause chaining. Not
a prominence claim.|
|Verb in final position|Typological. Not focus.|
|Context before predicate at every scale|Typological, reinforced by genre.|
|ན་/ནས་/སྟེ་ as the chaining inventory|Typological.ན་here is sequential-
consecutive, not hypothetical.|
|Flat juxtaposition of micro-movements;དེ་
ནས་as the visible joint|Typological and generic. The unmarked way
to build an exposition.|
|Repetition:བཞག་བཞག་,ཕྱི་མ་རྒས་×2,གནས་/བསྡད་
×3|Typological. Repetition is the_unmarked_
device for durativity, iteration and
intensification.|
|Concrete figure for abstract process (རྒས་of
a thought)|Typological. Low-abstraction encoding is the
default.|
|Point deferred to the end of the exposition|Genre. Shedra exposition does this as a
matter of course.|
|ཙམ་as measure-specifier|Typological. Routine.|
|ཚེ་ན་chosen to frame the recitation rather
than coordinating it|Marked.|
|བཅུག་, permissive, rather than a plain
causative|Marked.|
|V-གི་རེད་rather than V-ཡིན་or V-ཡོང་|Marked.|
|ཙམ་inསེམས་མ་ཡེངས་པ་ཙམ་དུ་, where no measure
was required|Marked, weakly.|
|ཐིང་ཐིང་recurring across both outcomes|Marked.|
|ག་ལེར་|Marked, lexical.|
Six marked items, not fifteen devices.
7.1 Consequences for the preliminary macro-reading
The reading offered in §4.8 — that the speaker dismantles the assumption that the result is manufactured by the technique, using five structural devices — over-reads the evidence. Three of those five devices belong to bin one:
-
“Subordinating the technique” was inferred from the finite/non-finite ratio in Period A. The ratio is fixed. What the speaker actually chose was to place the recitation under ཚེ་ན་ rather than coordinating it; that choice asserts a relation (the recitation is the container of the gazing), but it does not assert a ranking of importance.
-
“Pivoting on cessation” was inferred from the position and the converb at མཚམས་བཞག་ ནས་. Sequence is what ནས་ does. The cessation is in the content, not in the structure. “Withholding the name” was inferred from the final position of སེམས་མི་རྟོག་པ་. Terminal placement of the point is the convention of the genre, which the speaker is following rather than deploying.
The anti-effort reading is still present in the passage, but it rests on the marked items alone — the permissive བཅུག་, the unrequired ཙམ་, ག་ལེར་, and above all the generic morphology — which is a far narrower base than Section 4.8 claimed and a correspondingly narrower target for the comparison.
7.2 Consequences for the first comparison
Almost every finding in Section 6 concerns bin one. Level 1’s finiteness counts, Level 3’s rank inversions, Level 5’s period arithmetic and Level 6’s “near-inverted hierarchies” are all measurements of typological difference. Level 4’s junction count rests on construing ན་ as hypothetical, which it is not in procedural Tibetan — rendering every ན་ with si is itself a known over-translation and produces precisely the legalistic French that fluent renderings avoid.
What survives the sort as a candidate loss: the four marked items that the French did not reproduce, and the generic-to-promissory substitution. These are tested in Sections 8 to 10.
7.3 A fourth bin: functions discharged outside the running text
The three-way sort above is incomplete. Some features are present neither as typological default, nor as genre convention, nor as the speaker’s rhetorical choice, but because of the conditions under which the source was produced and received — and their function is not one that target text carries. Two cases, from different genres.
The clearest case is segment-final ཟེར་. Taken as a quotative it looks like a citation marker; taken in its setting it does three things at once — hands off the turn at a segment boundary, deflects authorship away from both parties present, and marks the segment as a summation. Only the second and third have anything to render. The first has an addressee who exists in the performance situation and in no written target, so a translator who renders it faithfully imports machinery from a step that does not occur. The corpus translator’s handling bears this out: a summative marker ( Voilà , Donc voilà , au fond ), and the draft’s est-il dit deleted.
The second case comes from the written tradition. A Tibetan śāstra renders its own outline in the running text : དང་པོ་ … གཉིས་པ་ … , the divisions announced དེ་ལ་གསུམ། and then in words as the text proceeds, because the pecha format and the recitational mode of transmission afford no typographic hierarchy. A Western edition discharges the same function by layout — headings, numbered sections, a table of contents, running heads — so a translator faces a question with no textual answer: render the sapche as sentences, convert it into structure, or both.
The two cases differ instructively. ཟེར་’s handoff has no counterpart in any written target and simply drops: addressee loss . The sapche’s navigation has a counterpart, genuinely needed, realised in layout rather than in words: channel migration . What they share, and what separates both from bins one to three, is that faithful textual rendering is the wrong operation.
That these are instances of one phenomenon rather than two oddities can be seen from a third observation: the same tradition realises the same navigational function through different channels according to medium. In the oral corpus of §10.4 the གང་རེད་ཟེར་ན་ question frames do the navigational work and stay in the text, because linear delivery affords no random access and a listener needs the structure spoken. In the written śāstra the sapche does it, and migrates to layout. One function, one tradition, the channel set by the medium. Bin four is therefore a systematic category with a medium variable, not a residue.
This bin behaves differently from the other three at assessment time. A bin-one feature is not a loss when absent because it was never an assertion; a bin-four feature is not a loss when absent because its addressee is gone or its function has moved channel. Both are indistinguishable from real loss in a correspondence table, and are told apart only by knowing how the source was produced and how the target is to be presented — which makes delivery conditions and target format part of the minimum context for assessment, not background.
The channel-migration case sharpens this to a point. Convert a sapche into a heading and a text-to-text comparison records a deletion, while the function sits intact in the layout where the comparison cannot see it. That is §11.2’s self-concealment in a second guise, and it is the strongest argument for keeping the bin on thin evidence: the category exists precisely where the standard instrument is blind. The open questions this raises are listed under Future work.
8. Second comparison, conducted by function delivered
The same correspondences, re-scored. The question at each point is not did the target match the source’s form but did the target deliver the source’s function, by whatever means its own language affords .
8.1 Period A and S1 — frame and figure
The frame-figure assignment is genuinely swapped: Tibetan frames the recitation under ཚེ་ན་ and predicates the gazing; French frames the gazing in a nominal absolute and predicates the recitation. The swap is real and is the one place where a marked source choice (§7) was not matched.
But the label anacoluthon , applied to the French absolute in the first pass, must be withdrawn. Read as a nominal absolute the construction is grammatical and performs exactly the backgrounding that a Tibetan subordinate clause performs. The first pass reached “anacoluthon” by importing the source’s structure as an expectation: a finite verb was looked for because the source had one there.
At the level of units delivered, both texts carry: one-pointed attention; directed at the Buddha’s form; concurrent with roughly a thousand recitations; of a specified mantra. Which element is foregrounded is an encoding fact. No idea lost.
