Census mock trial: take the 40 census programs from 8 whole to 40 of 40 accounted, measured by the tool only

Owner request 2026-11-04: treat the 40 census programs (9 public repos excluded from the blind study, never burned, counted by tools only under opaque labels) as a mock trial of the engine. Write-up: https://claude.ai/artifact/9CJkXtuprzxzRne4bty3dy

Baseline (last real survey 2026-10-24, `6758295` / #6361)

8 of 40 translate whole (20%). Of the 32 blocked: 13 only by source / clone problems (missing estate copybooks 9, undeclared items #7647, ISPF screens #7645), 17 only by translator gaps, 2 by both. On the 25 programs with no source or clone problem: 8 of 25 (32%).

Predictions (frozen before the first fix)

Translator work alone takes the census to 25 whole (62%); the other 15 end as named source / clone problems: 40 of 40 accounted.
The tool's coverage-if-closed (upper bounds; they count the missing-copybook gap as closable): top 1 -> 17, top 3 -> 20, top 5 -> 29, top 10 -> 32.
Cost: 9 translator gap classes x ~4-5 PRs per class (calibrated on Siblings of #5042: cics-crucible programs that do not translate whole (17 programs, 8 hole classes) #6876: 17 programs, 8 classes, 35 PRs, 6 days) = 35-45 PRs, 6-9 days at two workers.

Protocol

First (owner 2026-11-04): Translator consumes engine record layouts (fact_ownership.md follow-up): POINTER width as a parameter, shared stand-in copybooks, retire the duplicate layout computation #5649 -- the translator consumes the engine's record layouts (routed to nyx). The T0 re-survey waits until Translator consumes engine record layouts (fact_ownership.md follow-up): POINTER width as a parameter, shared stand-in copybooks, retire the duplicate layout computation #5649 has merged; T0 then measures main with Translator consumes engine record layouts (fact_ownership.md follow-up): POINTER width as a parameter, shared stand-in copybooks, retire the duplicate layout computation #5649 in, and the ledger notes it.
T0: re-survey the census at current main with no census-targeted change (cics_census.py survey over the census repos, then gap_prevalence.py --census-survey ... --census-corpora ... --write); post the rate here. The 10-24 -> T0 delta is what the burned-estate work did for unseen code.
Nobody reads census source. A fix gets the gap label, the IBM manual page and the owning issue, and proves itself on our own test (a cics-crucible program or a fixture), never a census excerpt. Tightened by the owner 2026-11-06: no edited, patched or excerpted copy of census source either, even unprinted (no "move this line and re-parse" probes). Allowed: the tools run over the clone as is, with aggregate or scrubbed output. A finding that needs more than that is reported as "needs a fixture" and proven on one.
Measure with the tool only (aggregates, opaque labels), after each fix merges; append one ledger row below.
A gap that is the source's fault becomes a named source defect with its reason.
L1 only (translates whole); no claim about L2+.
Finish: 40 of 40 accounted (basis 39 since the owner ruling 2026-11-06: one of the 40 is a data-only copy member, census: ExprError is a named source defect, UnitRefused a copy member; both rows owned (Part of #7847, #7709) #7885 / census: inventory counts a data-only copy member as a program (basis 40 -> 39, #7709) #7901; report "of 39" from T9 on, with the change named) (whole or named), time box 14 days from T0; publish the ledger as it stands at the time box.

Gap classes (owners)

Owner rulings 2026-11-06: (1) code-page literals stay a named source defect as stored; #7526 closed as accounted. (2) The 2 IBM z/OS Connect programs (BAQ product copybooks) stay unaccounted: a product dependency is not a source or clone defect; frontier #7900. (3) Basis 39 (#7901). (4) Rule 2 tightened (above).

Ledger (one row per re-survey)

date	main	whole / 40	gap closed (PRs)	predicted vs actual	new gaps exposed	cost (PRs, worker-h, $)	named, not fixed
2026-10-24	`6758295`	8	baseline	-	-	-	-

Related: #6064 (no current non-burned translation rate), #5865 (coverage-if-closed never validated against a run), #3803 (the real blind trial this rehearses).

T0 re-survey (owner 2026-11-04: run T0 now, ahead of #5649)

8 of 40 census programs translate whole (20.0%) at main 098ddd5d1, unchanged from the 2026-10-24 baseline (8 / 40 at 6758295f2). So the 10-24 -> T0 delta, what the burned-estate work did for unseen code at L1, is 0 programs.

