Repository path: workshop/translations/petits-poemes/R07-v1/translation.md · rendered 2026-09-09

T-petits-poemes-R07-v1 — Baudelaire, four prose poems, under R07 (fluency)

• Source. Charles Baudelaire, Petits Poèmes en prose, in Œuvres complètes IV (Paris: Michel Lévy Frères, 1869), via French Wikisource (Textes validés). Four complete poems: III «Le Confiteor de l'artiste», XVII «Un hémisphère dans une chevelure», XXXV «Les Fenêtres», XLI «Le Port». 843 French words. Public domain. Text at ../source-fr.txt.
• Regime. R07 v1.0 (fluency), frozen 2026-07-28 at S048. Single pass, no revision, source only.
• Why this material. ARM-rule-coverage, step 1. Figure-dense, in a source whose world an English reader already holds — chosen to separate the rule set decides from the source is far away. The criterion and three predictions are frozen in ../prereg.md, committed at e983b84 before the source was read for translation.
• Dialect declared under F5. American English, held throughout — the third R07 artifact and the third to declare American, which is what keeps the three comparable to each other.
• Translated by the lead agent, in session. Cost $0; not ledgered (charter §3, A4).
• No published rendering of any of these four poems into any language was read before this text and its log were frozen. The comparator (below) was fetched to disk in the same session; only the counts of six heading strings were printed from it, and no line of its prose was displayed — note (bdn), the safe check is the one that prints no prose.
• Contamination suspected, MEASURED (appended below in a separate commit). Per-unit longest run 16 tokens; the whole-work 21 is a splice across two poems and is not a run. One-line basis: Baudelaire's prose poems are among the most-translated short prose in French and at least three public-domain English editions are freely reachable on Project Gutenberg, so a prior of suspected is the honest one to commit; against that, none of these four is among the two or three pieces (Enivrez-vous, L'Étranger) that circulate as quotations. The measurement cannot precede the translation — it needs this text.
• What contamination bears on here, and what it does not. This artifact's use in ARM-rule-coverage is to supply decision sites and coverage codes, which are claims about what the rule set does, not claims about how good the English is. A remembered phrase changes whether the rendering is the lead's; it does not change whether F4 names the feature at issue. Stated so that the declaration is not read as doing more work than it does.
• Evaluation. The lead never judges its own translation (charter §5). The study limb it is wired to asks non-lead readers to assign the coverage codes — a question about the rule set, on the factual side of the S015 line — and does not score the prose.

Translation

III. The Artist's Confiteor

How piercing the ends of autumn days are! Ah, piercing to the point of pain! For there are certain delicious sensations whose vagueness does not exclude their intensity, and there is no point sharper than the point of the Infinite.

What a delight it is to drown your gaze in the immensity of sky and sea! Solitude, silence, the incomparable chastity of the blue! A little sail shivering on the horizon, which by its smallness and its isolation imitates my own irremediable existence; the monotonous melody of the swell — all these things think through me, or I think through them (for in the grandeur of reverie the self is soon lost!). They think, I say, but musically and picturesquely, without quibbles, without syllogisms, without deductions.

All the same, these thoughts, whether they come out of me or dart up out of things, soon grow too intense. Energy inside pleasure creates a discomfort and a real suffering. My nerves are strung too tight, and they give out nothing now but shrill, painful vibrations.

And now the depth of the sky dismays me; its clearness exasperates me. The insensibility of the sea, the unchangingness of the spectacle, revolt me. Ah! Must a man suffer forever, or forever run from the beautiful? Nature, pitiless enchantress, rival who always wins, let me alone! Stop tempting my desires and my pride! The study of the beautiful is a duel in which the artist cries out with fear before he is beaten.

XVII. A Hemisphere in a Head of Hair

Let me breathe the smell of your hair a long time, a long time; let me bury my whole face in it, like a thirsty man in the water of a spring, and shake it out with my hand like a scented handkerchief, to shake memories loose into the air.

If you only knew everything I see! everything I feel! everything I hear in your hair! My soul travels on scent the way other men's souls travel on music.

Your hair holds a whole dream in it, full of sails and masts; it holds great seas whose monsoons carry me toward charming climates, where space is bluer and deeper, where the air is perfumed by fruit, by leaves, and by human skin.

In the ocean of your hair I half make out a harbor swarming with melancholy songs, with strong men of every nation and ships of every shape, cutting their fine, complicated architectures against an immense sky where the eternal heat lies sprawling.

In the caresses of your hair I find again the languors of long hours spent on a divan, in the cabin of a beautiful ship, rocked by the imperceptible roll of the harbor, among the flowerpots and the cooling water jars.

In the burning hearth of your hair I breathe the smell of tobacco mixed with opium and sugar; in the night of your hair I see the infinity of the tropical blue blazing; on the downy shores of your hair I get drunk on the mingled smells of tar, musk, and coconut oil.

Let me bite your heavy black braids a long time. When I nibble at your springy, rebellious hair, it seems to me that I am eating memories.