8.2 Joint 1 — consecution
The claim that the joint was lost does not hold. ན་ in procedural Tibetan is consecutive. French delivers consecution by procedural sequence after Puis , which is its own device for the same job, and the contingency on manner survives because sans aucune altération attaches to laissez . The idea — the state arises as a consequence of unaltered resting — is present in the French. Downgraded from failure to re-encoding.
8.3 The couplet — iteration as scalar
This is the clearest reversal, and it is the paper’s central demonstration.
| Tibetan | French | Operation |
|---|---|---|
| སྔ་མ་/ཕྱི་མ་ | les unes à la suite des autres |
ordinal pair → successive phrase |
| རྒས་(grow old) | faibles | concrete figure → abstracted attenuation |
| doubled clause (iconic) |
de plus en plus | iconic repetition → scalar adverbial |
The third row is the whole of §2.3 in one move. Repetition is the unmarked Tibetan device for iteration; a French translator who preserved it would have introduced a stylistic signal — in a pedagogical text, a signal of incompetence — absent from the source. The scalar adverbial is the faithful rendering. It follows that the first pass’s two findings here were both errors of the same kind: the demotion to a relative clause is the obligatory nesting operation, and “vanishing → scalar fading” mistook the rendering of iteration for a substitution of event type. Both withdrawn.
8.4 The ཐིང་ཐིང་ thread — repetition as progression
The first pass recorded the thread as broken by paisible / limpide . The distribution shows otherwise:
འཇམ་ + ཐིང་ཐིང་ + ལྷོད་དེ་བ་ (first outcome) → paisible + détendu
ཐིང་ཐིང་ (second outcome) → limpide
The French splits the compound, assigns འཇམ་ and ལྷོད་ to the first outcome, and reserves ཐིང་ཐིང་ for the second, where limpide renders the settled-clarity sense. The thread is not broken; it is converted from lexical repetition into a lexical progression — smooth and loose first, limpid at the end — with l’esprit holding constant as subject across both. This is the correct operation under §2.3, since repeating the same French adjective would have marked what the Tibetan leaves unmarked. Withdrawn, and recorded as a solution rather than a defect.
The same evidence dissolves the first pass’s “two separate outcomes rather than one maturing state”. Constant subject, anaphoric ce calme , and a graded quality progression chain the two phases.
8.5 S4 — flat repetition as anaphora
The source states the abiding three times: དེ་གནས་བསྡད་པ་ / དེའི་ཐོག་ལ་ / གནས་སུ་བཅུག་སྟེ་. French gives si vous posez votre esprit dans ce calme : one movement, the triple repetition collapsed, the referent carried by anaphoric demonstrative.
Below discourse level this looks like a rewrite — three predications become one, two of three mentions of གནས་ disappear. At discourse level it is exact: flat repetition is the Tibetan encoding of “the state you have just reached, on which you now build”, and anaphoric demonstrative plus a single predicate is the French encoding of the same thing. The encoding changed layer; nothing was lost.
8.6 Period D and S5 — positional focus as cleft
Tibetan obtains terminal identificational focus without cost: head-final order, final equative རེད་, deictic འདི་. French has no positional focus available and uses the cleft. C’est là que is the standard French presentative for “this is the point at which”; ce qu’on appelle with guillemets renders ཟེར་མཁན་ precisely, distancing included. The first pass’s “equative deixis degraded to a locative metaphor” was wrong twice: là is not read spatially here, and the cleft is not a degradation but the functional equivalent of what the source achieves by position. Withdrawn. This is arguably the best-handled moment in the target text.
8.7 Register particles
Eh bien , flagged in the first pass as a register clash against Arrivera un moment , is better read as the French rendering of oral Tibetan’s discourse-marker texture. The source’s own register is mixed, placing colloquial ཟེར་མཁན་ beside technical སེམས་མི་རྟོག་པ་. A nested French text retaining one or two oral particles holds both signals at once — nesting carries authoritative , the particles carry spoken . See §11.1 for why the nesting is not optional. Withdrawn.
8.8 What survives the second pass
Four of the six marked source items were not reproduced: ག་ལེར་, the ཚེ་ན་ framing, the permissive བཅུག་, and the generic V-གི་རེད་. ཐིང་ཐིང་ was reproduced by redistribution; ཙམ་ was reproduced as measure but not as limitative. These are carried into the audit.
9. Meaning-unit audit
The criterion of §2.5 is applied exhaustively. Twenty-six meaning units were identified in the source and each tracked into the target. Status values: present (delivered by any means), shifted (delivered in altered form), thinned (implied but not stated), absent .
| # | Meaning unit | Status in French |
|---|---|---|
| 1 | one-pointed attention | present |
| 2 | object: the Buddha’s form | present |
| 3 | mantra X given asexemplar of a class | shifted —_par exemple_reads onto the count or the act |
| 4 | count approximately one thousand | present (_un millier_carries the approximation) |
| 5 | recitation concurrent with attention | present (frame and figure swapped) |
| 6 | ocularབལྟ་and mentalདམིགས་as two faculties |
open— see §12.1 |
| 7 | breaking off the practice | present |
| 8 | gently, at ease (ག་ལེར་) | absent |
| # | Meaning unit | Status in French |
| 9 | without contriving | present |
| 10 | just leaving it (reduplicative durative) | thinned |
| 11 | consecution: unaltered resting → the state |
present (Puis+ procedural juxtaposition) |
| 12 | no thought proliferates | present |
| 13 | no thought of attachment or aversion at all |
present |
| 14 | soft, still, loose | present (distributed, §8.4) |
| 15 | a condition of settled mind comes about |
shifted —_un moment_rather than a state |
| 16 | the practitioner is not the agent of the arising |
present |
| 17 | building on that abiding | present (anaphoric_ce calme_) |
| 18 | letting it rest (permissiveབཅུག་) | shifted — agentive_poser_ |
| 19 | sustaining at the measure of non- distraction |
present as measure; limitative overtone thinned |
| 20 | continuative aspect of the sustaining | present (maintenez) |
| 21 | successive thoughts, each more worn | present (§8.3) |
| 22 | the thoughts become absent, away | thinned — implied by_limpide_, not stated; ཕར་absent |
| 23 | stillness and clarity result | present (limpide) |
| 24 | the outcome stated as generic, not promised |
shifted — addressee-directed simple future |
| 25 | at that juncture | present (là) |
| 26 | this is what is called non-conceptual mind |
present |
9.1 The availability test
For each divergence, §2.5 requires a second question: did French hold a resource for the function which the translator did not use?
|Divergence|Marked in
source?|Resource available in
French?|Counts as loss?|
|—|—|—|—|
|8 —ག་ལེར་absent|yes|yes (doucement, one word)|yes|
|3 — exemplification
scope|yes|yes (un mantra tel que …)|yes|
|18 — permissive to
agentive|yes|yes (laissez votre esprit
reposer dans ce calme)|yes|
|24 — generic to
promissory|yes|yes — see §9.2|yes|
|22 — extinction and
direction|no (default
figuration)|partly|marginal|
|19 — limitative
overtone|weakly|ne serait-ce que,
simplement|marginal|
|10 — reduplicative
durative|no (typological)|—|no|
|15 — state to moment|no|—|no (content
delivered)|
9.2 The modal substitution tested
Unit 24 is the one divergence that touches propositional content. V-གི་རེད་ states how things behave; vous plus simple future states what will happen to this hearer. These are different claims, and the difference is epistemic rather than stylistic.
