Writing · ELLIPSE
ρ .846
Score correlation
N=100. Evidence for writing on the Standard scale. Correlation is not a percentage of exact agreement.
Validation / The scale trust ledger
Evidence, with
its boundaries.
Some results have empirical support. Some scales inherit a validated score. Others are anchored to published boundaries. Those are different kinds of evidence.
Ledger snapshot · 2026-07-25 · Repository evidence, not a new independent audit
Essay-level CEFR · ICNALE
94%
Within one band
37% exact. The band estimate has a margin of uncertainty; it is not an official exam result.
The ladder, as recorded
What each status means.
“Trusted” is the repository’s status label. “Anchored” means a documented conversion or boundary, not a validation against independently scored work on that exam.
| Scale | Status | Basis |
|---|---|---|
| Standard 50 / 100, Percentage | TRUSTED | ELLIPSE ρ .846 (N=100); Percentage = Standard-100 renamed (pinned) |
| US Letter, P/M/D, Standards-Based 1-4 | TRUSTED (convention) | boundary labels on the validated /100; conventions documented; schools can override via custom_boundaries |
| CEFR | VALIDATED ±1 band | evidence-derived thresholds (FCE/ICNALE/ELLIPSE anchors); ICNALE: 94% within-1-band, 37% exact; FCE B2 exam → 67% B2 |
| TOEFL Writing /30, Independent /5 | ETS-ANCHORED | ETS's own rubric-mean table; chain-check: our B2 zone → ETS B2 range (17-23); 87% of FCE strong passes inside it. No TOEFL-scored gold set yet |
| GCSE /40 | BOUNDARY-ANCHORED | AQA 8700 June 2025 official boundaries (as % of max). Caveat: L1-English exam graded by an ESL-taxonomy engine |
| IGCSE /25 | BOUNDARY-ANCHORED | Cambridge 0500 June 2025 thresholds; Cambridge CEFR letter (B≈C1, E≈B2) as cross-check |
| IB MYP /32 | BOUNDARY-ANCHORED | official MYP 1-7 descriptor boundaries |
| General Speaking /10 | TRUSTED (remap) | pure RALE /50 ÷ 5 |
| IELTS family | Davy's lane | untouched tonight |
The table is reproduced from the ledger with its status wording intact. “Davy’s lane” identifies the owner of IELTS work; “untouched tonight” refers to that dated campaign.
Source record
docs/reference/SCALE_TRUST.md · “The ladder now”modules/common/scale_crosswalk.pyKnown caveats
Say these out loud.
Never hide them.
CEFR is ±1 band at essay level.
The corpora themselves overlap that much. C1/C2 thresholds are provisional, pending the 500-essay top-end run after recalibration.
Absolute grades can be harsh in a classroom.
ELLIPSE’s best essays (about B2) map to US letter “D+”. That is correct on the absolute scale. Schools wanting classroom-relative letters can use per-school custom_boundaries.
TOEFL conversions diverge at the extremes.
The rubric-mean path and ETS’s IELTS↔TOEFL population concordance disagree at the low and high ends. The mid-range (B2) agrees.
The taxonomy count was corrected.
The public impact field skills_per_scan was corrected from 100 to 50. A list of skills is not evidence that every skill is measured in every submission.
Open questions
What we have not
validated yet.
TOEFL-scored gold data
The ledger explicitly says there is no TOEFL-scored gold set yet. Agreement between conversion tables does not replace assessment against that data.
The top end of CEFR
C1/C2 thresholds remain provisional in this ledger. The planned top-end run is not a completed result.
IELTS-specific validation
The IELTS family was outside this campaign and remains Davy’s lane. This ledger supplies no new IELTS examiner-agreement result.
Transfer across modalities
The writing correlation does not validate speaking or receptive skills. Standard Reading has objective marking tests; that establishes marking behavior, not a broad proficiency claim.
Exam and task fit
The ledger leaves questions open about IB MYP task-appropriate criteria, GCSE top-end mechanics calibration, IGCSE criterion structure and BTEC criterion codes.
Current-form TOEFL Speaking
The ledger describes a historical speaking conversion. It explicitly says this must not be presented as a current-form exam claim.
Read the receipts correctly
A working pipeline is
not an accuracy study.
The ledger also records live report generation using one clean essay across several scales. That demonstrates routing, report generation and display conversions. It does not expand the validation sample or prove examiner agreement.
Likewise, automated regression checks protect calculations and known cases. They do not establish a universal accuracy rate. The “educational-board review” in the source used simulated boards; it is not external accreditation.
Source record
docs/reference/SCALE_TRUST.md · live proof, test coverage and deferred findings