MetacouplingLLM 0.2.1__tar.gz → 0.2.2__tar.gz
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/CHANGELOG.md +726 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/INTRODUCTION.md +11 -8
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/MANUAL.md +212 -50
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/PKG-INFO +12 -9
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/README.md +2 -1
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/docs/METHODS_adjacency.md +55 -6
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/docs/REPRODUCING.md +1 -1
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/docs/adm1_aliases_audit.csv +118 -7
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/pyproject.toml +1 -1
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/scripts/validate_adm1_aliases.py +263 -22
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/src/metacouplingllm/__init__.py +3 -1
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/src/metacouplingllm/core.py +132 -61
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/src/metacouplingllm/data/PROVENANCE.md +1 -1
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/src/metacouplingllm/data/adm1_aliases.csv +118 -7
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/src/metacouplingllm/knowledge/adm1_pericoupling.py +508 -9
- metacouplingllm-0.2.2/src/metacouplingllm/knowledge/countries.py +1237 -0
- metacouplingllm-0.2.2/src/metacouplingllm/knowledge/examples.py +1020 -0
- metacouplingllm-0.2.2/src/metacouplingllm/knowledge/places.py +292 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/src/metacouplingllm/knowledge/rag.py +55 -37
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/src/metacouplingllm/knowledge/websearch.py +12 -11
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/src/metacouplingllm/llm/client.py +38 -6
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/src/metacouplingllm/llm/parser.py +222 -39
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/src/metacouplingllm/output/formatter.py +30 -11
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/src/metacouplingllm/prompts/builder.py +110 -46
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/src/metacouplingllm/prompts/templates.py +3 -2
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/src/metacouplingllm/visualization/__init__.py +37 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/src/metacouplingllm/visualization/adm1_map.py +10 -1
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/src/metacouplingllm/visualization/worldmap.py +67 -10
- metacouplingllm-0.2.2/tests/test_adjacency_settings.py +561 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/tests/test_adm1_name_matching_regression.py +44 -2
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/tests/test_adm1_pericoupling.py +895 -11
- metacouplingllm-0.2.2/tests/test_anthropic_sdk.py +195 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/tests/test_builder.py +198 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/tests/test_client.py +126 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/tests/test_core.py +373 -0
- metacouplingllm-0.2.2/tests/test_countries.py +689 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/tests/test_docs_examples.py +47 -0
- metacouplingllm-0.2.2/tests/test_examples.py +248 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/tests/test_formatter.py +29 -1
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/tests/test_parser.py +630 -0
- metacouplingllm-0.2.2/tests/test_place_readings.py +179 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/tests/test_rag.py +145 -1
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/tests/test_relabel_sliver_corridors.py +2 -1
- metacouplingllm-0.2.2/tests/test_validation_notes.py +559 -0
- metacouplingllm-0.2.2/tests/test_visualization_wrappers.py +71 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/tests/test_websearch.py +82 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/tests/test_worldmap.py +111 -0
- metacouplingllm-0.2.1/src/metacouplingllm/knowledge/countries.py +0 -725
- metacouplingllm-0.2.1/src/metacouplingllm/knowledge/examples.py +0 -713
- metacouplingllm-0.2.1/tests/test_countries.py +0 -204
- metacouplingllm-0.2.1/tests/test_examples.py +0 -208
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/.github/workflows/test.yml +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/.gitignore +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/LICENSE +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/build_data/arc_and_combined_screens/SPEC_ba1_border_arc_and_screens.md +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/build_data/build_record/relabel_RESULTS.md +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/build_data/geodesic_distances/SPEC_gd1_geodesic_distances.md +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/build_data/ndlsa_ne/SPEC_da1_defacto_administrators.md +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/build_data/stage_order/SPEC_st1_stages.md +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/build_data/water_screen_rebuild/corridor_census_exact/SPEC_cc1_corridor_census.md +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/build_data/water_screen_rebuild/crossing_adjudication/SPEC_ca1_crossing_adjudication.md +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/build_data/water_screen_rebuild/crossing_width/SPEC_cw1_crossing_width.md +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/build_data/water_screen_rebuild/july_rejections/SPEC_jr1_july_rejections.md +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/build_data/water_screen_rebuild/rule_rejections/SPEC_mr1_rule_rejections.md +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/build_data/water_screen_rebuild/uniform_acceptance/SPEC_ua1_uniform_acceptance.md +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/docs/BRIDGE_CLASSIFICATION_METHODOLOGY.md +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/docs/FUTURE_EDGE_AUDITS.md +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/docs/FUTURE_WORK.md +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/docs/VALIDATION_SAMPLING_PLAN.md +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/docs/VALIDATION_STUDY.md +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/docs/ndlsa_tract_audit.csv +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/examples/brazil_soybean_adjacency.csv +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/examples/brazil_soybean_flows.csv +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/examples/fishing_eez/.gitignore +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/examples/fishing_eez/00_build_edge_table.py +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/examples/fishing_eez/01_build_region_map.py +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/examples/fishing_eez/02_compute_indicators.py +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/examples/fishing_eez/03_analysis.py +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/examples/fishing_eez/README.md +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/examples/fishing_eez/data/eez_region.json +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/examples/fishing_eez/data/fishing_edge_table.csv +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/examples/fishing_eez/data/s16_country_region.json +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/examples/fishing_eez/outputs/global_totals.csv +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/examples/fishing_eez/outputs/indicators_by_eez.csv +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/examples/fishing_eez/outputs/regional_catch.csv +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/examples/fishing_eez/outputs/regional_indicators.csv +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/examples/fishing_eez/outputs/telecoupled_share_timeseries.csv +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/examples/fishing_eez/run_all.py +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/scripts/apply_check_keywords.py +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/scripts/apply_overlays.py +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/scripts/audit_unreplaced.py +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/scripts/build_adm1_aliases.py +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/scripts/build_all.py +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/scripts/build_embeddings.py +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/scripts/build_paper_metadata_table.py +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/scripts/build_pericoupling_db.py +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/scripts/build_rag.py +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/scripts/check_no_stale_metacoupling.py +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/scripts/check_oa_status.py +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/scripts/copyright_audit.py +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/scripts/diagnose_anthropic_web_search.py +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/scripts/draw_validation_sample.py +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/scripts/extract_keywords.py +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/scripts/integrate_keywords.py +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/scripts/match_references.py +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/scripts/organize_pdfs.py +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/scripts/organize_pdfs_v2.py +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/scripts/organize_pdfs_v3.py +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/scripts/patch_doi_inventory.py +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/scripts/rebuild_rag_corpus.py +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/scripts/regenerate_paper_metadata.py +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/scripts/relabel_sliver_corridors.py +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/scripts/strategy_b_decision.py +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/scripts/trace_pipeline.py +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/scripts/verify_papers_vs_bib.py +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/src/metacouplingllm/data/Papers.zip +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/src/metacouplingllm/data/PeriTelecoupling_clean.csv +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/src/metacouplingllm/data/PeriTelecoupling_subset.csv +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/src/metacouplingllm/data/chunk_embeddings.manifest.json +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/src/metacouplingllm/data/chunk_embeddings.npy +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/src/metacouplingllm/data/denylist_pairs.csv +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/src/metacouplingllm/data/disputed_overlay_pairs.csv +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/src/metacouplingllm/data/land_gap_overlay_pairs.csv +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/src/metacouplingllm/data/pericoupled_adm1_edge_list.csv +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/src/metacouplingllm/data/sliver_corridor_relabel.csv +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/src/metacouplingllm/data/telecoupling_literature.bib +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/src/metacouplingllm/data/water_classification_pairs.csv +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/src/metacouplingllm/data/water_separated_pairs.csv +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/src/metacouplingllm/indicators/__init__.py +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/src/metacouplingllm/indicators/_math.py +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/src/metacouplingllm/indicators/_prompts.py +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/src/metacouplingllm/indicators/_schemas.py +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/src/metacouplingllm/indicators/adjacency.py +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/src/metacouplingllm/indicators/classify.py +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/src/metacouplingllm/indicators/core.py +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/src/metacouplingllm/indicators/llm.py +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/src/metacouplingllm/knowledge/__init__.py +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/src/metacouplingllm/knowledge/citations.py +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/src/metacouplingllm/knowledge/framework.py +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/src/metacouplingllm/knowledge/literature.py +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/src/metacouplingllm/knowledge/pericoupling.py +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/src/metacouplingllm/knowledge/references.py +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/src/metacouplingllm/llm/__init__.py +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/src/metacouplingllm/output/__init__.py +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/src/metacouplingllm/output/export.py +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/src/metacouplingllm/prompts/__init__.py +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/src/metacouplingllm/tracing.py +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/tests/__init__.py +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/tests/_helpers.py +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/tests/conftest.py +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/tests/data/adm1_resolve_golden.json +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/tests/data/avocado_main_analysis_response.md +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/tests/test_adm1_map.py +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/tests/test_apply_overlays.py +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/tests/test_build_pericoupling_db.py +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/tests/test_citations.py +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/tests/test_coupling_first.py +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/tests/test_data_csvs.py +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/tests/test_docs_capabilities.py +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/tests/test_framework.py +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/tests/test_gemini_grok_support.py +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/tests/test_indicators.py +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/tests/test_indicators_llm.py +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/tests/test_integration.py +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/tests/test_literature.py +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/tests/test_openai_integration.py +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/tests/test_output.py +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/tests/test_pericoupling.py +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/tests/test_rag_only_mode.py +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/tests/test_rag_pipeline.py +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/tests/test_structured_extraction.py +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/tests/test_templates.py +0 -0
- {metacouplingllm-0.2.1 → metacouplingllm-0.2.2}/tests/test_tracing.py +0 -0
|
@@ -7,6 +7,732 @@ file. The format is loosely based on
|
|
|
7
7
|
|
|
8
8
|
## [Unreleased]
|
|
9
9
|
|
|
10
|
+
## [0.2.2] — 2026-10-24
|
|
11
|
+
|
|
12
|
+
A maintenance release: the Claude adapter works with anthropic 2.x, two
|
|
13
|
+
copies of the package no longer clash over the paper cache, place names are
|
|
14
|
+
read as texts write them, analyses are parsed in more of the forms models
|
|
15
|
+
write, and the adjacency settings reach the prompt and the maps. No
|
|
16
|
+
documented API is removed or renamed; `AnthropicAdapter`'s default model is
|
|
17
|
+
now `claude-sonnet-5-5`.
|
|
18
|
+
|
|
19
|
+
Highlights:
|
|
20
|
+
|
|
21
|
+
- **The Claude adapter works with anthropic 2.x**: anthropic 2.0.0
|
|
22
|
+
(2026-10-21) targets API version 2026-10-21, which has no `temperature`,
|
|
23
|
+
so on a fresh install every `AnthropicAdapter` call raised `TypeError`
|
|
24
|
+
except with `claude-opus-4-7` (0.2.1 too). The adapter now sends
|
|
25
|
+
`temperature` only to SDKs and models that take it (not anthropic 2.x,
|
|
26
|
+
not Claude 4.7 and later), and its default model is `claude-sonnet-5-5`;
|
|
27
|
+
the old default reached its end-of-life on 2026-06-15. New tests run the
|
|
28
|
+
adapter and Claude web search through the installed SDK.