The availability test is decisive, because French holds the resource and the translator declined it at all five opportunities:
Il arrive alors un moment où plus aucune pensée ne s’élève … les pensées deviennent de plus en plus faibles et l’esprit devient limpide.
Impersonal il arrive with the present of general truth is idiomatic French — arguably more idiomatic here than the original — and carries the genericity. On the evidence of one passage this looked like a translator’s choice. The larger corpus shows it is not. §10.6 sets out the mechanism: the substitution is the terminus of a four-step cascade, each step locally forced, which begins wherever the source stacks nominalised complements under a single copula. The drift is therefore systematic rather than occasional. The availability test was correctly applied and gave the right answer about the resource while giving the wrong answer about agency — the construction exists, and nothing in the language routes a translator to it.
Combined with the two other unreproduced marked items that pull the same way (ག་ལེར་, the permissive བཅུག་), it tilts the passage toward a method-with-payoff reading — the conclusion of the first pass, but reached on one mechanism instead of five. On the corpus evidence that one mechanism is more consequential than the first pass’s five would have been, because it fires at a frequent construction rather than at one passage’s clauses; and less culpable, because it is not elected.
10. Results
10.1 The loss inventory
Of twenty-six meaning units, one is absent, four are shifted, three are thinned, one is undecidable on the available evidence, and seventeen are delivered unaltered. Applying the two-condition test of §2.5:
| Loss | Kind | Avoidable |
|---|---|---|
| Generic morphology → | propositional | the construction exists; not |
| addressee-directed future | (epistemic modality) | elected (§10.6) |
| ག་ལེར་“gently, at ease” | lexical, one word | yes |
| Exemplification scope (the mantra as exemplar) |
scope, small | yes |
| Permissiveབཅུག་→ agentive poser |
lexical-semantic, small | yes |
| Extinction and direction of the ceasing (ཕར་) |
thinned | partly |
| Limitative overtone ofཙམ་ | thinned | marginally |
One loss touches the message. Four are small and individually arguable. All five that are avoidable are repairable without restructuring any sentence.
10.2 The two verdicts compared
| First pass (source form) | Second pass (function delivered) |
|
|---|---|---|
| Source devices identified | 5 | 6 marked items |
| Scored as reversed or severed |
3 | 0 |
| Scored as lost | “the argument of the passage” |
1 propositional unit, 4 minor |
| Verdict | architectural failure | good discourse-level translation |
The difference between the two rows is the measurement of the bias. It is not a difference of degree: the first pass identified the target text’s best solutions — the scalar for iteration, the cleft for positional focus, the anaphora for flat repetition, the redistributed ideophone — and recorded each of them as a defect.
10.3 Direction of error
The first pass’s errors were not randomly distributed. Every one of them ran in the same direction: toward judging the target deficient for not resembling the source. None ran the other way. A method whose errors are unidirectional is not noisy but biased, and its outputs cannot be corrected by averaging or by taking more care — only by changing the procedure.
The bias also concentrated at a specific point. The independent descriptions of Sections 4 and 5 required no retraction; the preliminary macro-reading of §4.8 required partial retraction, since it attributed devices to a speaker who was following his grammar; and everything else withdrawn was generated after comparison was licensed. The defect is in the comparative operation.
10.4 Design of the corroboration study
The findings above rest on nine lines. They were then tested against roughly twenty minutes of transcribed oral teaching in the same genre, same speaker, same register: two Tibetan transcriptions with a French rendering by a translator working in the genre.
One fact bearing on inference: the §3.1 French and the corpus French are by the same hand — the specimen an in-progress draft, the corpus finished work. This strengthens the corroboration wherever it concerns settled practice, since a pattern appearing in both is a convention rather than an occasion. It also answers the question §9.2 left open: the specimen’s addressee-directed futures are not an isolated choice but the translator’s norm, visible across twenty minutes of finished output. What the corpus still does not establish is that the norm is representative of the genre at large.
10.5 Corroborated
The definition pattern (§4.3, §7.1). At least ten instances of … ཟེར་མཁན་དེ་རེད་ closing a definiens. The paper’s strongest structural claim now has a frequency rather than a single instance. The corpus also yields a variant the passage did not contain and which is more frequent still: the rhetorical question frame X གང་རེད་ཟེར་ན་ … Y རེད་ , past twenty instances, rendered by the corpus translator as a French question with the answer introduced by a particle.
The finiteness ratio (§2.1). Corroborated, and the paper’s description was over-specified while its conclusion was understated. The ratio is not one finite verb per period: four consecutive periods at A_01 subs 83–85 contain no verb at all, other periods carry a twoword finite clause, others long converb chains. A quantity that swings from zero to several within one speaker in one genre cannot encode prominence under any reading.
ན་ (§8.2). Corroborated with better evidence and refined. Across one file the corpus translator’s renderings divide roughly evenly between si and four other constructions: a relative ( pour celui qui possède… ), a temporal ( tandis que , lorsque ), a causal ( étant donné que ), a manner ( en procédant ainsi ), with one instance absorbed into an alternation ( soit … soit ). So ན་ selects among at least five French constructions and si is right about half the time. The claim that ན་ is not automatically si stands; any advice to avoid si does not. Limitatives (§10.1). Corroborated. ཐ་ན་ → ne fût-ce que ; ཙམ་གྱིས་ → en rester là . The devices exist and are used in the genre, which retrospectively confirms the classification of the ཙམ་ loss as avoidable.
ཟེར་མཁན་ and figurative predication (§8, §2.4). Both corroborated: ཟེར་མཁན་ dropped where it carries an established category and kept where the naming is the point; སྐྱག་ལང་ demetaphorised at every occurrence.
10.6 Revised: the mechanism behind the single propositional loss
The corpus contains the construction that produces the drift, in a form clear enough to trace. At A_01 subs 17–20 four nominalised complements stand under one རྒྱུ་ཡིན་པས་ན་ — that is the cause of [blocking the door], [pacifying suffering], [freeing from the ocean], [approaching liberation]. The French renders them as three finite futures, with three different subjects, across three sentences: suffira , vous apaiserez , vous permettra . The cascade, each step locally forced:
-
French nominalisation is heavy — le fait de bloquer la porte , la pacification de la souffrance du karma . Three or four stacked under c’est la cause de produce administrative prose.
-
So the complements are broken out as finite clauses.
-
A finite clause requires a tense, and the only tense that fits a consequence is the future.
-
A future consequence requires a subject, and the one available is the addressee.
