The figures we quote come from one claim ledger, in its wording. Each is listed with the condition under which it holds and where it was measured.
Pre-registered. The hypotheses were fixed before the held-out books were scored: seven of ten confirmed, all ten reported.
+2.1
+2.1 Gini points over the strongest open model of its class at 100 matured loans, across 208,443 held-out loans on 13 lending books from 12 lenders (+1.9 on test partitions alone); +1.3 across 100–500 loans, confirmed in a pre-registered held-out test (95% interval +0.5 to +2.2; 9 of 12 lenders).
Comparator: TabICLv2 with eight members. Ranking quality, not calibration; outcome-stratified contexts. Gini = 2·AUC − 1.
+0.8
Specialisation never costs you at full data: +0.8 Gini over the same open model at full context, where a fixed specialisation lost 0.7.
At full data the gain comes from the serving procedure, not from the specialised weights.
+1.7
On 11 public credit tables: +1.7 Gini over the strongest open model of its class at full context, and ahead of TabPFN-3 at every context size (+0.6 at 100–500 rows, +2.0 at full context).
TabPFN-3 at library defaults; TabPFN-3.5 untested. On private books the difference with TabPFN-3 is unresolved.
0
Pretrained on synthetic credit economies; no borrower record enters pretraining.
Own-brand instances add continued pretraining on the bank's data, inside the bank's environment. Synthetic-only pretraining and bank-local serving are separate properties.
~50 ms
Scores an application in about 50 ms (54 ms median, 57 ms p95) on a 16-thread server CPU and about 20 ms on a GPU, matching the research model's predictions.
After a one-time portfolio encoding (about 90 s on CPU, 5 s on GPU, can run offline); 12,000-row context; one server.
4–8
Runs on commodity CPU servers: 4–8 cores and about 7 GB RAM per worker.
Memory scales with portfolio size and member count. This is production sizing, not the 16-thread host the latency figure was measured on.
0.98–1.03×
Credit specialisation adds no serving cost.
Measured at equal members, precision and engine settings: 0.98–1.03× of the base.
0.651 → 0.699
In replay on 10 private books, automatic Handover lifted served AUC from 0.651 to 0.699, recovering 70% of the hindsight gap (63% on the 7 books unseen in training).
A retrospective replay between Day-Zero and zypl.Foundation — not a live deployment, and not yet with Zehn. Always serving zypl.Foundation reached about 0.711: Handover is a governed, reversible policy, not an accuracy gain over the foundation model. Switches occurred at application index 2,862 to 11,182.
≈ 39
Zehn matches default CatBoost, LightGBM and EBM on 120 datasets with about 39 effective parameters.
Default baselines; public datasets; under review.
Calibrated probabilities out of the box. Ranking is validated; probabilities are recalibrated per portfolio before use as PDs.
Reason codes for zypl.Foundation scores. Not shipped. Zehn's reason codes describe Zehn itself.
Higher approvals or lower losses. Only from a lender's own matured pilot evidence, with its permission.