|
|
29
|
+
- **The paper cache is safe across installs**: two copies of the package
|
|
30
|
+
on one machine (a checkout and a venv, say) re-extracted the bundled
|
|
31
|
+
papers over each other, so a copy reading them could see part of the
|
|
32
|
+
corpus and fall back to TF-IDF. Each corpus now has its own folder, named
|
|
33
|
+
after the zip's contents and filled before it is put in place.
|
|
34
|
+
- **Framework examples checked against the corpus**: the five bundled
|
|
35
|
+
examples were replaced; each now summarises one paper in the literature
|
|
36
|
+
corpus, is laid out coupling-first and gives each system's role per flow.
|
|
37
|
+
The constants of the four retired cases (`SOYBEAN_TRADE`, `PANDA_SDGS`,
|
|
38
|
+
`RIFA_INVASION`, `BEIJING_WATER`) are gone from
|
|
39
|
+
`metacouplingllm.knowledge.examples`.
|
|
40
|
+
- **Country and region names read as texts write them**: region names with
|
|
41
|
+
their type ("Narok County"), with a country's name inside ("New South
|
|
42
|
+
Wales") or with other separators, letters or apostrophes resolve; free
|
|
43
|
+
text no longer finds Wales in "New South Wales" or Andorra in "and"; the
|
|
44
|
+
ADM1 alias table grows from 1,145 to 1,256 entries (863 → 922 regions,
|
|
45
|
+
136 → 140 countries); `find_country_codes` is exported from the package.
|
|
46
|
+
- **The adjacency settings reach the prompt and the maps**:
|
|
47
|
+
`de_facto_borders` and `coupling_standard` now also apply to the
|
|
48
|
+
pericoupling hint in the system prompt and to the maps, and the
|
|
49
|
+
`metacouplingllm.visualization` map functions take every keyword of the
|
|
50
|
+
functions they wrap.
|
|
51
|
+
- **Parsing**: a §6 Research Gaps written as a numbered list or as
|
|
52
|
+
paragraphs, a system scope after a dash and a colon inside a bold label
|
|
53
|
+
are parsed, and a flow whose heading names two categories gets the same
|
|
54
|
+
category in every run.
|
|
55
|
+
- **Validation and hints**: a validation note calls the classification
|
|
56
|
+
consistent only when its check classified a pair; the prompt's ADM1 hint
|
|
57
|
+
names a region the description writes; the automatic map's text fallback
|
|
58
|
+
follows the query's framing.
|
|
59
|
+
|
|
60
|
+
### Changed
|
|
61
|
+
|
|
62
|
+
- **The five bundled framework examples were replaced.** Each now summarises
|
|
63
|
+
one paper in the literature corpus and was checked against it: Wolong Nature
|
|
64
|
+
Reserve (Liu 2017), South American food trade (Herzberger et al. 2019),
|
|
65
|
+
global marine fishing (Carlson et al. 2020), groundwater bottling and Michigan
|
|
66
|
+
trout streams (Carlson, Taylor and Hughes 2020) and transatlantic wood pellets
|
|
67
|
+
(Parish et al. 2018). The old set misstated some sources, for example the
|
|
68
|
+
fire-ant case cited Liu et al. (2013) for a USA-to-China invasion the paper
|
|
69
|
+
does not describe. Examples are now laid out coupling-first like the analysis
|
|
70
|
+
they model, give each system's role per flow, and use the template's
|
|
71
|
+
cross-coupling vocabulary; tests check that every source paper is in the
|
|
72
|
+
bundled corpus and that country pericouplings match the adjacency data.
|
|
73
|
+
- **`AnthropicAdapter`'s default model is `claude-sonnet-5-5`.** The old
|
|
74
|
+
default, `claude-sonnet-4-20250514`, reached its end-of-life on
|
|
75
|
+
2026-06-15 according to the anthropic SDK's deprecation table, so an
|
|
76
|
+
adapter built without `model=` called a retired model. The MANUAL and
|
|
77
|
+
INTRODUCTION examples name the new default.
|
|
78
|
+
|
|
79
|
+
### Fixed
|
|
80
|
+
|
|
81
|
+
- **Region names written with their type resolve: "Narok County",
|
|
82
|
+
"Kagoshima Prefecture", "Rangpur Division".** `resolve_adm1_code` returned
|
|
83
|
+
`None` for many names as researchers write them, so adjacency validation
|
|
84
|
+
never engaged for those regions: with the country hint, "Narok County" gave
|
|
85
|
+
`None` while "Narok" gave `KEN033`, and "Rajshahi Division" resolved only
|
|
86
|
+
because the alias table happened to list it, while "Rangpur Division" did
|
|
87
|
+
not. When no strategy matches, the resolver now removes one trailing
|
|
88
|
+
designator and resolves the rest. Region, province, state and oblast count
|
|
89
|
+
in any country, as they already did after a canonical name; county,
|
|
90
|
+
prefecture, division, department, district, governorate and voivodeship
|
|
91
|
+
count only in the countries whose World Bank first-level units carry them,
|
|
92
|
+
so "Washington County" does not become the State of Washington, nor
|
|
93
|
+
"Khulna District" the Khulna Division. The result is refused when the rest
|
|
94
|
+
also names another unit of the same country ("Almaty Province": "Almaty"
|
|
95
|
+
is Almaty City) or matches several regions ("La Paz Department" without a
|
|
96
|
+
country). A name that resolved before is unchanged: of the 183,917
|
|
97
|
+
queries recorded before the change (each World Bank name and alias with
|
|
98
|
+
and without its country hint, alone and followed by a designator, and the
|
|
99
|
+
recorded name-matching queries), none of the 43,883 that resolved gives a
|
|
100
|
+
different code. The alias table gains 18 curated entries, each checked
|
|
101
|
+
against the source its basis cites (1,145 → 1,163 entries, 863 → 872
|
|
102
|
+
regions):
|
|
103
|
+
"Mexico City", "Silesian Voivodeship", "North-West District" (Ngamiland),
|
|
104
|
+
"Barisal", the English names of five more Polish voivodeships and two
|
|
105
|
+
Chilean regions, the full official names of three Chilean regions that the
|
|
106
|
+
World Bank names shorten ("Libertador General Bernardo O'Higgins" for
|
|
107
|
+
"Libertador Gral. Bernardo O'Higgins"), and four spellings the World Bank
|
|
108
|
+
names get wrong ("Lubelskie" for "Lubeiskie", "Małopolskie" for
|
|
109
|
+
"Malopolske", "Arica y Parinacota" for "Arica y Painacota",
|
|
110
|
+
"Zachodniopomorskie" for "Zachodnio-Pomorskie").
