Repository path: workshop/translations/bargamot/R04-v1/variant-F1.md · rendered 2026-09-09
Page metadata (front matter)
| type | manifest |
|---|---|
| id | T-bargamot-R04-v1-F1 |
| status | frozen |
| created | 2026-07-28 |
| updated | 2026-07-28 |
| links | workshop/translations/bargamot/R04-v1/translation.md, workshop/experiments/E-20260728g-scale-usage/design.md, wiki/findings/results/RS-20260727b-tierD-rules.md, workshop/experiments/E-20260726d-tierD-heldout/design.md |
Variant F1 — the known-difference member of the pilot pair
What this is. T-bargamot-R04-v1 Unit B, carrying eight accuracy faults. Together with the unmodified Unit B it forms a pilot pair whose difference is known by construction and documented site by site — the material a positive-control gate needs, and the thing the Tier D run has never had: its sham pair is "no real difference" and its held-out pair is "unknown difference".
Built by build_variant.py, which is the record. The text is variant-F1.txt; the per-site measurements are variant-F1.json. Re-running the script reproduces both from translation.md alone, and fails loudly if any anchor is not unique.
Freeze order, which is a constraint and not a courtesy (E-20260728g-scale-usage/design.md §4.7, from the pre-run critic's finding E2). This variant is used by no statistic on the page that motivated it. Any future design that uses this pair as a control must be frozen after this commit.
The operator is not lead-invented, and that is what licenses the construction
Charter §5.2 forbids lead-invented operators, and D-20260725-06 forbids an experimenter from constructing the parity a negative control needs. The pre-run critic asked, correctly, whether constructing a known-different pair falls under the same prohibition. It does not, on one condition, and the condition is met here: the fault types are external to the lead.
Every site instantiates one of the four failure types of O4, whose catalogue is the project's record of accuracy failure in published translations — the Shaw Christianisation (A-shaw-spider-thread), Swann 1974 on Turney's lapses, and A-chekhov-pari on Koteliansky & Murry 1915 («за пять часов» → "five minutes", «в 12 часов дня» → "twelve o'clock midnight", «Евангелие» → "the New Testament"). A site instantiating none of the four is not admissible, and the type is named per site below.
And the limit the critic attached, adopted verbatim rather than softened: a juror passing a positive control built this way demonstrates sensitivity to this fault class, at this dose, on this passage — not a general capacity to express a 0.75-point margin across six senses. Anything that uses this pair must carry that sentence.
The eight sites
Composition constraint of E-20260726d-tierD-heldout/design.md §4 — at least two of each of the four types, so no single type carries the dose — met exactly: two of each.
| # | type | reference | variant | Δ words |
|---|---|---|---|---|
| 1 | wrong referent | "let the greasy collar slip mechanically out of his hand" | "…out of Garaska's hand" | 0 |
| 2 | wrong referent | "Bargamot, still at a loss, begins to feel…" | "Garaska, still at a loss, begins to feel…" | 0 |
| 3 | wrong word sense | "even insolence with meekness" («дерзость») | "even daring with meekness" | 0 |
| 4 | wrong word sense | "having lost his point of support" («точку опоры») | "having lost his point of view" | 0 |
| 5 | invented detail | "hauled him off towards the destination aforesaid" | "hauled him off through the mud of the unlit roadway, past the shuttered shops, towards the destination aforesaid" | +11 |
| 6 | invented detail | "some kind of slime… there clung fragments of dyed eggshell." | "some kind of yellow slime, still warm… there clung four fragments of red-dyed eggshell, none of them larger than a copeck." | +11 |
| 7 | dropped negation | "it was somehow not the thing to set them to work" | "it was somehow the thing to set them to work" | −1 |
| 8 | dropped negation | "without managing to show Bargamot the object" | "managing to show Bargamot the object" | −1 |
Site 2 is the story's last sentence and is the largest fault on the page. In Andreyev it is Bargamot — the policeman, the man with all the power in the scene — who begins to understand; moving that to Garaska destroys the reversal the story exists for. It is a wrong referent by the same mechanical rule as site 1 and it is not the same size of error, which is a fact about this fault class that a site count cannot express and that anything using this pair should say out loud.
Site 8 changes what happens. Dropping without means Garaska succeeds in showing the egg before he falls, which contradicts the following paragraph. That is exactly the Koteliansky & Murry failure mode — an error that survives because nobody checks the English against the Russian.
Characterisation on the two axes RS-20260727b §6.3 requires
| axis | F1 (this variant, 8 sites) | Tier D targeted O4 (24 sites) | Tier D sham (16 sites) |
|---|---|---|---|
| reference length | 441 words | 292–394 words | 292–394 words |
| sites changing the word count | 4 of 8 (50%) | 10 of 24 (42%) | 0 of 16 (0%, by constraint) |
| net words added | +20 | +44 over 3 items (+13/+15/+16) | 0 |
| as a fraction of the reference | +4.54% | ≈4% per item | 0% |
| sites changing no lexeme at all | 0 of 8 | 0 of 24 | 6 of 16 |
F1 is matched to the targeted operator on both axes and is therefore usable where the sham is not. RS-20260727b §3's finding was that a sham constrained to be word-count neutral cannot bound the false-alarm rate of an operator that adds a clause at a quarter of its sites. F1 adds words at half its sites, at the same per-item rate as the operator it is modelled on.
One thing was not matched and is stated rather than buried. The Tier D targeted items carry 8 sites over 292–394 words (2.0–2.7 per 100 words); F1 carries 8 sites over 441 words (1.8 per 100). The dose is slightly lighter in rate while the length effect is matched in percentage terms. Nothing here corrects for that, and any design using this pair should either accept the difference or add a ninth site.
What this pair is not
- Not a Tier D run, not a control that has been used, and not evidence about any juror. No model has seen either text.
- Not a negative control. It is the opposite: the difference is real, deliberate and documented. It says nothing about whether two independent good translations can be told apart.
- Not a claim about
T-bargamot-R04-v1's quality. The lead never judges its own translation (charter §5). Unit B is the reference here in the technical sense — the text that was not damaged — and that is not an evaluation of it.