Repository navigation
Answer keys: 5 drafting-reader errors found by the referee panel (signed VALUE, PIC as a name, truncated hex, VALUE from the PROCEDURE DIVISION, continued literals) #4393
Description
Problem
The #4377 referee panel (PR #4384) cross-checked the mainframe answer keys (tests/cobol_mainframe/answer_key/*.json) against the engine, tree-sitter-cobol-enterprise and Che4z. In 5 patterns the key is wrong and the engine is right. That puts noise into every engine-vs-key score (ground-truth ledger, #4318's extent and edge rows, the referee scorecard).
| Pattern | Items | Example | What the key should hold |
|---|---|---|---|
| Signed numeric VALUE read as null | 83 | VALUE +18 / +0 / -1 (GenApp lgacdb01 L83, CBSA ABNDPROC L40, CardDemo CBPAUP0C L82) |
the signed literal (the key-side twin of engine fix #4277) |
10 PIC X(20) VALUE ... named PIC |
10 | CardDemo COPAUS1C L58-67 | FILLER (an unnamed entry is an implicit FILLER, #4246) |
| Hex literal truncated to its prefix | 3 | VALUE X'0D25' → X (zECS ZECS001 L45, ZECS003 L49) |
X'0D25' (key-side twin of #4354) |
| Last DATA DIVISION item takes a VALUE from the PROCEDURE DIVISION | 10 | COMM-OVERDR value 'TO' (CBSA BNK1CAC L157 and 9 more) |
no VALUE (key-side twin of #4329) |
| Continued literal keeps the raw continuation line | 11 | CardDemo CBSTM03A HTML-L08 (STYL\n ...) |
the joined literal (see the engine-side issue filed alongside) |
Acceptance criteria
- The key drafting reader (
tests/tools/cobol_answer_key.py) fixes each pattern at its root, not by hand-editing keys. - Keys are re-drafted, and the fields are re-verified per the key's own verification discipline.
-
ground_truth_ledger.py updateratchets out the entries these patterns caused; no engine change is needed. - A rerun of the referee panel shows these patterns gone from the engine-vs-key disagreements.
Related: #4377, #4318, #4277, #4354, #4329, #4246.
🤖 Generated with Claude Code
Activity
fix: train -- #4329, #4391, #4392, #4393 (#4398)