XXXV. Windows

A man looking in from outside through an open window never sees as much as a man looking at a closed window. There is no object deeper, more mysterious, more fertile, more shadowy, more dazzling than a window lit by a candle. What you can see in the sunlight is always less interesting than what goes on behind a pane of glass. In that hole, black or bright, life lives, life dreams, life suffers.

Beyond waves of roofs I make out a woman getting on in years, wrinkled already, poor, always bent over something, who never goes out. Out of her face, out of her clothes, out of the way she moves, out of almost nothing at all, I have made this woman's story over again — her legend, rather — and sometimes I tell it to myself in tears.

If it had been a poor old man, I could have made his over just as easily.

And I go to bed proud of having lived and suffered in someone other than myself.

Perhaps you will say to me: "Are you sure that legend is the true one?" What does it matter what the reality outside me may be, if it has helped me to live, to feel that I am, and what I am?

XLI. The Harbor

A harbor is a charming place to stay for a soul tired out by life's struggles. The breadth of the sky, the moving architecture of the clouds, the changing colors of the sea, the flashing of the lighthouses, make a prism wonderfully suited to amusing the eyes without ever tiring them. The slender shapes of the ships, with their complicated rigging, to which the swell gives a harmonious rocking, keep alive in the soul a taste for rhythm and beauty. And then, above all, for a man who has no curiosity and no ambition left there is a kind of mysterious, aristocratic pleasure in watching, as he lies in the belvedere or leans on the jetty, all these movements of those who are leaving and those who are coming back, of those who still have the strength to want, the desire to travel or to grow rich.

894 English words (source 843).

Translator's log

Written as the draft was written and frozen with it. Decisions and alternatives, not quality claims. A site is logged when two or more renderings were live at the moment of writing; sites where nothing was live are not decisions and are not logged, so the totals are over contested sites only.

Coverage codes per R07 §5: D — a rule names the feature and only one live option satisfies it. P — rules bear but two or more live options satisfy them all. S — no rule bears.

IR1, carried over and then found not to reach

IR1 (from T-postmaster-R07-v1) — F10 governs the translator's latitude, not the reproduction of source figures. Carried here unchanged, applied at every figurative site, and it decided nothing that was in doubt: on this material the question may I keep the source's metaphor arises constantly and IR1 answers it the same way every time.

What is new, and it is site 24: IR1 was ruled about F10 and does not reach a figure that a DIFFERENT rule flattens. «Dans ce trou noir ou lumineux vit la vie, rêve la vie, souffre la vie» is a figure of word order — three inversions with the subject held to the end each time. F6 (idiomatic syntax before close syntax) forbids the inversion in English outright. So the rule set protects a metaphor the translator could have deleted, and requires the deletion of a syntactic figure the translator would have kept. IR1 says F10 does not license flattening what the source says; it says nothing about F6 doing so, and F6 is not a rule about latitude — it is a rule about English. Recorded, not resolved: resolving it would be adding a rule, which R07 §2 forbids.

Missing rules, recorded per R07 §2

• M1 — nothing bears on typography or layout. Recurs from the Bengali run, on entirely different material. Here: the French spaced ! ; ?, the guillemets at site 28, and the suspension points of «me révoltent…». The rule set is silent on all of it. Two runs, two languages, two scripts, same gap.
• M2 — nothing bears on tense. Recurs. «vit la vie, rêve la vie, souffre la vie» is a gnomic present inside a text that moves to a compound past at «j'ai refait»; F7 is about clause completeness, not sequence of tenses.
• M3 — NEW. Nothing bears on third-language matter. F4 is worded for source-language words and source-language proper-noun conventions. «Confiteor» is Latin in a French text, and F4's clauses do not reach it while F4's heading ("No foreign matter") plainly does. The heading and the clauses come apart, and the rule set gives no way to choose. Site 1.
• M4 — NEW. Nothing bears on lexical consistency across a text. Four rules constrain each choice locally; none says that a word chosen at one site binds the next. «l'azur» (sites 6, 21), «port» (sites 19, 32) and «charmant» (sites 17, 31) each occur twice, and holding them steady was the translator's decision, not the rule set's. R05's binding register exists precisely because this is not automatic.

Decisions

What the log counts

36 logged decisions. Coverage: D 13 (36.1%), P 21 (58.3%), S 2 (5.6%).

F4 is first-and-deciding at 2 of 36 sites (5.6%) — site 22 «gargoulettes» and site 28 the tu/vous collapse. On the Bengali run the same rule was first-and-deciding at 12 of 30 (40.0%).

Four missing rules are recorded: M1 (typography) and M2 (tense) recur from the Bengali run on unrelated material; M3 (third-language matter) and M4 (lexical consistency across a text) are new.

The registered predictions, scored against these codes and by the agent that wrote both: P1 holds (36.1% < 40.5%, and nearer the Italian 14.3% than the Bengali 66.7%); P2 holds (5.6% < 20%); P3 fails as stated — IR1 bore on sites 1, 4, 5, 20 and 24, which is 5 and clears the threshold, but at four of the five it permitted both options and decided nothing, so counting the sites it "bears on" was the wrong test and the prediction was cheap. The prediction that should have been registered is whether IR1 ever decides anything, and the answer in this text is that it does not.