Nobody elects to convert a characterisation into a promise; the morphology arrives there. Three further losses ride along. The single predication becomes three claims. The grounding function of ཡིན་པས་ན་ disappears, so the premises are spent as promises and what they were grounding — that such a Dharma is being taught and heard is extraordinary fortune — arrives unmotivated two sentences later. And the attribute becomes an outcome.
One operation repairs all of it: infinitival or relative subjects under a single predication, in the present.
Entendre ces mots ferme la porte des renaissances inférieures et du samsara ; en pratiquer le sens apaise la souffrance du karma et des émotions perturbatrices ; en garder ne fût-ce que quatre vers libère de l’océan du samsara et rapproche de la libération et de l’omniscience.
One predication, four complements, no future, no addressee subject, and the list remains available as a premise. The construction is idiomatic French; it is simply not where the language lands by default.
10.7 Revised: the encoding findings
དེ་ནས་ . §6.4 and Appendix D record the A→B joint as matched (དེ་ནས་ → Puis ). The corpus shows no fixed correspondent: Alors , Ensuite , Puis , Donc , Maintenant , and frequently nothing at all. Since དེ་ནས་ is the corpus’s most frequent discourse marker, “matched” was an artefact of a single token.
Oral features do not uniformly falsify register. The rhetorical question-answer chain and its answer particle transfer wholesale — about a dozen instances of eh bien in the corpus French occupying the ཟེར་ན་ answer slot, as a settled device rather than a lapse. §5.7’s charge of register oscillation against the §3.1 French and §8.7’s rehabilitation of it were both reasoning from one token; the rehabilitation reached the right verdict for the wrong reason.
Repetition and the abstraction differential are refined in place at §2.3 and §2.4. §11.1 is recalibrated in place.
10.8 The self-concealing property, corroborated unintentionally
During the corroboration the author assembled a six-item critique of the corpus translator’s French. Four items were withdrawn under correction, each for the same reason: an element of the Tibetan was sought at the position its French correspondent should occupy, not found there, and recorded as a loss — while the correspondent stood elsewhere in the sentence.
| Item | What was actually there |
|---|---|
| ཡག་པོ་རེད་“deleted” | absorbed into_il convient absolument_, withན་re- placed around the whole |
| causalཡིན་པས་ན་“dropped” | resolving two sentences later at_Alors, comme un_ affamé… écoutez |
| contrast “imported” by_tandis que_+ cleft |
licensed byརང་གི་inཆོས་ཉན་པ་པོ་རང་གི་རྒྱུད་ལ་against གཉིས་ཀའི་ |
| topicalisation “stranded” | a licensed French construction;le sandwich, c’est _moi qui l’ai mangé_places the resumptive inside the relative of a cleft, two levels below the topic |
The six items had already accumulated into a pattern — five of six clustering where the matrix verb stood furthest from the clause — which collapsed with the withdrawals. §11.2 predicts precisely this: granular, individually checkable findings whose accumulation produces confidence, every one of them a measurement of the method rather than of the text.
Two features of the episode bear on §11.3 and §11.5.
First, the procedure was written and was not run. It was bypassed twice in one working session, both times while moving quickly and feeling well evidenced. A procedural constraint that depends on the assessor electing to apply it does not bind; it has to gate output rather than sit beside it as advice.
Second, the findings that survived were countable. A lexeme repeated across adjacent sentences; a predication count that did not match the source’s; an agreement mismatch; a dual reference flattened to a universal. Every finding that rested on a judgement of whether a target construction was licensed or too heavy was withdrawn — including one that asserted a constraint on French topicalisation which does not exist. This is a sharper division of competence than §11.5 proposed, and it cuts against that section’s suggestion that judgement of the target as a text is the reliable half. It is not. The reliable half is narrower: countable checks within the target and against the source, and no register judgement in either language.
11. Discussion
11.1 Register value, and why structural preservation falsifies it
A natural objection to §2.3 and §2.4 is that French also possesses a flat, juxtaposed, loosely-linked mode — the mode of ordinary speech. Vous savez, mon cousin, eh bien son père, il a un vélo. Et alors, son vélo, le guidon de son vélo, la couleur du guidon, eh bien, c’est du bleu stands to la couleur du guidon du vélo du père de mon cousin est bleue roughly as chained Tibetan stands to nested French. If French has the flat mode, why not use it, and preserve the source’s organisation exactly?
Because the social value of the two flatnesses is opposite. Flat juxtaposition in Tibetan expository speech is the register of authoritative oral teaching; flat juxtaposition in French is the register of the uneducated speaker. The two languages distribute the same structural resource across opposite points of the prestige scale. A French text that preserved the source’s organisation wholesale would therefore read as produced by someone who cannot organise his thoughts, where the source reads as produced by a master explaining something. The corpus requires this to be stated with more care than it first received; see the recalibration below.
The consequence is strong and worth stating plainly: preserving the structure would have falsified the register. Nesting is not a concession to French taste; it is the operation that holds constant the one thing a reader most reliably infers — who is speaking and with what authority. And it is this that rehabilitates the register particles of §8.7: a nested French carrying one or two oral markers signals authoritative through its syntax and spoken through its lexicon, which is exactly the source’s own combination.
The same argument covers abstraction. Raising the abstraction level of རྒས་ — a thought growing old — to de plus en plus faibles is not the loss of a figure. In Tibetan the concrete predication is unmarked and carries no poetic charge; in French it would be marked and would carry one. Transposing the figure intact would have added imagery the source does not have. This is the structural reason literal translations from Tibetan sound, as practitioners report, childish at worst and awkward at best: they transfer unmarked source forms into marked target slots, and the markedness is read by the target reader as authorial intent.
Recalibration. The argument form holds: the same structural resource can carry opposite social value in two languages, and where it does, preserving it falsifies register. The conclusion first drawn from it was too broad. What does not transfer into French is flat predicate juxtaposition across successive clauses, and repetition. What does transfer — and in the corpus transfers wholesale — is the rhetorical question-answer chain with its answer particle, and topicalised information packaging, which French shares with the source’s own left-branching order: la joie que procurent ces quatre considérations, elle est semblable à… ; ces pensées, elles sont comme les bulles qui se forment d’un moment à l’autre à la surface de l’eau . French left dislocation with clitic resumption is licensed, and permissive about where the resumptive lands — le sandwich, c’est moi qui l’ai mangé places it inside the relative clause of a cleft, two levels below the topic.
So the operation is not flat-to-nested. It is better described as gathering the clause chain into one complex topic and predicating once , which stays left-branching like the source, and which fails only where the topic’s internal material is itself a calque: les comportements de nos corps, parole et esprit , one determiner over a plural and two singulars, is licensed by no French register. The distinction matters because the two failure modes call for opposite repairs — one asks for restructuring, the other only for determiners.
11.2 Why source-form assessment is self-concealing
The bias documented here is unusually durable because the method that produces it feels more rigorous than the method that corrects it.
