Repository navigation
Run two or three fresh public batch estates through the trial protocol #3806
Description
Activity
- added a commit that references this issue
on Sep 28, 2026 Rescan of the opensourcecobol4j estate on origin/main
6991970d(pre-trial, no porting)This rescan follows the fixes for #3955-#3958 (PRs #3961, #3965, #3966, #3967). The pipeline is unchanged: refract with
--scan(declared encoding per suite), generate Java,mvn -o compile(JDK 17). Each run is checked against an independent Unicode-aware reading of every program.before (2026-09-28) now Japanese PROGRAM-IDs not refracted 12 0 data names missing 51 0 Japanese data names truncated 59 25 Japanese sections / paragraphs missing 36 28 ASCII names truncated ( _in names, jp-compat)7 0 generated project compiles 3 of 4 suites 3 of 4 (jp-compat fails) Remaining defects, filed:
- COBOL: full-width hyphens (－ U+FF0D / − U+2212) and full-width digits still truncate Japanese data names and drop sections (#3955 follow-up) #3991: full-width hyphens (
－U+FF0D /−U+2212) and full-width digits still truncate Japanese data names and drop sections. All 25 remaining truncations have this cause. - Generated REST controller references an undefined <File>DTO for a batch program without JCL: jp-compat still does not compile after #3957 #3992: in jp-compat, 10 controllers reference an undefined
TestFileDTO. This happens for a batch program with no JCL. It was previously hidden behind the Generated REST controller does not compile when a file is ASSIGNed to a literal path ("./TEST-FILE" -> @RequestBody ./testFileDTO) #3957 parse error.
Not filed: two small i18n_utf8 programs are still excluded as "Statistical Anomaly" (previously 7 across suites), and one name is missing in
095-data-description-unterminated, a deliberately malformed program.Next: fix #3991 / #3992 and rescan.
trial startwaits until the harness extension that raises eligibility (stdout as a proved channel) has landed, because eligibility is frozen at start.🤖 Generated with Claude Code
- COBOL: full-width hyphens (－ U+FF0D / − U+2212) and full-width digits still truncate Japanese data names and drop sections (#3955 follow-up) #3991: full-width hyphens (
Rescan on origin/main
71d33c52(after #3994 / #3995), pre-trial, no porting2026-09-28 6991970d71d33c52Japanese PROGRAM-IDs not refracted 12 0 0 data names missing 51 0 0 Japanese data names truncated 59 25 0 Japanese sections / paragraphs missing 36 28 3* ASCII names truncated 7 0 0 generated project compiles ( mvn -o compile, JDK 17)3 of 4 3 of 4 4 of 4 * The 3 remaining "missing" entries are all
003-long-field-name, and all 3 are false positives of the comparison script. A data name sits alone on a line as a MOVE/DISPLAY operand, and the script reads that line as a paragraph. The engine records the data item whole.Still open, not filed:
- Two i18n_utf8 programs are excluded as "Statistical Anomaly" (019, 126). i18n_utf8 is not in upstream CI and is out of scope for the trial.
- jp-compat
095-data-description-unterminatedis a deliberately malformed program.
The scan side of the estate is clean. The next step before
trial startis the harness extension that raises eligibility; eligibility is frozen at start.🤖 Generated with Claude Code
- added this to the New Challenger Approaches 🥊 — Fresh Estates milestone
on Sep 30, 2026
Part of epic #3803. Needs input generation and the trial tooling.
Pick estates no development corpus contains: public COBOL batch applications with JCL and copybooks (ideally VSAM / sequential I/O), under a licence that lets us publish results. For each one: freeze, trial, iterate to identical or out of scope, and publish the histogram.
State on the chart that public code may be in a model's training data. A private estate (e.g. a pilot's, if its terms allow) is the uncontaminated data point.
🤖 Generated with Claude Code