Repository path: workshop/translations/hafez-daasht/dependence-note.md · rendered 2026-09-09
Page metadata (front matter)
| type | note |
|---|---|
| id | dependence-hafez-daasht |
| status | frozen |
| created | 2026-09-01 |
| updated | 2026-09-01 |
| links | workshop/translations/hafez-daasht/R58-v1/translation.md, tools/dependence_check.py, wiki/method-notes.md |
Contamination measurement — T-hafez-daasht-R58-v1
python3 tools/dependence_check.py runs/RS-20260901-inversion-habit/dep_cells.json, raw output at
runs/RS-20260901-inversion-habit/dep.json. The lead's 16 rendered lines of غزل ۷۷ against the
whole of each comparator, not against a matched excerpt — the conservative direction, since a
whole-book comparator gives every run in the book its chance.
| comparator | shared 7-grams | 12-grams | longest common run | verdict |
|---|---|---|---|---|
| Payne 1901, vol. 1 entire (Brockhaus I–CC, 199 odes, and it contains this poem as ode 69) | 0 | 0 | 3 tokens | clean |
| Leaf 1898 entire (28 odes) | 0 | 0 | 3 tokens | clean |
lead, T-hafez-sahar-bolbol-R55-v1 |
0 | 0 | 3 | clean |
lead, T-hafez-darad-R56-v1 |
0 | 0 | 4 | clean |
lead, T-hafez-amad-R57-v1 |
0 | 0 | 2 | clean |
lead, T-hafez-bekonad-R57-v1 |
0 | 0 | 3 | clean |
lead, T-hafez-ghazals-radif-R55-v1 |
0 | 0 | 2 | clean |
Both figures are reported, run length and 7-gram count, because run length alone has failed as a
proxy three times (CLAUDE.md; note (bcd); RS-20260731g §4.2). Here they agree: zero shared
7-grams against every comparator, which is the stronger signal and the one S028 showed is normally
non-zero even between independent pairs (75 of 78 independent pairs share at least one).
The Payne row is the load-bearing one and it is stronger than the usual case, because Payne's volume contains this ghazal. His ode 69 was not opened until the rendering and its frozen log were committed, and the measurement is consistent with that: three tokens is the shortest run this project has recorded against a comparator that contains the same poem.
Note (bhb) does not fire. The lead's five earlier Hafez renderings produced this project's 37-token self-match record; against this poem the longest self-run is 4 tokens, and the poems are different poems, which is the whole of the explanation.
The tool's non-ASCII warning fired at 2.4% and is expected: the five prior-rendering comparator files carry the Persian source lines alongside the English, and ASCII tokenisation deletes them. The lead's own file is pure ASCII, and deleting Persian from a comparator can only reduce the run lengths reported against it — the wrong direction to worry about, since every run reported is already at or below 4.
What this licenses and what it does not. It licenses the statement that the rendering was made
from the Persian. It does not make the lead an independent third translator of anything, and no
figure in RS-20260901-inversion-habit depends on the lead's independence from Payne: the study
limb's counts are coded off Payne's and Leaf's own printed pages, and the log's role is to supply a
claim to be tested, not a measurement to be pooled.