A literal translation shares surface structure with its source. Laying the two side by side produces repeated local confirmations, which the assessor experiences as evidence of fidelity. A discourse-level translation diverges in surface structure, so the same operation produces repeated local discrepancies, which the assessor experiences as detected error. Both experiences are artefacts of the comparison method, and neither is information about the target’s fidelity.
The experience is also granular and cumulative, which makes it persuasive. The first pass recorded eight predicate-level divergences, three rank inversions, a lost junction, a shifted boundary and a modal substitution, each individually checkable. It was the accumulation of checkable small findings that produced confidence in the macro verdict — and all but one of them were measurements of typology.
11.3 The proposed procedure
Four steps, in order. The order is not optional; steps 1 and 2 must complete before any evaluative vocabulary is admitted.
-
Describe each text independently , in isolation, without the vocabulary of loss, addition or correspondence, and without the fact of translation entering the description of the target.
-
Establish the unmarked encoding of each discourse function in each language, and subtract it. Only deviations from the unmarked are carried forward.
- Compare functions delivered , not forms matched. For each surface divergence, state first what function the target construction performs in its own language; a loss may be recorded only after that statement has been attempted and has failed.
- Apply the availability test. A divergence counts as loss only if the source choice was marked and the target language demonstrably held a resource for the same function that went unused. Where the first condition fails, the assessor is measuring typology. Where the second fails, the assessor is recording an untranslatability, which is a property of the language pair.
Step 3’s requirement that the charitable hypothesis be stated first is what catches the ཐིང་ ཐིང་ case. The first pass never asked what function the paisible / limpide distribution served; it recorded a broken thread and moved on. Step 4 is what isolated the one real loss, and would have cleared the other five candidates before any of them was written up.
11.4 Implications for corpus-based norm extraction
The result bears directly on projects that extract translation norms from aligned corpora, whether for human reference or to condition a model.
Norm extraction inherits whatever its exemplar corpus contains. An exemplar set weighted toward the philologically literal tradition — which is where the bulk of published Tibetan-to-Romance translation sits — will yield norms encoding precisely the bias this paper documents: preserve repetition, preserve finiteness distribution, preserve clause order, preserve concrete figuration. Those norms will moreover appear well attested, because they will be frequent. Frequency in such a corpus measures the prevalence of a convention, not its fidelity.
The implication is that curation of the exemplar set is prior to, and more load-bearing than, the extraction model applied to it. Selecting for discourse-level transposition rather than for philological transparency is a selection criterion, not something a downstream layer can correct, because the layer has no independent signal against which to recognise that the corpus is skewed.
11.5 Implications for machine-assisted assessment
A language model performing this assessment reproduces the bias from a different source than a human assessor does, and the difference matters for how it can be controlled.
A human assessor holds the source-form norm as a methodological commitment, which can be argued with. A model holds it as a statistical prior over what translations from Tibetan look like — and because that corpus is small and skewed, a competent discourselevel translation resembles the prior less than a poor literal one does. The model is then liable to score the better translation worse, not because it applies a theory of fidelity but because it is pattern-matching to a corpus. The failure mode is therefore not a reasoning error that further reflection corrects; it is a calibration error. Two controls follow from §3.2. First, the division of labour: a model’s descriptive capacity and its comparative judgement are differently reliable, because its competence in the target language rests on a vast corpus of native prose while its sense of that language as a translation target rests on a tiny one. It may be used for description and for assessing a target text as a text; it should be structurally distrusted on whether a rendering is faithful. Second, sealing: analysing the target before and independently of the source removes the input that activates the prior.
It should also be noted that better training data is not available as a remedy here. The correction applied in Section 7 requires linguistic training, native Romance intuition, fluency in the spoken source language, and formation in its exegetical tradition, held simultaneously. That combination is rare enough that the correction is largely unwritten, so the gap in any model’s training reflects a real gap in the literature rather than an accident of sampling. The remedy is procedural constraint and expert supervision, not more corpus.
11.6 Recalibration across Romance languages
The findings divide by which side of the operation they describe, and that division determines how far they travel.
The source-side findings travel unchanged to any target, because they concern comprehension and arise before a target language is involved: the definition pattern and its question-answer variant, the markedness sort of §7 including its fourth bin, the range of ན་, zero-argument resolution, the epistemic particles read literally, doublets against repetitions, the pro-verb དེ་འདྲ་བྱས་, and the ambiguity of the agentive-genitive particle series. A translator working into any Romance language can get these wrong in the same way, and the colophon — a single head-final period whose matrix verb stands last after a conventional slot order — is the genre where the failure is most reproducible.
The target-side findings are calibrated to French and must be re-derived for each other target. Three differences in the other major Romance languages are large enough to change the results. They are flagged here, not argued.
-
Null subjects. Spanish, Italian and Portuguese permit unexpressed subjects; French does not. Since the source drops arguments freely, a French translator must resolve a zero subject at every occurrence and can resolve it wrongly, as §4.1’s inverted kinship relation shows. A null-subject target can sometimes leave it unresolved and carry the source’s underspecification forward. The forcing that produces that error class is specific to French.
-
Nominalisation weight. The cascade of §10.6 begins in French’s resistance to stacked nominalised complements. Spanish, Italian and Portuguese have more productive infinitival-subject constructions — and Portuguese has the inflected infinitive, which French lacks entirely — so step 1 may not fire, and with it neither the future nor the addressee subject. The paper’s single propositional loss may be a French loss rather than a Romance one.
-
Topicalisation tolerance. Left dislocation with clitic resumption is more freely available in the written registers of Italian, Spanish and Portuguese than in written French. The §11.1 recalibration is therefore pitched at the most restrictive member of the family, and a target further along that scale would need less restructuring, not more.
Untested and not to be assumed: whether the markedness of repetition (§2.3) and the abstraction differential (§2.4) hold at the same strength across the family. Both are stated here for French, on French evidence, from one translator’s practice.
The practical consequence for the procedure of §11.3 is that step 2 — establish the unmarked encoding of each function in each language, and subtract it — has to be carried out twice, once per language, and cannot be inherited from a neighbouring pair. A set of transposition rules derived for Tibetan–French is a set of hypotheses for Tibetan–Spanish, not a set of results.
12. Limitations and open questions
12.1 The anaphoric opening
The passage begins with དེར་, pointing to a referent established in discourse not available here. Given སངས་རྒྱས་ཀྱི་སྐུ་ in the same period, the referent is plausibly a physical image or statue, in which case the source instructs ocular attention to an external object together with mental focus on the Buddha’s form, and the French regard intérieur has collapsed two faculties into one and dropped the external object. If དེར་ instead refers to a visualisation already established in prior discourse, the French is exact and unit 6 of the audit is present , not open.
This is the only place in the audit where a whole meaning unit may be missing, and it is decidable — by anyone holding the preceding context — rather than genuinely indeterminate. It is recorded as open rather than resolved because resolving it here would require guessing.