Ledger row:

date	main	whole / 40	gap closed (PRs)	predicted vs actual	new gaps exposed	cost (PRs, worker-h, $)	named, not fixed
2026-11-04	`098ddd5`	8	T0 (no census fix counted; see below)	-	-	-	LayoutError 9 now named a source defect (#7714)

Census-targeted changes that merged before T0. Option 1 bent the T0-after-#5649 order to stop more of these leaking into the baseline:

gap_prevalence: census LayoutError programs are source defects; regenerate census page (Part of #5865, #7644, #7709) #7714 (17:52Z): the 9 bare LayoutError programs are keyed as a source defect (text past column 72, det: name source defects (text past column 72 in EXEC SQL, ISPF screen captures, undeclared items) instead of 'no such item' / SystemExit #4735). This changes how they are named, not the whole count.
det: translate(zos_comp5=True) holds COMP-5 big-endian and lifts the byte-use refusals (Part of #5865, #7646, slice 2 of 3) #7673 (18:09Z): COMP-5 big-endian behind translate(zos_comp5=True). The survey does not set the switch, so bytes overlap another item still blocks 2 programs.
Translator consumes engine record layouts (fact_ownership.md follow-up): POINTER width as a parameter, shared stand-in copybooks, retire the duplicate layout computation #5649 now becomes the trial's first fix row instead of part of the baseline. det: IBM's default DFHCOMMAREA and DFHAID/DFHBMSCA constants leave the census 'no such item' class (Part of #7709) #7723 (no such item, 3 programs) was a draft with no auto-merge, held until this T0 was posted, and may now proceed.

Coverage if closed at T0 (upper bounds, unchanged from the frozen predictions): top 1 -> 17, top 3 -> 20, top 5 -> 29, top 10 -> 32.

Method (tools only, no census source read):

cics_census.py survey --baseline --no-fetch --sha 098ddd5d1 --root <shared-scratch> (194 s).
gap_prevalence.py --survey L --census-survey L --census-corpora <shared-scratch>/census --top 1 3 5 10, where L is a temporary directory linking only before-b and before-nb (the 9 provisioned census corpora). The cache's before-nb-local (2 local non-burned corpora, 12 programs) was left out, as the 10-24 survey did.
Not run with --write: the committed census page was regenerated by gap_prevalence: census LayoutError programs are source defects; regenerate census page (Part of #5865, #7644, #7709) #7714 today. Run --write with the first fix row.

The burned half of the same run: 126 of 129 burned programs translate whole.

🤖 Posted by Claude Code

det: IBM's default DFHCOMMAREA and DFHAID/DFHBMSCA constants leave th…

#7723 merged: a partial slice of [c:bced86e8] "No such item, 3 programs (no owner yet)"; left: the 3 programs leave the no such item class, but none becomes whole: they still carry 36, 11 and 2 other holes, each in its own gap class (promoter-bot)

T1 re-survey: first fix row

8 of 40 census programs translate whole (20.0%) at main 7f73c5700, unchanged from T0 (098ddd5d1).

date	main	whole / 40	gap closed (PRs)	predicted vs actual	new gaps exposed	cost (PRs, worker-h, $)	named, not fixed
2026-11-04	`7f73c57`	8	no such item, 3 programs (#7723); #7713 (part of #5649)	gap gone as expected; +0 whole (each of the 3 programs has another gap)	none	2 PRs	-

det: IBM's default DFHCOMMAREA and DFHAID/DFHBMSCA constants leave the census 'no such item' class (Part of #7709) #7723: the <name>: no such item row (3 programs, 1 estate) is gone from the census table. None of the 3 became whole, because each still carries at least one other gap.
Engine lays out the five shipped system members (Part of #5649) #7713 (the engine lays out the five shipped system members, part of Translator consumes engine record layouts (fact_ownership.md follow-up): POINTER width as a parameter, shared stand-in copybooks, retire the duplicate layout computation #5649): no census row changed.
Coverage if closed (upper bounds): top 1 -> 17, top 3 -> 20, top 5 -> 31 (T0: 29), top 10 -> 32. Closing the no-such-item gap concentrated the remaining blockers.
Not in this row: open PRs equivalence: declared_differences case field and COMP-5 oracle byte reversal in the comparator (Part of #7646) #7725 (COMP-5 oracle byte reversal) and det: EXEC SQL COMMIT / ROLLBACK in a CICS program answer -925 / -926 (Part of #7709, #7527) #7727 (EXEC SQL COMMIT / ROLLBACK in CICS).

Method: identical to T0 (cics_census.py survey --baseline at 7f73c5700, 177 s; gap_prevalence.py over before-b and before-nb only, the 9 census corpora). Tools only, no census source read.

🤖 Posted by Claude Code

Unmasked survey at T1 (`7f73c5700`): what the whole-program refusals hide

Whole-program refusals stop the translator, so their programs' other gaps were invisible to T0 and T1, and coverage-if-closed counts them as upper bounds. cics_census.py blockers --baseline --unmask switches one refusal check off in survey mode only. This is a what-if, not a translation. The run is the census survey only (whatif-nb). Aggregates only.

