Repository path: workshop/experiments/E-20260806-published-loss/materials/statements-check.md · rendered 2026-09-09
Page metadata (front matter)
| type | note |
|---|---|
| id | E-20260806-statements-check |
| status | frozen |
| created | 2026-08-06 |
| updated | 2026-08-06 |
| links | workshop/experiments/E-20260806-published-loss/design.md |
The yardstick statements — leak screen and role-direction check
Both checks were made after the yardstick returned and before stage 2, i.e. before POSITIVE
and WRONG were authored and before any grading call. Design amendments A5 (role check, from the
critic's note N1) and §6 (leak screen, in the frozen design).
Seat and provenance. x-ai/grok-4.5, provider xAI, two calls ($0.01458840 + $0.00611040),
shown the German span, the two names, and nothing else — no English of any kind, no gloss, no
description of the study.
1. Leak screen
The screen is a mechanical substring test over the words listed in design.md §6. First pass: four
of twelve statements flagged — A2, A7, N1 on formal, A8 on du. All four were
regenerated once with the constraint restated, as the frozen design requires.
After regeneration, one statement still flags: N1. Adjudicated and recorded rather than
suppressed:
N1— "…addresses Luther with formal deference due to a figure of high public and moral standing…". Two hits, and neither is a leak.duis a substring false positive: it matches inside the English word due. The screen is a substring test and this is its known cost.- formal describes the speaker's manner, not a German category. And
N1is a NULL control: the span carries no second-person address form at all, so there is no device at that site for a statement to name.
F3 requires more than three of twelve to fail the screen twice. One does, and it is adjudicated
above. F3 does not fire.
2. Role-direction check (amendment A5)
Each statement read against its site's speaker and addressee. A statement that assigns the roles the
wrong way round would poison every arm at that site, including the POSITIVE and WRONG derived
from it.
| site | speaker → addressee | statement assigns | verdict |
|---|---|---|---|
| A1 | Luther → Kohlhaas | "a spiritual superior correcting a man he regards as erring and subordinate" | correct |
| A2 | Kohlhaas → Luther | "Kohlhaas addresses Luther as a revered authority of markedly higher standing" | correct |
| A4 | Luther → Kohlhaas | "a moral superior demanding an account from a man beneath him" | correct |
| A5 | Luther → Kohlhaas | "a spiritual mentor speaking to a wayward but still approachable man" | correct |
| A6 | Luther → Kohlhaas | "a superior who may command a moral decision" | correct |
| A7 | Kohlhaas → Herse | "Kohlhaas addresses Herse as his own subordinate servant" | correct |
| A8 | Kohlhaas → Herse | "Kohlhaas continues to treat Herse as a household dependent under his authority" | correct |
| A9 | Herse → Kohlhaas | "explains himself to his master with the respectful deference owed by a servant" | correct |
| A10 | Kohlhaas → Herse | "a master who feels entitled to extract a confession" | correct |
| N1 | Kohlhaas → Luther | "addresses Luther with formal deference due to a figure of high public and moral standing" | correct |
| N2 | Herse → Kohlhaas | "reports events to his master with the earnest deference of a subordinate" | correct |
| N3 | Herse → Kohlhaas | "continues his report to his master … employee to employer" | correct |
Twelve of twelve correct. No site is removed.
This is a reading, not a machine test, and it was made by the lead — the same party that authors
POSITIVE and WRONG. The blinding in this design protects the graders; it does not protect this
check, and the critic's note N9 applies to it as much as to stage 2.
3. One thing the statements show, recorded before grading
The seat was shown one decontextualised turn of German and the two names. It nonetheless separated
the two du dyads: Luther's is "the blunt personal authority of a spiritual superior" and
"scold him intimately and without ceremony"; Kohlhaas's to Herse is "the easy familiarity and
corrective rights of a master speaking to a dependent". That is the distinction T-kohlhaas-R20-v1's
log declares its English cannot make — its device is the word man in both cases. Whether any of
the four Englishes conveys the distinction is what the grading measures; that the yardstick draws
it is a property of the German, and of a reader with access to the German.