12.2 Provenance of the target text
The provenance is now established, and is neither of the two possibilities this section originally weighed. The French is a working draft, still in progress, by the translator whose finished renderings form the corroboration corpus. It is not a published rendering, and not a transcription of live interpretation. The profile described next is real nonetheless; what changes is its explanation.
The target’s specific profile — the verbless opening, the unmotivated imperative-toindicative shift, the missing connective at S3, eh bien , and a register swinging between Arrivera un moment and vous vous arrêtez — is the signature of speech transcribed rather than prose composed. If that is its provenance, the losses recorded in §10.1 should be weighted differently: an interpreter working in real time cannot hold a two-stage consecutive architecture across forty seconds of speech while also tracking a mantra citation and a count, and promoting the countable element to the main verb is what processing load forces. Judged as live interpretation this is respectable work and the losses are a cost of the medium. Judged as a published translation, the five avoidable items in §10.1 should be repaired.
Those features are better explained by the draft state and by the constraints of §10.6 and §11.1 than by conditions of delivery: an unfinished draft carries unresolved decisions, and the cascade and the dislocations are forced by French rather than by speed.
One qualification follows, and it is not trivial. Where §10.1 records a marked source item as absent — ག་ལེར་, the limitative ཙམ་ — a draft in progress may simply not have reached it. Absent and not-yet-rendered are indistinguishable from outside the work, and only the translator can say which. The loss inventory should be read with that qualification against those two rows.
The methodological finding stands either way: the first pass’s verdict was wrong about the text, whatever its provenance.
12.3 Single-passage evidence base
One passage cannot establish the frequency of the bias, only its mechanism. The claim argued here is that source-form assessment produces unidirectional error on a well-made translation and that the four-step procedure removes it in this case. Establishing prevalence would require a corpus study: a set of Tibetan passages with two renderings each, one philologically literal and one discourse-level, assessed blind under both procedures. The prediction is specific and falsifiable — source-form assessment should rank the literal rendering higher, and function-delivered assessment should reverse the ranking. That study remains undone. §10 reports something weaker and differently shaped: a corroboration of the same findings against twenty minutes of the same genre, which establishes frequencies and corrects two mechanisms but compares no two renderings of one passage.
12.4 The couplet’s text
ཕྱི་མ་རྒས་, doubled, may be emphatic iteration or a transcription artefact of hesitation; རྒས་ itself sits one letter from རྒྱས་. The reading adopted here — senescence, hence attenuation — fits the passage’s vector, and the alternatives do not change the audit’s result for units 21 and 22. But a collation against other witnesses of the same teaching would settle it, and has not been attempted.
12.5 Scope of the typological claims
The SOV and head-final characterisations used in §2 are deployed as heuristics for wellattested distributional tendencies, not as a typological taxonomy with explanatory force. The inference from head-finality to discourse-level deferral of the point is a correlated tendency reinforced by genre convention, not an entailment. Nothing in the argument requires the stronger claim: what the procedure needs is only that the baseline be established empirically for the pair in hand, by whatever description.
12.6 A defect in the corroboration corpus
The machine first drafts supplied alongside the corpus and the Tibetan transcription supplied with them are not from the same state of the transcription. One draft annotation quotes a reading it marks incomplete (འཁྲུལ་པས་གཟིགས་) that the supplied transcription does not contain, reading འཁྲུལ་པས་གཟིར་ instead; and one apparent comprehension error in that draft, an inverted kinship relation, traces to a genitive particle standing in the transcription where the audio has an agentive. The source was revised between the draft and the comparison.
This yields a condition on any use of draft-against-revision pairs as evidence: the source must be unchanged between the two. A pair whose source was corrected in between credits the earlier version with an error it did not make, or penalises a reading that was correct when it was made. The findings in §10.5 to §10.7 rest on the finished French against the Tibetan and are unaffected. Observations about the drafts are provisional, and are not relied on in this paper.
12.7 One translator, one genre
The corpus and the §3.1 specimen are by one translator. That strengthens the inference where §10 reports a settled practice, and narrows it elsewhere: one translator’s practice is not shown to be representative of the genre. Where §10 reports a rendering as the genre’s solution, the claim is that it is a competent solution attested in the genre, not that it is the norm.
Future work
Three questions the present material raises and does not settle.
Populating the fourth bin
§7.3 proposes the bin on two cases, one per sub-type. Both need populating and the boundaries tested, and both questions are tractable.
Candidates visible in the corpus, none confirmed: the segment-final recapitulations that close a stretch by restating its point, which read as recaps before a handoff; the distribution of ཟེར་ itself between segment-final position and the ཟེར་མཁན་ and ཟེར་ན་ constructions, which are not scaffolding and must be told apart from it; whether segment length is itself shaped by interpretation; and whether the frequency of དེ་ནས་ at segment onsets is partly a re-entry marker after a handoff rather than a connective. Audiencedirected tags are not in this bin: རེད་བ་ is addressed to listeners, and n’est-ce pas ? renders it correctly.
The method would be a corpus of transcribed teachings of known delivery conditions — interpreted consecutively, dictated to a scribe, or composed in writing — and a comparison of particle distribution across the three. A feature occurring only in the interpreted set is scaffolding; one occurring in all three is genre convention. This matters for assessment and not only for description, because bin-one and bin-four features are indistinguishable in a correspondence table and are told apart only by knowing how the source was produced. It follows that delivery conditions belong in the minimum context an assessor establishes before beginning, alongside the target language and the reader — a step this paper’s procedure did not contain and should.
The sapche question , which is better stated as a decision rule to be tested than as an open choice. Sapche are not uniform. Some assert a relation — དེ་ཡང་གཉིས་ཏེ། སྤྱིར་བཤད་པ་དང་བྱེ་ བྲག་ཏུ་བཤད་པའོ། — where the division itself carries doctrinal weight and the author’s terminology, and where a later passage may argue from the distinction it draws. Others only mark a position: a bare དང་པོ་ or གཉིས་པ་ with no descriptive content. The hypothesis worth testing is that the first is rendered as text, optionally mirrored in the typographic structure, and the second converted to numbering alone.
Two complications follow, both checkable against existing editions. Later text cross-refers to the second of those three , which needs the in-line numbering to remain legible, so conversion is not free even in the contentless case. And where an edition both translates a sapche and sets a heading from it, the two can state the same division in different words — a consistency failure that a text-to-text comparison will not catch either, for the same reason it cannot see the conversion.
Evidential warrant under transposition
ཟེར་ is testimonial. The corpus French renders the deflection of authority it carries with an inferential impersonal ( on voit que ), so the deflection survives and its warrant type does not. French has no neutral hearsay frame for running prose — dit-on and paraît-il import doubt, selon ce qui est dit is legalistic — which suggests the shift may be forced in the manner of §10.6’s cascade rather than chosen. Tibetan’s evidential system is rich and the Romance languages differ among themselves in reportative resources, so this is testable in the frame of §11.6: if the shift is forced, it should weaken in a target with a lighter reportative.