|
|
111
|
+
`scripts/validate_adm1_aliases.py` holds the curated list and still
|
|
112
|
+
reproduces the other 1,145 entries byte for byte. Tests +108 (suite
|
|
113
|
+
1,623).
|
|
114
|
+
- **Region names that contain a country's name resolve without a country
|
|
115
|
+
hint: "New South Wales", "British Columbia", "Mexico City".** Without a
|
|
116
|
+
hint, `resolve_adm1_code` returned `None` for every name in which
|
|
117
|
+
`resolve_country_code` found a country, so 52 of the 3,374 regions did not
|
|
118
|
+
resolve under their own World Bank names, some read as another country:
|
|
119
|
+
"New South Wales" and "East New Britain" as the United Kingdom, "Northern
|
|
120
|
+
Ireland" as Ireland, "Port Of Spain" as Spain, "New Jersey" as Jersey,
|
|
121
|
+
"Thai Binh" as Thailand. It now refuses only a name that is a country's ISO
|
|
122
|
+
code, name or alias (what `resolve_country_code` tests with the new
|
|
123
|
+
`substring=False`), alone or among the connector and administrative words
|
|
124
|
+
its matcher ignores ("of Mexico", "State of Mexico", "Niger State");
|
|
125
|
+
without that last part, n-grams of running text such as "region of
|
|
126
|
+
Mexico" (from "the Michoacán region of Mexico") would resolve to the State
|
|
127
|
+
of México. The 33 names with another word besides a country's name
|
|
128
|
+
resolve to their own regions, as do the curated aliases "Mexico City",
|
|
129
|
+
"Lesser Poland", "Greater Poland" and "Magallanes and the Chilean
|
|
130
|
+
Antarctica"; the 19 that are a country's name ("Georgia", "Niger",
|
|
131
|
+
"Wales", "Est", which is Estonia's ISO code) still need the hint, and so
|
|
132
|
+
now do "México" and "Panamá" (next entry). A research description naming
|
|
133
|
+
such a region ("Wool exports from New South Wales") now counts as
|
|
134
|
+
subnational for the validator and the map. Of the 14,421 recorded
|
|
135
|
+
name-matching queries, eight give a different result, each listed with
|
|
136
|
+
its reason in the regression test: "British Columbia", "Western
|
|
137
|
+
Australia", "Ile-de-France", "Île-de-France", "Ciudad de Mexico" and
|
|
138
|
+
"Mexico City" resolve without a hint, and "México" and "Estado de México"
|
|
139
|
+
no longer do. Of the 124,013 phrases with a
|
|
140
|
+
country's name in them among the one- to four-word phrases and capitalised
|
|
141
|
+
word groups of the 421 corpus papers and the avocado run, 23 resolve
|
|
142
|
+
differently without a hint: 22 to the region they name (New Jersey,
|
|
143
|
+
Western Australia, New South Wales and others) and "México" to `None`.
|
|
144
|
+
- **The alias table takes English names that contain a country's name:
|
|
145
|
+
"Upper Austria", "Inner Mongolia", "North Holland", "East Azerbaijan"
|
|
146
|
+
(1,163 → 1,201 entries, 872 → 897 regions, 136 → 138 countries).**
|
|
147
|
+
Validation rule G of `scripts/validate_adm1_aliases.py` dropped every
|
|
148
|
+
generated alias in which `resolve_country_code` found a country, 64 of
|
|
149
|
+
them, because the resolver read such names as the country without a hint
|
|
150
|
+
(entry above). It now drops only a key that is a country's code, name or
|
|
151
|
+
alias, 4 of them ("hong kong", "kuwait", "french guiana", "saint
|
|
152
|
+
martin"). Of the 50 aliases this admitted, each checked against
|
|
153
|
+
Wikipedia's lists of the countries' divisions and the units' own
|
|
154
|
+
articles, 38 ship and 12 went on the review denylist: the capital-city
|
|
155
|
+
namesakes "kuwait city", "luxembourg city" and "panama city"; "french
|
|
156
|
+
riviera", the coast of Provence-Alpes-Côte d'Azur; generic labels that
|
|
157
|
+
span several units, "northern", "eastern" and "western kazakhstan",
|
|
158
|
+
"north lebanon" (also Akkar), "south lebanon" (also Nabatieh) and
|
|
159
|
+
"southwest ethiopia" (the RAG corpus's "southwestern Ethiopia" studies lie
|
|
160
|
+
in Jimma Zone, Oromia); the former name "lower congo" (Bas-Congo, renamed
|
|
161
|
+
Kongo Central in 2015); and the calque "isle of france", once the name of
|
|
162
|
+
colonial Mauritius. Seven others are now redundant (rule R1): with the
|
|
163
|
+
country hint they resolve without an alias ("british columbia", "port of
|
|
164
|
+
spain"). Deduplication (rule D0) now keeps the curated record when a
|
|
165
|
+
generated alias has the same key and code, so "mexico city", "lesser
|
|
166
|
+
poland" and "greater poland", which the generation run had also
|
|
167
|
+
proposed, keep their cited sources in the audit table. The 14,421
|
|
168
|
+
recorded name-matching results are unchanged.
