XE
Xenova/bert-base-uncased
fill-masktransformers.jsunknown110M params · updated Jul 2025
Mainframe serving
Can a COBOL transaction program call this checkpoint, and under which transport? The model returns evidence; the host keeps the decision.
Export status
Already ONNX
Published with ONNX artifacts — no export step required
Recommended pattern
Pattern 01
Co-located in-memory
Latency class
< 2 ms
ONNX artifact
Not measured
measured once the harness posts results
Screening note
Compiled graph executes in the same address space as the COBOL program. No network hop; the host keeps the decision.
Static graph, per-record fill-mask output and 110M parameters fit a co-located sub-2 ms call
Zero egress — data stays inside the encrypted boundary
- ONNX artifact present: onnx/model.onnx
osai-screening/1.0.0 · verdict 30a97a1696f1d6ed…
Serving controlskill switch offshadow onconfidence floor 0.85GET /serving-status
Promotion gates
| Gate | What it checks | State |
|---|---|---|
| Structural check | Graph loads; node inputs, outputs and shapes resolve | Not run |
| Numerical parity | Same inputs agree with the source framework within 1e-4 | Not run |
| Eight-bit quantisation | INT8 weights with parity re-run afterwards | Not run |
| Shadow cycle | Full cycle scored against live traffic with no influence on outcomes | Not run |
Interface contract
The copybook is generated, versioned and never hand-edited on either side. Entry point SCORE_BERT_BASE_UNCASED.
Generated copybookmodel record · contract
* GENERATED BY osai-copybook-gen/1.0.0 — DO NOT EDIT (STANDARD S2)
* COPYBOOK FOR Xenova/bert-base-uncased
* REVISION ab680a327acc PATTERN 01 ENTRY SCORE_BERT_BASE_UNCASED
01 SCORE-BERT-BASE-UNCASED-REQUEST.
05 REQ-RECORD-ID PIC X(12).
05 REQ-TEXT-LEN PIC S9(4) COMP-5.
05 REQ-TEXT PIC X(512).
01 SCORE-BERT-BASE-UNCASED-RESPONSE.
05 RESP-SCORE COMP-1.
05 RESP-CONFIDENCE COMP-1.
05 RESP-LABEL PIC X(16).
05 RESP-MODEL-VERSION PIC X(16).
05 RESP-STATUS PIC S9(4) COMP-5.
SCORE-BERT-BASE-UNCASED-REQUEST / OsaiRequest526 bytes
| COBOL | C | Kind | Offset | Bytes | Note |
|---|---|---|---|---|---|
| REQ-RECORD-ID | record_id | alnum(12) | 0 | 12 | caller correlation key, logged with the inference |
| REQ-TEXT-LEN | text_len | binary16 | 12 | 2 | bytes used in REQ-TEXT |
| REQ-TEXT | text | alnum(512) | 14 | 512 | de-identified text; tokenized inside the graph |
SCORE-BERT-BASE-UNCASED-RESPONSE / OsaiResponse42 bytes
| COBOL | C | Kind | Offset | Bytes | Note |
|---|---|---|---|---|---|
| RESP-SCORE | score | float32 | 0 | 4 | model output; evidence, not a decision |
| RESP-CONFIDENCE | confidence | float32 | 4 | 4 | compared against the confidence floor by the host |
| RESP-LABEL | label | alnum(16) | 8 | 16 | e.g. REVIEW or APPROVED — never DENY |
| RESP-MODEL-VERSION | model_version | alnum(16) | 24 | 16 | revision that produced the score (standard S4) |
| RESP-STATUS | status | binary16 | 40 | 2 | 0 scored · 4 below floor · 8 transport or runtime failure — 4 and 8 fall through to host rules |
CLAIMPROC.cob — the callergenerated by the screening service
IDENTIFICATION DIVISION.
PROGRAM-ID. OSAICALL.
DATA DIVISION.
WORKING-STORAGE SECTION.
COPY OSAICPY.
77 CONFIDENCE-FLOOR COMP-1 VALUE 0.85.
PROCEDURE DIVISION.
CALL "SCORE_BERT_BASE_UNCASED" USING SCORE-BERT-BASE-UNCASED-REQUEST, SCORE-BERT-BASE-UNCASED-RESPONSE
ON EXCEPTION MOVE 8 TO RESP-STATUS
END-CALL
EVALUATE TRUE
WHEN RESP-STATUS NOT = 0
PERFORM APPLY-EXISTING-HOST-RULES
WHEN RESP-CONFIDENCE < CONFIDENCE-FLOOR
PERFORM APPLY-EXISTING-HOST-RULES
WHEN RESP-LABEL = 'REVIEW'
PERFORM HOLD-FOR-REVIEW
WHEN OTHER
PERFORM CONTINUE-ADJUDICATION
END-EVALUATE
PERFORM WRITE-AUDIT-RECORD
GOBACK.
The C-ABI wrapper is on the Usage tab under “C bridge”. Test the bridge on a workstation with GnuCOBOL before it reaches an LPAR — struct packing and numeric conversion are what actually break, and they are identical on both targets.
Standards this deployment inherits
Download ML-BOMS1Models ship as ONNX artifacts. No ML framework runtime is installed on z/OS.
S2The copybook is the interface contract — generated, versioned, never hand-edited on either side.
S3Latency class selects the pattern. Team preference and vendor enthusiasm do not.
S4Every inference is logged with the record state and model version that produced it.
S5PHI leaves the platform only tokenized, over mTLS, to an endpoint under an executed BAA.
S6A model influences payment only after a full shadow cycle and with a tested fallback path.