Whether the encoding findings are French or Romance
§11.6 lists three structural differences that could change the target-side results and argues none of them. Each is a small study against existing bilingual material. The corpus study §12.3 asks for also remains undone, and would now be better run on colophons than on continuous teaching: they are formulaic, so the slots are fixed and scoring is mechanical; they are numerous; and they are widely published with the Tibetan facing.
13. Conclusion
The translation examined here is good. Of twenty-six meaning units it delivers seventeen unaltered, shifts four, thins three, leaves one absent, and leaves one undecidable on the evidence available. Several of its boldest departures from source form are the reason it succeeds: a scalar adverbial for iconic repetition, a cleft for positional focus, an anaphoric demonstrative for flat repetition, a redistributed ideophone compound for a lexical thread that could not be repeated without marking what the source leaves unmarked. These are the operations of a translator working at the level of discourse rather than of the clause.
One loss touches the message: the substitution of an addressee-directed simple future for the source’s generic-predictive morphology. A larger corpus of the same genre shows that loss to be both greater and less culpable than one passage suggested. It is not a choice but the terminus of a four-step cascade beginning in French’s resistance to stacked nominalised complements, so it fires wherever that construction occurs rather than at one passage’s five clauses; and while an idiomatic French construction avoids it entirely, nothing in the language routes a translator to that construction. Four further losses are small, individually arguable, and repairable without restructuring a sentence. The first assessment of the same text, conducted in the ordinary way, returned a verdict of architectural failure. That verdict was an artefact. It arose from treating the source’s typological defaults — one finite verb per period, verb-final position, repetition as the unmarked device for iteration and intensification, concrete figuration of abstract process — as rhetorical assertions by the speaker, and then scoring the target for not reproducing them. The errors so produced were unidirectional: every one ran toward judging the target deficient for not resembling the source, and the method’s granular, checkable, locally confirmable character was what made the resulting verdict persuasive.
Three findings have application beyond this passage. First, the baseline must be subtracted before any divergence is recorded, and a divergence counts as loss only where the source choice was marked and the target language demonstrably held an unused resource for the same function. Second, preserving source structure can falsify register, because two languages may distribute the same structural resource across opposite points of the prestige scale. That argument holds and is the structural explanation for why literal translation from Tibetan reads as awkward or juvenile in a Romance language; the conclusion first drawn from it was too broad. Flat predicate juxtaposition and repetition do not transfer into French, while rhetorical question-answer chains, their answer particles and topicalised information packaging do, so the operation is better described as gathering a clause chain into one complex topic and predicating once than as hierarchical nesting. Which of these results are French and which are Romance is open, and §11.6 sets out the three differences — null subjects, nominalisation weight, topicalisation tolerance — that would decide it. Third, the bias is localised in the comparative operation rather than in descriptive capacity, which makes sealing — describing each text independently, and the target as a text in its own language before it is a target — an effective and cheap control. Cheap, but not sufficient: during the corroboration the author assembled a six-item critique of the corpus translator’s French and withdrew four of the six under correction, each because an element of the source was sought at the position its correspondent should occupy, not found, and recorded as a loss while the correspondent stood elsewhere in the sentence. The six had already accumulated into a pattern, which collapsed with them. §11.2’s account of why sourceform assessment is self-confirming therefore has a demonstration in this paper’s own working, carried out with the corrective procedure already written and not run. Two refinements follow. A procedure that depends on the assessor electing to apply it does not bind; it has to gate output. And of the findings that survived correction, every one was countable — a lexeme repeated in adjacent sentences, a predication count that did not match the source’s — while every finding resting on a judgement of whether a target construction was licensed or too heavy was wrong, which is a narrower division of competence than §11.5 proposed.
The last finding has a particular edge where assessment is machine-assisted. A model’s source-form bias is not a methodological commitment that argument can dislodge but a statistical prior over what translations from the source language look like; and because that corpus is small and dominated by the literal tradition, a competent discourse-level translation resembles the prior less closely than a poor literal one does. The failure is a calibration error rather than a reasoning error, and further reflection does not correct it. What corrects it is the procedure — and, for the foreseeable term, the supervision of the rare combination of competences that wrote the correction in the first place.
Appendix A — The Tibetan text
དེར་རྩེ་གཅིག་ཏུ་བལྟས་པ་དཔེར་ན་ཏདྱ་ཐ། ༀཾ་མུ་ནེ་མུ་ནེ་མ་ཧཱ་མུ་ནེ་ཡེ་སྭ་ཧ་ཟེར་མཁན་གྱི་སྔགས་དེ་ སྟོང་ཕྲག་གཅིག་ཙམ་བཟླ་བའི་ ཚེ་ན་སངས་རྒྱས་ཀྱི་སྐུ་ལ་དམིགས་པ་དེ་འདྲ་བྱས་ནས་ བལྟས། དེ་ནས་མཚམས་བཞག་ནས་ག་ལེར་རང་གི་སེམས་མ་བཅོས་པར་ བཞག་བཞག་ན་ སེམས་ལ་རྣམ་རྟོག་གང་ཡང་མི་འཕྲོ་བར་ཆགས་སྡང་གི་རྟོག་པ་གཏན་ནས་མེད་པ་ འཇམ་ཐིང་ཐིང་ལྷོད་དེ་བ་བྱས་ སེམས་གནས་སྡོད་མཁན་གཅིག་ཡོང་གི་རེད། དེ་གནས་བསྡད་པ་དེའི་ཐོག་ལ་གནས་སུ་བཅུག་སྟེ་སེམས་མ་ཡེངས་པ་ཙམ་དུ་སྐྱོང་ བསྡད་ན་ དེ་ནས་དེའི་རྣམ་རྟོག་སྔ་མ་འོང། ཕྱི་མ་རྒས། ཕྱི་མ་རྒས། སྔ་མ་འོངས་ཏེ་ཕར་མེད་པ་ཆགས་ནས་ཐིང་ཐིང་འགྲོ་ཡི་རེད་དུས་ དེའི་ནང་ན་སེམས་མི་རྟོག་པ་ཟེར་མཁན་འདི་རེད།
Segmentation into periods A to D, with internal boundaries marked, is given at §4.1.
Appendix B — The French text
Maintenant votre regard intérieur en un seul point sur le corps du Bouddha, récitez par exemple un millier de fois le mantra TADYATHA OM MOUNÉ MOUNÉ MAHA MOUNAYÉ SOHA. Puis vous vous arrêtez, et vous laissez votre esprit complètement détendu, sans aucune altération. Arrivera un moment où plus aucune pensée ne s’élèvera, l’esprit sera complètement paisible et détendu, sans aucune pensée d’attachement ou d’aversion. Alors, si vous posez votre esprit dans ce calme, et si vous le maintenez dans cette absence de distraction, eh bien les pensées qui arrivent les unes à la suite des autres deviendront de plus en plus faibles et l’esprit deviendra limpide. C’est là que se trouve ce qu’on appelle « l’esprit sans pensées ».
Sentence numbering S1 to S5 as at §5.1.
