oracle policy: IBM is the reference, GnuCOBOL the instrument — audit register entries + z/OS calibration oracle #4702

Owner policy (2026-10-08): IBM is the reference; GnuCOBOL is the instrument.

• Where IBM documents a result under the estate's compile options (TRUNC, NUMPROC, ARITH, ZONEDATA, ...), the det port follows IBM. The oracle (cobc 3.1.2 -std=ibm) is configured to match through cobc options, config or a wrapper; if that's impossible, the shape is a declared oracle-vs-IBM difference the comparator expects. Proofs are never loosened.
• Where IBM's result is undefined (e.g. non-digit data, divide by zero without ON SIZE ERROR, which can abend S0CB), the port refuses by name or follows cobc, and the register says so explicitly.
• GnuCOBOL stays the bulk oracle: free, in CI, reproducible by purchasers. A real z/OS (Enterprise COBOL) is a calibration oracle: run the probe and fuzz corpus there occasionally to replace "IBM side ASSUMED" with measured results. It is not in CI.

Part 1: audit register entries that chose "match GnuCOBOL"

For each entry: the IBM reference (cited), whether IBM documents it under which options, what cobc can be made to do, the chosen behaviour, and the register entry updated.

• C2 (det runtime: arithmetic intermediates are not truncated as the oracle's ARITHMETIC-OSVS (IBM's fixed-point decimal places) truncates them (register C2) #4287 / det runtime: truncate arithmetic intermediates as cobc's ARITHMETIC-OSVS does (#4287) #4654, merged): OSVS intermediate truncation, where the oracle departs from IBM's Appendix A rule; the 30-digit cap is not modelled.
• C4 (det runtime: a result overflowing a binary receiver's bytes (no ON SIZE ERROR) is stored unlike GnuCOBOL #4684 / det: an unsigned binary ADD / SUBTRACT wraps as cobc's native arithmetic does; lifted COMP-5 under TRUNC(STD) (#4684) #4689): unsigned binary ADD/SUBTRACT. Decided: IBM absolute value (on det: an unsigned binary ADD / SUBTRACT wraps as cobc's native arithmetic does; lifted COMP-5 under TRUNC(STD) (#4684) #4689).
• C11 (equivalence: LGTESTP4's add moves a space-filled CA-BROKERID / CA-PAYMENT to binary (GnuCOBOL 931773840, det port 0; C11) #4652 / equivalence: model the oracle's MOVE of a non-digit zoned item to a binary item (#4652) #4663 merged, equivalence: the rest of C11 -- non-digit zoned data in arithmetic, comparisons, edited and zoned MOVEs differs from the oracle #4662 / equivalence: non-digit zoned data as MOVE sender, operand, comparand; the oracle's sign rewrite (#4662) #4693): non-digit zoned data. Mostly undefined in IBM's docs; check ZONEDATA (Enterprise COBOL 6) and NUMPROC.
• C13 (det runtime: numeric-edited item compared with a numeric literal numerically; GnuCOBOL compares it as text #4665 / det: a numeric-edited item compared with a number is a text comparison (#4665) #4668, merged): numeric-edited vs numeric comparisons. The IBM side is ASSUMED; confirm against the Language Reference.
• C14 (det runtime: a zero divisor without ON SIZE ERROR throws in the port; the oracle leaves the receiver unchanged #4655 / det runtime: a zero divisor leaves the receivers unchanged, as the oracle (#4655) #4677, det runtime: a floating-point zero divisor still throws ArithmeticException (Hfp.divide) #4675 / det: a float zero divisor is libcob's NaN; a phrase binds to the innermost open statement (#4675, #4676) #4686): size errors and divide by zero (undefined or S0CB on IBM), and 0 ** -n (IBM says size error, the oracle returns 0).
• Batch Sysout: DISPLAY of a hex-literal byte >= X'80' is not byte-identical under a non-latin-1 record charset (cp037/cp1047/cp1252) #4698: non-latin-1 batch DISPLAY bytes and line terminator.

Part 2: z/OS calibration oracle (blocked on access and legal)

• Candidates: IBM Wazi as a Service (z/OS dev/test on IBM Cloud), ZD&T (licensed dev/test emulation), IBM Z Xplore (free; its terms likely don't cover commercial validation), or a customer's or partner's mainframe. z/OS on Hercules is not an option (licence ties z/OS to IBM hardware).
• Legal questions for counsel first: do the licence terms allow using outputs to validate a commercial translator, and may results be published (relates to the publishing hold, legal Q16)?
• Deliverable: a probe corpus (the register shapes plus the det fuzz kits) runnable on Enterprise COBOL with recorded results, and a "z/OS-calibrated" marker per register entry and in the evidence report.

🤖 Generated with Claude Code

Remaining slices (as of 2026-10-18)

Done so far: 3 merged PRs (#5308, #5384, #5396).

• C11 leftovers: quote or settle the remaining ASSUMED bullets (P-scaled items, UNSTRING numeric INTO) from the 6.4 Language Reference (C11 register page, cobolrt_rules.json). Check: register.py --check, test_cobolrt_rules.py. (1 PR, docs) (register: C11 UNSTRING / P-scaled bullets quoted from the 6.4 Language Reference (Part of #4702) #5488)
• C2 "To settle": probe the 31+ digit intermediate on the Docker cobc and record it as matched or a declared difference. Check: register.py --check + the probe output quoted. (1 PR) (C2: probe the 31+ digit intermediate on the oracle (Part of #4702) #5502)
• Batch Sysout: DISPLAY of a hex-literal byte >= X'80' is not byte-identical under a non-latin-1 record charset (cp037/cp1047/cp1252) #4698 / D2 re-check: non-latin-1 batch DISPLAY bytes and line terminator against D2 as updated in register: D2, F1 and F3 quote IBM's text and declare the oracle difference (Part of #4702) #5396. Check: register.py --check; a behaviour change needs a golden bless. (1 PR) (D2: Sysout writes X'25' under IBM037 as its own byte; low bytes proven against cobc (Part of #4702) #5737)
• (verify) ZWB/NOZWB (C13): only if a burned program needs it (register: C13, C14, C4 cite IBM's text instead of assuming it (Part of #4702) #5308 says none); else close as nothing to prove. (1 PR or none) (register: ZWB/NOZWB (C13) -- no burned program needs it, nothing to prove (Part of #4702) #5771)
• (verify) Packed/binary vs nonnumeric comparison (IBM rejects, the port compares digits): grep the burned corpora before any change, since refusing moves goldens. (1 PR or none) (det: a packed or binary item compared with a nonnumeric operand is refused, IBM rejects it (C13) (Part of #4702) #5797)
• Register sweep for other "match GnuCOBOL" / ASSUMED entries outside the C/D/F members done so far. Check: register.py --check + a list of entries with no IBM quote or declared difference. (1 PR) (register: strict IBM citation for X16, X23, X24; STRICT_BASELINE emptied (Part of #4702) #6044)
• Part 2: z/OS calibration probe corpus and a "z/OS-calibrated" marker per register entry and in the evidence report. Blocked on access and legal; do not start.
  Done when: every register entry that chose "match GnuCOBOL" cites IBM's text and matches it or declares the difference, and Part 2 has recorded z/OS results (or the owner descopes Part 2).
  Owner decision needed: z/OS access (Wazi as a Service, ZD&T or a customer mainframe) and counsel on licence/publishing (Q16); and whether Part 2 stays here or splits off, since it blocks closure indefinitely.

Drafted 2026-10-18 by a Sonnet triage pass from this issue, its comments and the bodies of its "Part of" PRs; (verify) items were not checked against main. Tick boxes as slices merge; edit freely.