|
|
169
|
+
- **`resolve_country_code` ignores accents and reads the World Bank's
|
|
170
|
+
country names.** "México", "Perú", "Panamá", "Curaçao" and "Réunion"
|
|
171
|
+
returned `None`, as did "España", "Brasil", "Bangladeshi", "Kyrgyz
|
|
172
|
+
Republic", "Slovak Republic" and "Lao People's Democratic Republic", and
|
|
173
|
+
the World Bank's "Congo, Dem. Rep." and "Korea, Dem. People's Rep." gave
|
|
174
|
+
the Republic of the Congo and South Korea. Case, accents, curly
|
|
175
|
+
apostrophes and repeated spaces are now ignored, and a leading "the" when
|
|
176
|
+
the text is a whole name ("the Gambia"). New aliases add the two names in
|
|
177
|
+
the country's own language, "Bangladeshi", and the World Bank's names that
|
|
178
|
+
resolved to `None` or to another country. In its country list
|
|
179
|
+
(api.worldbank.org, read 2026-10-16) 12 of the 217 economies did, and 11
|
|
180
|
+
now resolve, among them "St. Lucia", "Virgin Islands (U.S.)" (read as the
|
|
181
|
+
United States), "St. Martin (French part)" (as France) and "Puerto Rico
|
|
182
|
+
(US)" (next entry); the Channel Islands have no ISO code. In the
|
|
183
|
+
boundary data behind the ADM1 database, "Democratic Republic of Congo" and
|
|
184
|
+
"D. P. R. of Korea" gave COG and KOR, so for their regions web search
|
|
185
|
+
wrote the Republic of the Congo and South Korea into its queries. Tests
|
|
186
|
+
+78 with the two entries above (suite 1,701).
|
|
187
|
+
- **"Puerto Rico" resolves to Puerto Rico, not the United States.** The
|
|
188
|
+
alias table listed "puerto rico" twice, as PRI and further down as USA,
|
|
189
|
+
and the later key wins, so `resolve_country_code` read Puerto Rico as the
|
|
190
|
+
United States, although the country pericoupling table, the ADM1 database
|
|
191
|
+
and the World Bank boundaries of both maps give it its own code, PRI. A
|
|
192
|
+
map with Puerto Rico as its focal country showed the United States, with
|
|
193
|
+
Canada and Mexico pericoupled, and arrows from Puerto Rico started in the
|
|
194
|
+
middle of the contiguous United States; `lookup_pericoupling` called
|
|
195
|
+
Puerto Rico and Mexico pericoupled, and the database hint for a research
|
|
196
|
+
description naming only Puerto Rico gave the United States as its focal
|
|
197
|
+
country, with Canada and Mexico as neighbours. With the country hint
|
|
198
|
+
"Puerto Rico", `resolve_adm1_code` found none of the island's eight regions
|
|
199
|
+
("San Juan" gave `None`, "Montana" the US state), nor did the map's
|
|
200
|
+
region scan of an analysis about Puerto Rico, and web search named their
|
|
201
|
+
country the United States ("Ponce, United States"). "Puerto Rico", the
|
|
202
|
+
World Bank's "Puerto Rico (US)" and its boundary data's "Puerto Rico
|
|
203
|
+
(U.S.)" now resolve to PRI, which the pericoupling table gives no
|
|
204
|
+
pericoupled neighbour. A new test fails on a repeated alias key; the two
|
|
205
|
+
other repeats in the table gave the same code twice.
|
|
206
|
+
- **ISO 3166 names written inverted resolve to their own country.** In
|
|
207
|
+
Debian's iso-codes data (the names pycountry gives) and Wikipedia's table
|
|
208
|
+
of ISO 3166-1, "Korea, Democratic People's Republic of" gave South Korea,
|
|
209
|
+
"Congo, The Democratic Republic of the" and "Congo, Democratic Republic
|
|
210
|
+
of the" the Republic of the Congo, "Virgin Islands, British" and "Virgin
|
|
211
|
+
Islands (British)" the United Kingdom, and "Virgin Islands, U.S." and the
|
|
212
|
+
official name "Virgin Islands of the United States" the United States,
|
|
213
|
+
each through a shorter name inside it. Seven aliases add them. Of the 433
|
|
214
|
+
names of the 249 entries in iso-codes (short, official and common), all
|
|
215
|
+
now give their own code except "Pitcairn" and the official names of
|
|
216
|
+
Gabon, Greece, Rwanda and Togo ("Gabonese Republic"), which give `None`,
|
|
217
|
+
and the names of three places the package has no code for: Antarctica
|
|
218
|
+
gives `None`, British Indian Ocean Territory the United Kingdom and South
|
|
219
|
+
Georgia and the South Sandwich Islands Georgia. Tests +10 with the
|
|
220
|
+
previous entry (suite 1,711).