LayoutError: source text past column (check cut-literal): the switch reaches 7 of the class's 9 programs.

Would translate whole with it lifted: 0 of 7. Holes per program: 24-48.
Gaps hidden behind it, by programs: undeclared item 7, EXEC SQL ROLLBACK in a CICS program 5, MOVE literal beyond the single-byte code page 4. Then 1 program each: EXEC CICS map field missing (6 distinct fields), no generated screen for a map, bare EXEC CICS DUMP, GO TO form, STRING literal beyond the code page, grammar does not parse.

missing copybook (check missing-copybook): the switch reaches 1 of the 9 estate-owned programs; it takes only a single named member.

Would translate whole with it lifted: 0 of 1. It hides a LayoutDisagrees refusal.
The other 8 stay masked.

What it changes for the trial:

The frozen top-N coverage-if-closed upper bounds treat these 9 + 9 programs as one gap from whole. At least 8 of them are not.
EXEC SQL ROLLBACK in CICS sits behind the LayoutError refusal in 5 programs, on top of its 1 visible program. So det: EXEC SQL ROLLBACK in a CICS program (census: 1 non-burned program) #7527 / PR det: EXEC SQL COMMIT / ROLLBACK in a CICS program answer -925 / -926 (Part of #7709, #7527) #7727 touches 6 programs, not 1, though none becomes whole while the source defect stands.
Both groups already count as accounted (named source / clone problems), so the 40-of-40 accounting goal is unaffected. Only the 25-whole prediction rests on the upper bounds.

Tool gap (sibling, for the class sweep): blockers --unmask prints census repository and program names in its per-program rows. gap_prevalence's census half scrubs these (scrub_key, opaque census-* labels), but unmask mode does not, so a run breaks protocol rule 2 by default. It should scrub census rows the same way. I did not record or repeat the names.

🤖 Posted by Claude Code

docs: RT-SQL-UNIT-OF-WORK title and det_port_design note state CICS C…

#7728 merged: a partial slice of [c:4112a597] "EXEC SQL ROLLBACK in CICS, 1 program (#7527)"; left: proof case and census re-survey remain (promoter-bot)

Only-gap breakdown at T1 (`7f73c5700`): census only, all 40 programs

cics_census.py blockers DIR --label before --all-programs over the census survey (before-nb) alone. Names are scrubbed.

Distinct gap classes per program: 0: 8 (whole), 1: 26, 2: 1, 3: 3, 4: 1, 7: 1.

The 26 single-gap programs:

only gap	programs	kind
LayoutError, source text past column (cuts a literal / truncates)	9	source defect (named, #7714); unmask: 7 of them hide 24-48 holes
missing copybook (estate-owned)	9	clone problem; unmask reaches 1
ISPF editor screen, not a COBOL program	2	source (named, #4735)
UnitRefused: a member to be copied, not a program	1	inventory / source
undeclared item	1	source defect (#7647)
bytes overlap another item (COMP-1/2/5)	2	translator (#7646)
LayoutDisagrees	1	translator (#6926)
ExprError: the PROCEDURE DIVISION does not parse	1	translator or source, unknown

The 6 multi-gap programs (2-7 classes each) hold the long tail: JSON GENERATE / PARSE, the three EXEC CICS WEB forms, EXEC CICS map fields, EXEC SQL ROLLBACK, MOVE / STRING literals beyond the single-byte code page, a STRING numeric sender, SQLCA undeclared, PERFORM undeclared paragraph. One away: MOVE code-page literal 1, STRING numeric sender 1.

What it means for the PR estimate:

22 programs need naming only, not translator work.
4 single-translator-gap programs: the bytes-overlap pair is closest, because the COMP-5 work is in flight.
6 multi-gap programs: about 10 classes. This is where the remaining ~30-45 PRs go.
COMP-5 alone has taken 6 PRs so far, above the 4-5 per class calibrated on Siblings of #5042: cics-crucible programs that do not translate whole (17 programs, 8 hole classes) #6876.

