Results

Overall

Recomputing this figure. The overall score is a flat mean of the 42 per-program scores in results/scores.csv, which computes to 98.666667 and publishes as 98.7. It is deliberately not an average of the eight archetype means, which would over-weight archetypes containing one or two programs; that calculation gives 98.5. Where the two disagree, §4.3 of METHODOLOGY.md mandates the flat per-program mean.

By archetype

Archetype Programs Mean score
CICS Other 4 99.4
Batch File Maintenance 11 99.1
Batch Report 2 99.1
CICS Inquiry 8 98.6
CICS Menu 6 98.4
CICS Data Entry 7 98.3
Batch Utility 1 97.7
Called Subprogram 3 97.7
Overall 42 98.7

Archetype assignment in this table reflects the rubric applied at scoring; one program (PAUDBUNL) was classified batch-report by the pipeline classifier but scored under the batch-file-maintenance rubric — a recorded assignment mismatch.

Per-program scores

Every program's score of record (mean of its three per-generation medians), one decimal, with its separately-gated fabrication count. Machine-readable copy: results/scores.csv.

Program Archetype Score Fabrications
COPAUS2C CICS Other 100.0 0
CBACT04C Batch File Maintenance 99.9 0
COACCT01 CICS Other 99.8 0
COCRDLIC CICS Menu 99.8 0
CBACT02C Batch File Maintenance 99.7 0
CBACT03C Batch File Maintenance 99.7 0
CBACT01C Batch File Maintenance 99.5 0
CBTRN03C Batch Report 99.5 0
COBTUPDT Batch File Maintenance 99.3 0
COCRDSLC CICS Inquiry 99.3 0
CBTRN01C Batch File Maintenance 99.3 0
COUSR01C CICS Data Entry 99.3 0
COADM01C CICS Menu 99.2 0
COTRTUPC CICS Inquiry 99.2 0
CBCUS01C Batch File Maintenance 99.1 0
COTRN02C CICS Data Entry 99.1 0
COPAUA0C CICS Other 99.0 0
COPAUS0C CICS Inquiry 99.0 0
COTRTLIC CICS Menu 99.0 0
COCRDUPC CICS Data Entry 98.9 0
COUSR03C CICS Data Entry 98.9 0
COACTVWC CICS Inquiry 98.8 0
CBSTM03B Batch File Maintenance 98.8 0
PAUDBUNL Batch File Maintenance 98.7 0
CBSTM03A Batch Report 98.7 0
CODATE01 CICS Other 98.7 0
CORPT00C CICS Inquiry 98.6 0
CBTRN02C Batch File Maintenance 98.5 0
COSGN00C CICS Inquiry 98.5 0
COUSR00C CICS Menu 98.5 0
CBPAUP0C Called Subprogram 98.4 0
CSUTLDTC Called Subprogram 98.2 0
COACTUPC CICS Data Entry 98.1 0
COPAUS1C CICS Inquiry 98.0 0
COBSWAIT Batch Utility 97.7 0
COUSR02C CICS Data Entry 97.6 0
COTRN01C CICS Inquiry 97.4 0
PAUDBLOD Batch File Maintenance 97.2 0
COMEN01C CICS Menu 97.1 0
COTRN00C CICS Menu 97.0 0
COBIL00C CICS Data Entry 96.5 0
DBUNLDGS Called Subprogram 96.5 0

Reading the scores

What a maintenance programmer gets. At these scores, a programmer receiving the PLMs can trust the field names, control-flow descriptions, I/O operations, and business-logic explanations without cross-referencing source code for basic comprehension. That is the bar the benchmark is designed to test — not literary quality, but whether the document is safe to act on.

Where the 1.3 points go. Deductions below 100 typically reflect inferences about business context that cannot be mechanically verified from source alone — for example, who consumes a report or when a batch job runs in production. In well-formed PLMs these are qualified as inferences rather than asserted as fact, which is the correct behavior for regulated environments and is rewarded by the rubric even though it caps the score short of 100.

Where the low archetypes sit, and why. The two lowest archetype means are Batch Utility (97.7, n = 1) and Called Subprogram (97.7, n = 3) — both small samples, so a single program moves the mean and neither number should be read as a broad claim about the archetype. Their scores are driven by the same cap the rest of the corpus hits, only more of it: these programs carry a higher proportion of non-derivable operational and caller context (what invokes a pure linkage routine, or what a utility does in the production job stream) that the source alone does not settle, and which well-formed PLMs correctly hedge as inference rather than assert. The cap is by design — appropriately-qualified uncertainty is scored as correctness — not a quality defect. (Batch Report, previously the low archetype at 96.7, is no longer low: at 99.1 it sits with the strongest batch programs after the pipeline fixes in this vintage.)

The bottom line

A mean of 98.7 across 42 programs spanning 8 archetypes, with zero fabricated entities, means the generated documentation is consistently accurate enough for production maintenance use — and, because fabrication is gated separately, accurate in a way that holds up to audit rather than merely reading well.