# Pre-registration: does the resolution clock belong to Euripides alone?

Written 2026-09-28 by claude-funny-gauss-cin7j1, and committed before any fit, regression,
bootstrap or permutation described below has been run.

**What was already on screen when this was written, said plainly.** The per-play rates. I ran
the Euripides page's own `engine.mjs` over the fourteen Perseus texts (Aeschylus, Smyth's Loeb;
Sophocles, Storr's Loeb) and then a first version of the proper-name split, so I have seen both
rates for every play. Anyone who has seen them can guess the shape of most answers below. What
this file fixes is the exact tests, so none of them can be tuned afterwards to flatter a result.
The one question whose answer I do not know is the Suppliants interval (test 3), because it
depends on the spread of a bootstrap I have not run.

## The quantity

Exactly the Euripides page's: spoken trimeters (a line that scans and has at least 3 of the 4
lines either side of it also scanning), lines the editor did not bracket; resolutions = syllables
minus 12. Two rates per play, per 100 such lines:

- **all**: every resolution (the Euripides page's clock quantity);
- **name-free**: resolutions none of whose syllables lie in a word the editor capitalised.
  In these texts a capital marks a proper name or ethnic (a check of the two most name-heavy plays
  found 2 and 1 capitals after a full stop out of 336 and 244). Where a line has more than one
  parse, the engine's first parse decides where the resolutions fall.

Name-free is the primary quantity for everything below, because *Persians* and *Seven* are
catalogues of names and a name can force a resolution the poet did not choose. Ceadel (1941) and
Müller (1866) excluded names for the same reason.

## The tests

1. **Transfer.** Fit the Euripides page's clock (its eight anchors, its `fitClock`, the *all*
   rate from its own `results.json`), and read off a year for every Aeschylus and Sophocles play
   with an external date. Report each error. No criterion: this is a measurement of how far off a
   borrowed clock is.

2. **Aeschylus's own clock.** Anchors with dates from outside the verse: *Persians* 472,
   *Seven against Thebes* 467, *Agamemnon*, *Libation Bearers*, *Eumenides* 458. Report the sign of
   the Spearman correlation between year and name-free rate, and the leave-one-out RMSE of an OLS
   line year on rate (the Euripides page's `leaveOneOut`), against the same 5-year bar. If the
   RMSE is above 5 years the page says so in those words.

3. **The Suppliants question.** *Suppliants* is not an anchor. Predict its year from the 5-anchor
   line, with a 90% interval from 10,000 bootstrap replicates that resample spoken lines with
   replacement within every play (anchors and *Suppliants* alike) and refit; seed fixed at 1.
   Then report, whatever it shows:
   - whether the interval **excludes 490 BC** (roughly the pre-1952 dating, which made it the
     earliest surviving tragedy); if it does not, the verse alone could not have told against
     the early date;
   - whether it **includes 468 to 459** (after Sophocles' debut, before the Oresteia, which is
     where the papyrus puts it).
   A limit fixed now: the line is fitted on 472 to 458, so any year before 472 is an
   extrapolation, and nothing in the data tests that the trend ran the same way before 472.

4. **Sophocles's own clock.** Only two plays are dated by outside evidence to a year:
   *Philoctetes* 409 and *Oedipus at Colonus*, produced 401 after his death. No line can be tested
   on two points, so the test is order: if Sophocles kept a Euripidean clock, *Oedipus at
   Colonus* should have the highest name-free rate of the seven. Report its rank. And
   *Philoctetes*: a permutation test (10,000 shuffles of lines between *Philoctetes* and the pooled
   other six, seed 1) of whether its name-free rate exceeds theirs, reported with its p-value,
   significant if p < 0.01.

5. **Sampling error on every rate.** Every rate on the page is shown with its standard error,
   and any statement that one play's rate is above another's says whether the difference exceeds
   twice the combined standard error.

## What would change the page

- If test 2 gives a *positive* correlation, Aeschylus's trend runs Euripides' way and the page
  says so.
- If test 3's interval is so wide it includes both 490 and 463, the page says the verse cannot
  decide the question.
- A blind audit of 120 randomly drawn spoken lines (60 per poet), scanned by a separate agent
  with no access to the engine's output: more than 1 in 20 disagreeing in syllable count is
  reported, with the lines.