🤖 Posted by Claude Code

tests: pin the whole 'no such item' class -- modelled EIB fields, cen…

det: STRING refuses a numeric literal sender / delimiter by name (Par…

#7735 merged: a partial slice of [c:b383e52f] "STRING numeric / decimal sender, code-page gaps, 3 programs (#7523)"; left: the code-page row (register D4) is refused by design and has nothing to implement; numeric-sender and numeric-literal STRING operands are now refused by name. (promoter-bot)

cics web: WEB WRITE HTTPHEADER SESSTOKEN translates on an assumed seq…

gap_prevalence: empty KNOWN_UNOWNED now that the census has no unowne…

census: --unmask runs skip a copy member; survey.md lists copy member…

det: JSON PARSE takes COMP / COMP-3 / COMP-5 numeric receivers (Part of

#7920 merged: a partial slice of [c:9bd1fd57] "JSON GENERATE / PARSE, 2 programs (#7524, reopened 2026-11-05 by the owner; work it there)"; left: JSON GENERATE exception JSON-CODE value (census GENERATE row), PARSE tables, JSON-CODE/JSON-STATUS after PARSE exception, CONVERTING, NAME OMITTED, OCCURS DEPENDING ON, SUPPRESS WHEN (#7905) (promoter-bot)

census: ExprError banner finding re-proved on an authored fixture (Pa…

cobolrt API manifest: add JsonParse.tryParse status overload as publi…

cobolrt API manifest: the emitter calls JsonParse.tryParse's six-argu…

census: gap_prevalence census half reports programs over the correcte…

harness: read the oracle's COMP-5 container bytes as z/OS holds them …

Merge origin/main into churn/7847-20261106T1234Z; regenerate the cens…

T9 re-survey (`7dc1d0138`), basis 39

Basis change: 40 -> 39 (owner ruling 2026-11-06, decision 3a). The 40th member is a data-only copy member (#7885); the inventory now lists it and does not count it (#7901, #7921, #7926).

11 of 39 census programs translate whole (28.2%); 35 of 39 accounted; ceiling under the rulings 37 of 39 (the 2 z/OS Connect programs excluded). Burned half: 126 of 129, unchanged.

Coverage if closed: top 1 -> 12, top 3 -> 23, top 5 -> 31, top 10 -> 36 (upper bounds).
Against the page census: gap_prevalence census half reports programs over the corrected basis (40 -> 39) (Part of #7901, #7709) #7926 committed: accounted 36 -> 35, top 1 13 -> 12. The WEB rows now read as one WEB CONVERSE POST gap, and one program moved from undeclared item to other. Not attributed further.
Tools: cics_census survey --baseline, blockers --baseline, gap_prevalence. blockers and the history line now print basis: ... copy member(s) listed, not counted (PR census: T9 re-survey on the 39 basis; blockers and history name the copy members (Part of #7901, #7709) #7934). No census source read or copied.

🤖 Posted by Claude Code

census: JSON PARSE / GENERATE refusals no longer collapse to the unow…

det: run a COMP-5 group through LINK CHANNEL between two translated p…

cobolrt API manifest: keep JsonParse.tryParse status overload public …

Navigation Menu

Repository navigation

Census mock trial: take the 40 census programs from 8 whole to 40 of 40 accounted, measured by the tool only #7709

Baseline (last real survey 2026-10-24, `6758295` / #6361)

Predictions (frozen before the first fix)

Protocol

Gap classes (owners)

Ledger (one row per re-survey)

T0 re-survey (owner 2026-11-04: run T0 now, ahead of #5649)

T1 re-survey: first fix row

Unmasked survey at T1 (`7f73c5700`): what the whole-program refusals hide

Only-gap breakdown at T1 (`7f73c5700`): census only, all 40 programs

87 remaining items

T9 re-survey (`7dc1d0138`), basis 39

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

Census mock trial: take the 40 census programs from 8 whole to 40 of 40 accounted, measured by the tool only #7709

Description

Baseline (last real survey 2026-10-24, 6758295 / #6361)

Predictions (frozen before the first fix)

Protocol

Gap classes (owners)

Ledger (one row per re-survey)

Activity

squid-protocol commented on Nov 4, 2026

T0 re-survey (owner 2026-11-04: run T0 now, ahead of #5649)

squid-protocol commented on Nov 4, 2026

squid-protocol commented on Nov 4, 2026

T1 re-survey: first fix row

squid-protocol commented on Nov 4, 2026

Unmasked survey at T1 (7f73c5700): what the whole-program refusals hide

squid-protocol commented on Nov 4, 2026

squid-protocol commented on Nov 4, 2026

Only-gap breakdown at T1 (7f73c5700): census only, all 40 programs

squid-protocol commented on Nov 4, 2026

87 remaining items

squid-protocol commented on Nov 6, 2026

squid-protocol commented on Nov 6, 2026

T9 re-survey (7dc1d0138), basis 39

Metadata

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

Issue actions

Baseline (last real survey 2026-10-24, `6758295` / #6361)

Unmasked survey at T1 (`7f73c5700`): what the whole-program refusals hide

Only-gap breakdown at T1 (`7f73c5700`): census only, all 40 programs

T9 re-survey (`7dc1d0138`), basis 39