Every number on this page has a source: the deterministic scanner, the MAT
export, or the oracle's byte-diff. Confidence is stated the Quarry way — greens are
machine-verified, ambers are known gaps, reds are honest limits.
100%
pilot CBACT01C verified P1→P2→P3 end to end
byte-for-byte oracle match, twice (hand port + Gemini's)
13.8%
P2 · already runs on Linux 3,167 of 22,904 code lines · 11 of 44 programs compile
measured by the GnuCOBOL compile sweep — same source, new machine
1.6%
P3 · rewritten in PHP 358 of 22,904 code lines · 1 of 44 programs
the pilot — plus 1 assembler module replaced and 3 inquiry screens rebuilt
100%
of translated code tested every port must pass the oracle
50 records · 5,350 bytes byte-compared · on 2 machines (Mac + VM)
whole-system confidence P2 → P3 PHP / target language
same weights − language risk; oracle-gated
How to read this in one breath: the two "translated" numbers are
different phases. P2 = 13.8%: 11 of 44 programs already compile and
run on Linux with GnuCOBOL (same COBOL, no mainframe). P3 = 1.6%: 1 program (the
pilot) has been fully rewritten in PHP. We tested 100% of what was rewritten (oracle
byte-diff — not code review), and the pilot passed completely. Scaling that evidence
across all 44 programs by category difficulty puts whole-system confidence at
≈71% for the Linux replatform and ≈66% for the language rewrite.
The pilot numbers are measurements; the whole-system numbers are projections that turn
into measurements as each program passes the oracle (clean batch first, CICS screens next,
DB2/IMS last).
Progress by phase — where each number belongs
P1 is finished: every program was analyzed. The conversion phases (P2, P3)
are at the pilot stage — that is where the 1.6% lives.
P1 · Analysis (mainframe COBOL understood)MAT + Quarry judgments over the whole system
100%
44/44 programs · 46 JCL · 83 datasets — COMPLETE
P2 · COBOL on Linux (replatform)same source compiled with GnuCOBOL, no mainframe
P3 · PHP rewrite (target language)AI-translated, only counts after passing the oracle
1.6%
1/44 programs — two ports (hand + Gemini), both byte-verified
The translation factory — watch the process
A replay of the measured pipeline over all 44 programs: each one is analyzed,
then attempts the Linux compile, and lands on its real, measured status — nothing is
painted green that did not actually pass. Press replay during a demo.
verified end-to-end compiles on Linux copybook fix needed CICS → screen rewrite IMS/DB2 → database port
The whole system — everything scanned
From the deterministic scan + MAT assessment of CardDemo.
44
COBOL programsdeterministic scan
22,904
code linesscanner-counted
65
copybooks
46
JCL jobs
83
datasetsMAT-inferred
44/44
judgment coverageQuarry LLM 5 + MAT 44
Translation verification
The oracle: same input into the COBOL and the PHP port, outputs byte-compared.
2/2
ports pass oraclehand-written + Gemini
50
records compared
5,350
bytes comparedper side, per run
100%
byte equivalenceOUT-FILE, both ports
1
asm module replacedCOBDATFT → COBOL
2
runtimes provenmacOS arm64 + VM x86_64
Confidence — stated honestly
✓
Output equivalence — proven
Oracle byte-diff: COBOL OUT-FILE vs PHP port = identical, 50 records / 5,350 bytes, zero differing bytes. Holds for the hand port and for Gemini's clean-room port.
✓
Hidden business rule carried
The 2525.00 zero-cycle-debit quirk was reproduced by both ports; Gemini discovered it from the COBOL source alone (it was never mentioned in the prompt).
✓
Binary encodings correct
Zoned decimal (sign overpunch) and COMP-3 packed decimal produced byte-identical results — verified by the same diff, not by inspection.
✓
Portable across machines
Same GnuCOBOL binary output verified on macOS arm64 and on the cloud VM poc-smk (Linux x86_64).
◐
Data coverage is partial
The 50 sample accounts contain only positive balances — the negative-sign encoding path is written but has never been exercised by the oracle. Seed negative-balance records before claiming it.
◐
Only OUT-FILE is compared
CBACT01C also writes ARRYFILE and VBRCFILE and prints DISPLAY output; those are produced but not yet oracle-compared.
✗
Scale: 1 of 44 programs
One batch program (CBACT01C) has completed the full pipeline. 43 programs remain, including CICS online programs that need a different runtime strategy.
✗
First Gemini run was invalid — and was caught
Gemini initially read the reference port from the repo workspace (verbatim comments). Flagged, quarantined, re-run in an empty sandbox. The clean run is the one reported.
The 12-step pipeline — where each step stands
1
Import flat file into MATphase1-legacy zip, 1.4 MB