|
|
221
|
+
- **A word that only begins a region's name no longer resolves to that
|
|
222
|
+
region: "other" is not Other Territories, nor "Puerto" Puerto Plata.**
|
|
223
|
+
Without a country hint, `resolve_adm1_code` took a single word that began
|
|
224
|
+
a longer World Bank name, or that a slash or an accented letter set off
|
|
225
|
+
inside one, for that region. The prompt's ADM1 hint, web search and the
|
|
226
|
+
map's region scans try each word of a description on its own, so "How
|
|
227
|
+
does soybean trade between Brazil and China affect other countries?" got
|
|
228
|
+
Other Territories (Australia) as its focal region, "the European Union"
|
|
229
|
+
La Unión (El Salvador), "Federal subsidies" Distrito Federal (Brazil),
|
|
230
|
+
"the lower Mekong" Lower River (The Gambia), "Coffee exports from Puerto
|
|
231
|
+
Rico" Puerto Plata (Dominican Republic) and "San Juan, Puerto Rico" San
|
|
232
|
+
Juan/Laventille (Trinidad and Tobago), and web search wrote these regions
|
|
233
|
+
into its queries. Of the 51,848 distinct words, as written, of the 421
|
|
234
|
+
corpus papers, 648 resolved without a hint and 177 of them no longer do:
|
|
235
|
+
"other", "area", "under", "long", "cross", "sector", "River", parts of
|
|
236
|
+
longer names ("Minas", "Addis") and given names and surnames ("María",
|
|
237
|
+
"Khan", "John"). A single word now resolves without a hint only as a
|
|
238
|
+
whole name, name variant or alias; with a hint it may still begin a
|
|
239
|
+
longer name ("Badakhshan" with "TJK"). The short names that are a
|
|
240
|
+
region's own resolve through 48 new aliases ("Beijing", "Shanghai",
|
|
241
|
+
"Xinjiang", "Michoacan", "Veracruz", "Madrid", "Tbilisi"): rule R1 of the
|
|
242
|
+
alias build no longer drops as redundant a one-word key that the resolver
|
|
243
|
+
found only as the start of a longer name, which admitted 50 generated
|
|
244
|
+
aliases, and the common English words "northwest", "oriental" and "udon"
|
|
245
|
+
went on the review denylist; "Brussels" is curated. Names written with
|
|
246
|
+
spaces for their hyphens now match ("Schleswig Holstein", "Baden
|
|
247
|
+
Württemberg"), which text scans need, as they split words at hyphens;
|
|
248
|
+
seven aliases this makes redundant are gone ("dar es salaam"). The table
|
|
249
|
+
goes from 1,201 to 1,242 entries (897 → 920 regions, 138 → 140
|
|
250
|
+
countries), and rule C1 caps only a region's generated aliases at three,
|
|
251
|
+
so Magallanes keeps its two curated ones. Of the 14,421 recorded
|
|
252
|
+
name-matching queries, ten change, each listed in the regression test:
|
|
253
|
+
eight words no longer resolve ("other", "Union", "federal") and "Ile de
|
|
254
|
+
France" and "KwaZulu Natal" now do. Tests +30 (suite 1,741).
|
|
255
|
+
- **Region names written with other separators, letters or apostrophes
|
|
256
|
+
resolve: "Trans-Nzoia", "Skåne", "Murang’a".** `resolve_adm1_code` returned
|
|
257
|
+
`None`, even with the country hint, for names that differ from a World Bank
|
|
258
|
+
name or alias only in how their words are joined or spelled: a hyphen, en
|
|
259
|
+
dash or nothing where the World Bank writes something else ("Trans-Nzoia"
|
|
260
|
+
for "Trans Nzoia", "Taita–Taveta", "Bío Bío" for "Biobio", Botswana's
|
|
261
|
+
"North-East District" for "North East"), a curly, other or no apostrophe
|
|
262
|
+
("Murang’a" for "Murang'a", "Valle d’Aosta", "Hawkes Bay" for "Hawke's
|
|
263
|
+
Bay"), and the å, ä, ö, ø and ü that the World Bank spells out in Swedish,
|
|
264
|
+
Norwegian and some German names ("Skåne" for "Skaane Laen", "Thüringen" for
|
|
265
|
+
"Thueringen"). With the hint, the World Bank's multi-word names returned
|
|
266
|
+
`None` when written with hyphens in 769 of 1,038 cases, with en dashes in
|
|
267
|
+
1,209 of 1,211 and with no separator in 1,205 of 1,211, and its 57 names
|
|
268
|
+
with an apostrophe or grave accent in 266 of 280 variants written with ’, ‘,
|
|
269
|
+
ʼ, `` ` `` or none; all now resolve, except names that two units of a
|
|
270
|
+
country share. When neither the strategies nor the designator retry resolve
|
|
271
|
+
a name, the resolver compares it with every name and alias by a spelling key
|
|
272
|
+
that drops spaces, hyphens, dashes and apostrophes and takes those letters
|
|
273
|
+
both as written and spelled out; failing that, it repeats the designator
|
|
274
|
+
retry with the rest compared the same way ("Trans-Nzoia County"). A key must
|
|
275
|
+
identify one region, in the hinted country when a hint is given, so "Nord
|
|
276
|
+
Ouest" without a hint (Haiti's Nord-Ouest and Cameroon's "Nord - Ouest")
|
|
277
|
+
stays `None`, and without a hint a name that is a country's is refused as by
|
|
278
|
+
the strategies. On the 200 research descriptions of the MetacouplingChat
|
|
279
|
+
set, the prompt builder now finds the labelled focal region of "The cherry
|
|
280
|
+
boom in O'Higgins, Chile", where it found none; the other 199, the
|
|
281
|
+
subnational-scope check on all 200 and the validator's ADM1 and flow-map
|
|
282
|
+
results on 63 saved analyses are unchanged. A name that resolved before is
|
|
283
|
+
unchanged: of the 1,440,634 queries recorded on main before the change (each
|
|
284
|
+
World Bank name, name-index key, alias key and alias display with and
|
|
285
|
+
without its country hint, alone and followed by each designator; the
|
|
286
|
+
recorded name-matching queries; and 1,177,308 separator, spelled-letter and
|
|
287
|
+
apostrophe variants of the same names), none of the 213,720 that resolved
|
|
288
|
+
gives a different code. Of the 359,366 that now resolve, 359,356 are
|
|
289
|
+
variants resolving to the region they were made from, and the other 10, dash
|
|
290
|
+
and joined forms of "North West District", are Botswana's North-West
|
|
291
|
+
District. The alias table gains 14 curated entries, each checked against the
|
|
292
|
+
source its basis cites (now 1,256 entries for 922 regions): the Swedish
|
|
293
|
+
counties Halland, Kronoberg, Jönköping, Gävleborg, Örebro and Skåne by name
|
|
294
|
+
alone (the World Bank names keep the Swedish genitive, "Hallands Laen", and
|
|
295
|
+
spell out the letters, so the unaccented "Orebro" missed "Oerebro Laen";
|
|
296
|
+
Stockholm is left out because in research text it mostly means the Stockholm
|
|
297
|
+
Resilience Centre or Convention), the Danish names of the five Danish
|
|
298
|
+
regions ("Region Midtjylland"), "Fryslân", "Hamadan" and "Khanty-Mansi
|
|
299
|
+
Autonomous Okrug". Tests +126 (suite 2,017).
