Repository path: workshop/translations/petits-poemes/prereg.md · rendered 2026-09-09

Pre-registration — the R07 coverage rate on Baudelaire

Frozen before the source was read for translation. The commit that adds this file contains no translation. workshop/translations/petits-poemes/source-fr.txt was fetched, its length measured, and its four titles chosen on one stated criterion; no sentence of it has been rendered, and no decision site has been identified, at the moment this is written.

Why register anything

ARM-rule-coverage exists because R07's headline statistic differs 4.7-fold between its only two runs, and because the same agent writes the rules, makes the choices and assigns the codes. A third run by that same agent, with a hypothesis in mind, is worth very little unless the hypothesis is written down first and unless somebody else assigns the codes afterwards. This file does the first. The study limb does the second, and the study limb is the test; this is the thing being tested.

The hypothesis

The R07 coverage rate is a property of the distance between the source's world and the target reader's, not of the rule set. Mechanism: F4 (no foreign matter) is the only one of the ten rules that is nearly self-applying, because whether a word is source-language matter is a fact about the word rather than a judgment about English. On a source full of culture-bound items F4 fires constantly and produces D codes; on a source whose world an English reader already holds it has nothing to bite on, and what remains are rules that permit rather than decide.

Materials, and why these

Four complete prose poems from Petits Poèmes en prose (1869): III Le Confiteor de l'artiste, XVII Un hémisphère dans une chevelure, XXXV Les Fenêtres, XLI Le Port. 843 French words. Chosen on figurative density in a source whose world is already the English reader's — Paris, the sea, ships, a window, hair. That combination is what breaks the confound: on this material F10 should fire everywhere and F4 almost nowhere, which is the opposite of the Bengali run.

The project's twelfth source language is not being added here; French is T-mare-au-diable-*'s language. Breadth is not what this material is for.

Registered predictions

Reference values, from the two existing R07 artifacts: T-osso-di-morto-R07-v1 D 3/21 = 14.3%; T-postmaster-R07-v1 D 20/30 = 66.7%. Midpoint 40.5%.

P1 is the one that matters and P1 is the one I can steer, because I assign the codes and I have just written the prediction down. Two things are done about that and neither is sufficient. (i) The codes go to non-lead readers afterwards, and their rate is the arm's evidence; mine is the claim under test. (ii) The log records the live options at every site, so a later reader can re-code from the log without trusting me.

A registered prediction that comes out right in a design where the predictor supplies the measurement is not a result, and this session will not report it as one.

What is not registered

The study limb's predictions. They need the site list, which does not exist yet, and they are frozen in E-20260730e's own design document before any call is dispatched.

Every judgment in this file is internal-judgment-only: no anchor supports the claim that F4 is "nearly self-applying", which is the mechanism the hypothesis rests on.

Erratum, appended S064 after the item builder parsed the logs it had read

The Italian reference value above is wrong, and it was wrong when it was frozen. T-osso-di-morto-R07-v1 has 44 logged decisions, not 21: its D-rate is 6/44 = 13.6%, not 3/21 = 14.3%. The error came from reading a truncated view of that log's table and taking the first 21 rows for the whole of it. The page's own tally section states 44 / 6 / 30 / 8 correctly; it was not read.

Consequences, stated rather than tidied away:

• The midpoint that P1 was registered against moves from 40.5% to 40.15%. The Baudelaire D-rate is 36.1%, so P1 still holds, and it holds by 4.05 points instead of 4.4.
• The 4.7-fold spread quoted in ARM-rule-coverage is 4.9-fold.
• T-postmaster-R07-v1's own tally line was also wrong and is corrected in place: P 7 / S 3, not P 8 / S 2. Its D count and its 66.7% are unaffected.

The threshold was not moved after seeing the result — it is a midpoint between two published figures and it is recomputed here from the corrected one, which moves it against the prediction, not toward it. Three of this project's five R07 tally lines have now been wrong on first writing, and every one was caught by parsing rather than by rereading. The corrections are all in the P/S split; no D count has ever been wrong.