Report Comparison

Engine: PMLC = python3 <skills-root>/learnings-coach/scripts/pmlc.py (this skill needs the learnings-coach skill installed alongside it). Read learnings-coach/references/comparison-method.md before starting, and learning-principles.md.

Inputs

• the ProbeMatrix workspace (ingested and classified)
• the other report (PDF, Word or text). Extract its text with whatever tools the host provides.

Steps

1. Establish sighted or blind. Search the findings for quotations of the other report's recommendation or lessons sections, e.g. PMLC select <ws> --text "recommendation|remedial" --brief, and look for page references that fall inside those sections. Report how many findings cite them. If any do, the comparison is sighted; say so on page 1.
2. Inventory the other report: observations (OBS-nn), root causes (RC-nn), recommendations (REC-nn), each with its section or discipline group. Count them.
3. Learnings on the ProbeMatrix side: use an existing Learnings Register if present; otherwise build a compact one (see learnings-register, steps 1–2) so both sides are compared at the level of learnings, not finding instances.
4. Map both ways (yes / partial / none) with reason classes for "none".
5. Compare dimensions: level reached, ownership (entity and tier named?), actionability, control tier of recommendations, evidence basis, interfaces.
6. Cross-tabulate the other report's discipline groups against Learning Clusters to show where each concentrates.
7. Names do not bleed in. Replace personal names in quoted comparator text with roles (and company names with their function) before they reach a table; see the name-bleed rule in comparison-method.md. Check with scan-names or check-language --names using the comparator's names. Put a conflict-of-interest line on page 1 when the commissioning organisation is one of the compared parties.
8. Say where the other report is stronger. Specific engineering recommendations, site findings, primary evidence: credit them.

Output

Write the pack <run>/outputs/report-comparison.json (schema: learnings-coach/references/pack-schema.md), then:

PMLC check-pack <ws> <run>/outputs/report-comparison.json   # fix every problem it lists
PMLC pack <ws> <run>/outputs/report-comparison.json         # -> report-comparison.html (primary) and .md

The pack supplies the status line and the footer. Sections and the kinds that fill them:

1. Basis: a para with the sighted or blind statement and the "same evidence, different analytical architecture" framing.
2. Bottom line: five entries.
3. Coverage: one table per direction (Other report, ProbeMatrix learning, Match = yes, partial or none, Reason if none; row_cites), plus a counts table.
4. Depth and ownership: a table (Dimension, Other report, ProbeMatrix).
5. Control-tier profile: a table.
6. Side by side: a table (Topic, Other report says, ProbeMatrix says; row_cites) with three to five rows.
7. Where the other report is stronger: para.
8. Limits: para.

Use the cross-tab of discipline groups × Learning Clusters as a further table. The other report's IDs (OBS-nn, RC-nn, REC-nn) are plain text, not citations. Put data-quality corrections in data_quality, findings left out on purpose in exclusions, and same-condition findings in supporting_cites.

If the user will share the comparison with the organisation that wrote the other report, also write <run>/outputs/report-comparison-client-safe.json with "pack": "report-comparison-client-safe". Keep the same content, use neutral wording and make no product claims.

{"kind": "table", "columns": ["Other report", "ProbeMatrix learning", "Match", "Reason if none"],
 "rows": [["REC-01 …", "…", "partial", ""]], "row_cites": [["FND-…"]]}