|
|
300
|
+
- **Free text names a country only where it names the country: "New South
|
|
301
|
+
Wales" is not Wales, "South America" not the United States, "and" not
|
|
302
|
+
Andorra.** `resolve_country_code` returns the country of the longest name
|
|
303
|
+
or alias it finds in a text, so the callers that read countries out of
|
|
304
|
+
free text read a place named after a country as that country, and a word
|
|
305
|
+
that is a country code as that country. On the 421 papers of the RAG
|
|
306
|
+
corpus, read as each caller other than web search splits text, 162 to 182
|
|
307
|
+
papers per caller had the first kind of misread:
|
|
308
|
+
- continents: "South America" and "Latin American" as the United States;
|
|
309
|
+
- regions of other countries: "Inner Mongolia" as Mongolia in 30 papers,
|
|
310
|
+
"British Columbia", "New Mexico", "New England";
|
|
311
|
+
- seas and basins: "Indian Ocean", "Gulf of Mexico", "Congo Basin".
|
|
312
|
+
|
|
313
|
+
As web-search queries, 129 of the 421 paper titles took Andorra as their
|
|
314
|
+
national focus through the word "and", and 14 searched for a wrong
|
|
315
|
+
country's exports ("Andorra ... export destinations"). The pronoun "us"
|
|
316
|
+
read as the United States in 83 to 89 papers for each caller that reads
|
|
317
|
+
lower-case words.
|
|
318
|
+
|
|
319
|
+
The new `free_text=True` skips a country's name or alias found only inside
|
|
320
|
+
the name of another place. The other places are the ADM1 database's
|
|
321
|
+
region names and aliases that hold another country's name, and a list of
|
|
322
|
+
continents, seas, lakes, rivers, basins and regions. A code, and the alias
|
|
323
|
+
"us", count only when written in capitals. These callers now read with it:
|
|
324
|
+
- the map's market and partner chunks, the focal-country pick, the system
|
|
325
|
+
roles, and the country after a comma in a region's name ("Sydney, New
|
|
326
|
+
South Wales") in `core`;
|
|
327
|
+
- both world map extractors.
|
|
328
|
+
|
|
329
|
+
The country mentions of an analysis, the prompt's country hint and web
|
|
330
|
+
search's national focus read with `find_country_codes`, which applies the
|
|
331
|
+
same rules (see its entry below).
|
|
332
|
+
|
|
333
|
+
New aliases cover names a shorter alias read as another country: "North
|
|
334
|
+
Korean" and "DPR Korea" (South Korea), "South Sudanese" (Sudan) and
|
|
335
|
+
"Guinea Bissau" (Guinea). "Lao" is added too: free text now reads the code
|
|
336
|
+
LAO only in capitals. With it, Vietnam's "Lao Cai" holds a country's name,
|
|
337
|
+
like "Thai Binh"; it still resolves to its own region with or without a
|
|
338
|
+
hint.
|
|
339
|
+
|
|
340
|
+
On the same papers, no chunk caller loses a correct reading, and the
|
|
341
|
+
country mentions' wrong countries summed over papers fall from 266 to 12.
|
|
342
|
+
In the committed avocado run, "Central American receiving markets" no
|
|
343
|
+
longer reads as the United States, and no output changes. Of the 14 title
|
|
344
|
+
queries, none now names a wrong country. Still read as countries:
|
|
345
|
+
- homonyms (Chad Wong, Panama disease, "Indian" for Indigenous peoples);
|
|
346
|
+
- acronyms written as codes (LCA).
|
|
347
|
+
|
|
348
|
+
The alias build still reproduces both alias tables byte for byte. Tests
|
|
349
|
+
+73 (suite 1,814).
|
|
350
|
+
- **A §6 Research Gaps written as a numbered list is parsed.**
|
|
351
|
+
`parse_analysis` read §6 with the bullet extractor, which takes only dash,
|
|
352
|
+
asterisk and `•` items, so a §6 written as `1. **Label**: ...` gave an empty
|
|
353
|
+
`research_gaps`: the full report and the Markdown and Word exports left §6
|
|
354
|
+
out, the summary, comparison and component views reported no gaps, and the
|
|
355
|
+
countries and terms it named were missing from every scan of the analysis
|
|
356
|
+
text (the maps' text-scan fallback, ADM1 resolution, `recommend_papers()`).