No sentence in this log is an evaluative claim about the translation. Where the log says a thing is "lost" it means the source states it and the English does not — a describable relation between two texts, not a judgment of the rendering.

Contamination — measured, appended after the freeze

Declaration stands at suspected, and it is now a measurement rather than a prior. This is the project's highest per-unit overlap on a lead translation to date, and the reading below is that most of it is forced by the language pair.

Comparator. Baudelaire: His Prose and Poetry, ed. T. R. Smith (New York: Boni & Liveright, 1919), Project Gutenberg #47032, public domain. Fetched to the session scratchpad, never stored in this repository (charter §7.5) and never displayed: the only things printed from the file before the measurement were the character counts of six heading strings and the word counts of the four sliced bodies (232 / 280 / 207 / 136, against French 230 / 279 / 200 / 134 — the agreement is what establishes the slices are the right poems and are clean). tools/dependence_check.py read the text; the translator did not.

The whole-work figure of 21 is spliced across two different poems and is not a run. Its text is "if it has helped me to live to feel that I am and what I am | a harbor is a charming" — the last words of XXXV followed by the first words of XLI. This is the paragraph-crossing inflation S047 measured at 3 tokens on one span, appearing here at 5 tokens across a boundary between separate works, which is the largest instance the project has recorded and the most clearly spurious. The honest figure for this artifact is the per-unit maximum, 16. Recorded so that no later page quotes the 21.

The three long runs, and why the reading is forced rather than remembered:

• III — "gaze in the immensity of sky and sea solitude silence the incomparable chastity of the" ← «noyer son regard dans l'immensité du ciel et de la mer ! Solitude, silence, incomparable chasteté de l'azur !»
• XVII — "of tobacco mixed with opium and sugar in the night of your hair i see" ← «l'odeur du tabac mêlé à l'opium et au sucre ; dans la nuit de ta chevelure, je vois»
• XXXV — "if it has helped me to live to feel that i am and what i am" ← «si elle m'a aidé à vivre, à sentir que je suis et ce que je suis ?»

All three are lists of cognate nouns and closed-class words in a language pair where the cognates are the ordinary renderings: immensité/immensity, solitude, silence, incomparable, chasteté/chastity, tabac/tobacco, opium, sucre/sugar. Of the three runs, exactly one contains a word from any of the 36 logged decisions — "gaze" (site 5), and it is the run's second token; the run then ends at site 6, where a choice was live. The runs occupy the stretches where the log records nothing, because nothing was contested there.

Two things this does NOT establish, and they are the same two the project has been careful about since S045. (i) Forced is an inference from cognate density here, not a measurement: the only instrument that would settle it is the three-model probe ARM-forced-defence owes, and it was not run this session. (ii) The measured figure is a convergence statistic and says nothing about avoidance.

What it does add, and it belongs to ARM-forced-defence rather than here. Every forced-run measurement this project holds is on Russian (21, 13, 11), Old English (20), Polish (13, 14), Bengali (11) — pairs with low cognate density. French is the language in which a forced run should be longest, and on the first French pair measured it is: 15, 15, 16 on three of four units, from a translator who had not opened the comparator. That is a prediction of the forcing account coming out right on a language chosen for a different reason, and it is filed as material for that arm's step 2.

What contamination bears on for ARM-rule-coverage: very little, and the reason is structural. The arm uses this artifact for its coverage codes — claims about which of ten rules names a feature. The three shared runs contain one logged decision between them. A remembered phrase would have to sit at a contested site to move a code, and the measurement says the runs sit exactly where the contest was absent.

Follow-up, S064 — what the study limb did to this log's own numbers

RS-20260730e-rule-coverage. Appended after the freeze; nothing above is rewritten.

1. The 36.1% D-rate is an upper bound, and so are the registered predictions it satisfied. R07 §5 defines D as a rule names the feature and only ONE live option satisfies it. Two of the three sites the study limb checked — both from other R07 texts — turned out to be P on that test while coded D, because F4 excluded one option and left two standing. This log's D codes were assigned by the same translator applying the same looser test, and none of the 13 has been re-derived. P1 and P2 in ../prereg.md hold against the numbers as recorded and the numbers as recorded are of unknown correctness.

2. IR1, which this text inherited, is unsupported. Sites 4 and 20 were among the four put to three non-lead readers. 8 of 12 cells: F10 REQUIRES the plain rendering. 4 of 12: F10's text does not settle it. 0 of 12: F10 favours keeping the figure. The registered refutation criterion did not fire, so IR1 is unsupported rather than refuted. The translation stands as executed under a declared ruling.

3. What this log contributed that survives both: M1 (typography) and M2 (tense) recur here from the Bengali run on wholly unrelated material, which makes them properties of Venuti's list; M3 and M4 are new; and site 24 shows F6 flattening a syntactic figure that IR1 protects from F10 — the rule set pulling in two directions on one sentence.