Appendix C — Predicate correspondence
The table records surface correspondence only. The operation column states what the target construction does in its own language; the status column gives the second-pass verdict of §8.
| Tibetan | French | Operation | Status |
|---|---|---|---|
| བལྟས་པ་…བལྟས་ | — (votre regard intérieur | nominal absolute as | re-encoded |
| (envelope, finite) | …, verbless) | circumstantial frame | |
| Tibetan | French | Operation | Status |
| དམིགས་པ་ (nominalised, manner) |
fused into the same NP (regard intérieur) |
lexical encoding of the ocular/mental distinction |
open (§12.1) |
| བཟླ་བའི་ཚེ་ན་ (subordinate frame) |
récitez(imperative, matrix) |
frame/figure swap | marked choice unmatched |
| མཚམས་བཞག་ནས་ (converb) |
vous vous arrêtez | converb → finite; présent de procédure |
re-encoded |
| བཞག་བཞག་ན་ (consecutive) |
vous laissez+Puis+ juxtaposition |
consecution via procedural sequence |
re-encoded |
| ག་ལེར་ | — | — | absent |
| མ་བཅོས་པར་ | sans aucune altération | negative converb → PP | present |
| ཡོང་གི་རེད་(generic; subject = a state) |
arrivera un moment (future; subject = a moment) |
genericity not carried | loss (§9.2) |
| མི་འཕྲོ་བར་ | plus aucune pensée ne s’élèvera |
attribute → finite assertion |
present |
| མེད་པ་ | sans aucune pensée d’attachement… |
attribute → PP | present |
| འཇམ་ཐིང་ཐིང་ལྷོད་དེ་བ་ བྱས་ |
paisible et détendu | compound split;ཐིང་ཐིང་ held back |
re-encoded (§8.4) |
| དེའི་ཐོག་ལ་+གནས་སུ་ བཅུག་སྟེ་(3 mentions) |
si vous posez … dans ce calme |
flat repetition → anaphora |
re-encoded (§8.5) |
| བཅུག་(permissive) | posez(agentive) | — | loss, small |
| ཙམ་དུ་+སྐྱོང་བསྡད་ན་ | si vous le maintenez dans cette absence… |
measure carried; limitative thinned |
thinned |
| སྔ་མ་འོང། ཕྱི་མ་རྒས།×2 | les pensées qui arrivent les unes à la suite des autres … de plus en plus faibles |
iconic repetition → scalar adverbial |
re-encoded (§8.3) |
| ཕར་མེད་པ་ཆགས་ནས་ | (implied by_limpide_) | extinction stated → implied |
thinned |
| Tibetan | French | Operation | Status |
| ཐིང་ཐིང་འགྲོ་ཡི་རེད་ | l’esprit deviendra limpide | ideophone → quality adjective; subject supplied |
re-encoded |
| དུས་དེའི་ནང་ན་+ཟེར་ མཁན་འདི་རེད་ |
c’est là que se trouve ce qu’on appelle … |
positional focus → cleft |
re-encoded (§8.6) |
Appendix D — Junction correspondence
| Joint | Tibetan | French | Second-pass status |
|---|---|---|---|
| opening | དེར་(anaphoric, outside) |
Maintenant (cataphoric) |
shifted; bears on §12.1 |
| A → B | དེ་ནས་ | Puis | matched |
| B protasis → apodosis |
ན་(consecutive) | asyndeton after_Puis_ | re-encoded (§8.2) |
| B → C | ཐོག་ལ་(basis anchor) | Alors+ce calme/cette absence |
re-encoded (§8.5) |
| C protasis → apodosis |
ན་+དེ་ནས་ | si … et si …+eh bien | matched |
| C → D | དུས་དེའི་ནང་ན་(temporal) | _là_inside the cleft | re-encoded (§8.6) |
| within A | ཚེ་ན་,བྱས་ནས་,དཔེར་ན་… ཟེར་མཁན་ |
_par exemple_only | thinned; scope shifted |
| within B | ∅×2 (asyndetic co- rank) |
où/ comma /sans | re-encoded; scope leak (§5.4) |
Note on apparatus
This is a first-order analysis of two primary texts, and its claims stand or fall on those texts rather than on citation. The theoretical vocabulary used in Section 2 — clause chaining, converb, ideophone, markedness, text linguistics, the distinction between propositional and stylistic content — is standard and is not attributed to particular pages. Readers wanting the standard treatments will find the relevant frameworks in the works listed below; the list is indicative of where the concepts are developed, not a record of sources consulted for any specific claim here, and no claim in this paper should be credited to any of them. Specific attributions, page references and engagement with the Tibetanstudies literature remain to be supplied , and the paper should be treated as a working analysis until they are.
The Tibetan is reproduced from the transcription supplied; no collation against other witnesses of the teaching was attempted (§12.4). The French is reproduced as supplied, with no editorial intervention. Neither text’s provenance, date or authorship is established here (§12.2).
Two gaps in the literature are worth recording, since they bear on §11.4 and §11.5. First, discourse-level and text-linguistic analysis of Tibetan-to-Romance translation appears to be very thinly represented; the bulk of methodological reflection in the field concerns terminology and philological transparency. Second, the metalanguage available for discussing the problem is itself largely source-oriented in its classical formulations, which may be one reason the bias is durable.
Indicative background
-
Beaugrande, R. de, and W. Dressler. Introduction to Text Linguistics. 1981. — text linguistics; the unit of fidelity argued in §2.5.
-
Beyer, S. The Classical Tibetan Language. 1992. — clause chaining, converbs, nominalisation.
-
Halliday, M. A. K., and R. Hasan. Cohesion in English. 1976. — cohesion and anaphora, relevant to §8.5.
-
Nida, E. Toward a Science of Translating. 1964. — formal against dynamic correspondence.
-
Reiss, K., and H. Vermeer. Grundlegung einer allgemeinen Translationstheorie. 1984. — skopos; register value as a translation constraint (§11.1).
-
Tournadre, N., and S. Dorje. Manual of Standard Tibetan. 2003. — spoken Tibetan morphology; evidential and epistemic auxiliaries, relevant to §9.2.
Contributions
The descriptive analyses of §§4–5, the two-pass comparison of §§6–9, the corroboration study of §10, the assessment procedure of §11.3 and the drafting of this paper were produced by Claude Opus 5 (Anthropic) in extended dialogue with the author.
The following originate with the author: the typological correction of §7, and its consequences throughout the paper; the fourth bin of §7.3, and the sapche case that required its criterion to be restated; the mechanism identified in §10.6; and the register argument of §11.1, which reaches the paper by way of a criticism made to the author by a senior translator in the field. The withdrawals recorded in §10.8 followed from the author’s corrections to the model’s analysis, as did the correction of provenance in §12.2 and the qualification entered in §12.6.
Authorship rests with the author alone. The model is not named as an author because authorship entails accountability for the work after publication — approving the final version, answering for its claims, standing behind them under dispute — which it cannot carry.

BuddhistAI Knowledge base