|
|
357
|
+
§6 now also takes numbered items (`1.` or `1)`); the other sections still
|
|
358
|
+
take bullets only, since a numbered line there can be a heading such as
|
|
359
|
+
`1. **Matter Flow**`. Of 31 gpt-5.5 analyses saved on 2026-10-15, 7 wrote §6
|
|
360
|
+
as a numbered list; each now yields its 6 to 8 gaps, and the other 24 parse
|
|
361
|
+
exactly as before. Tests +4 (suite 1,454; the 1,441 in REPRODUCING predated
|
|
362
|
+
the nine tests #167 added).
|
|
363
|
+
- **A system whose scope follows a dash is parsed.** `parse_analysis` read a
|
|
364
|
+
system's scope only in parentheses, as the output template writes it
|
|
365
|
+
(`**Receiving System (distant)**: Nigeria`), so a heading such as
|
|
366
|
+
`**Receiving System — distant**: Nigeria` was not read as one: the system was
|
|
367
|
+
lost, its lines dropped or, after another system's heading in the same block,
|
|
368
|
+
merged into that system, overwriting its fields. The full report and the
|
|
369
|
+
Markdown and Word exports left it out, the maps' sending, receiving and
|
|
370
|
+
adjacent country lists and the country-level pericoupling check against the
|
|
371
|
+
coupling database missed its countries, and its text was missing from the
|
|
372
|
+
scans of the parsed analysis (ADM1 resolution, `recommend_papers()`). System
|
|
373
|
+
headings, and the flat-bullet fallback, now also take a scope after an em
|
|
374
|
+
dash, en dash or hyphen when it starts with `adjacent` or `distant`, and read
|
|
375
|
+
it as a scope in parentheses is read. Of 31 gpt-5.5 analyses saved on
|
|
376
|
+
2026-10-15, 4 wrote 18 such headings, each ahead of any other system heading
|
|
377
|
+
in its block; each of the 4 now gains its 3 to 6 systems, and everything else
|
|
378
|
+
in the 31 parses exactly as before. Tests +9 (suite 1,463).
|
|
379
|
+
- **A §6 Research Gaps written as paragraphs is parsed.** `parse_analysis`
|
|
380
|
+
took only §6's list items (dash, asterisk, `•` or numbered), so a §6 written
|
|
381
|
+
as plain paragraphs or as `**1. Label**` lines gave an empty
|
|
382
|
+
`research_gaps`: the full report and the Markdown and Word exports left §6
|
|
383
|
+
out without notice, and the scans of the analysis text (the maps' text-scan
|
|
384
|
+
fallback, ADM1 resolution, `recommend_papers()`) missed it. A §6 without
|
|
385
|
+
list items is now read as paragraphs: each paragraph is a gap, a paragraph
|
|
386
|
+
that is only a bold label or a heading opens the paragraph after it, a
|
|
387
|
+
number at the start of a bold label is dropped as a list marker is, and
|
|
388
|
+
rules (`---`, `═══`) are skipped. A §6 with list items still yields only
|
|
389
|
+
them, except that a spaced rule such as `- - -`, which read as a dash item,
|
|
390
|
+
no longer does. None of the 31 gpt-5.5 analyses saved on 2026-10-15 writes
|
|
391
|
+
§6 without list items, and all 31 yield the same gaps as before.
|
|
392
|
+
- **A system scope that says more than "adjacent" or "distant" is read as
|
|
393
|
+
that word.** `parse_analysis` stored a qualifier such as `(adjacent
|
|
394
|
+
interaction)` whole as the system's `system_scope`, which `core.py` and the
|
|
395
|
+
maps compare with "adjacent", so a system headed `**Sending System (adjacent
|
|
396
|
+
interaction)**: Mexico` was not treated as adjacent: the maps put Mexico in
|
|
397
|
+
the sending role, and a flow from "neighboring producers" got no map arrow.
|
|
398
|
+
The scope is now the qualifier's leading "adjacent" or "distant", and a
|
|
399
|
+
qualifier that says more is kept whole as `system_scope_qualifier`, which
|
|
400
|
+
the report's system label shows as before (`Sending System (adjacent
|
|
401
|
+
interaction)`); a qualifier that starts with neither word ("regional", "not
|
|
402
|
+
adjacent") is the scope as written. Of the 31 saved analyses, 4 head 8
|
|
403
|
+
systems with such a qualifier, 7 distant (`distant coal importer`) and 1
|
|
404
|
+
adjacent (`adjacent interaction`); their scopes now read "distant" or
|
|
405
|
+
"adjacent", and their reports, exports, map roles, map arrows, coupling
|
|
406
|
+
checks and paper-search terms are unchanged (old/D12_r0's Mexico, the
|
|
407
|
+
adjacent case, was already adjacent on its maps through another system).
|
|
408
|
+
- **A label with its colon inside the bold is parsed.** `**Focal System:**
|
|
409
|
+
Michigan` gave the system name `** Michigan`, and `- **Human subsystem:**
|
|
410
|
+
Farmers` under it the value `** Farmers`. A flow heading such as `**Matter
|
|
411
|
+
Flow:**` was not read as one, so its flow lost its category and split in
|
|
412
|
+
two, and `- **Direction:** Brazil → China` gave the direction `** Brazil →
|
|
413
|
+
China`. An agent written `[Individuals / households] **Farmers:** ...` got
|
|
414
|
+
the name `**Farmers` and a description starting with `**`, and a cause or
|
|
415
|
+
effect heading `**Economic:**` opened a category `economic:` apart from
|
|
416
|
+
`economic`. System headings and the lines under them, the flat-bullet
|
|
417
|
+
systems fallback, flow headings and their `Direction` and `Description`
|
|
418
|
+
lines, the one-line flow bullets (`[Matter] **Brazil → China:** ...`,
|
|
419
|
+
`**Category:** Matter | **Direction:** ...`), agents, and cause and effect
|
|
420
|
+
headings now read the same with the colon inside the bold as after it.
|
|
421
|
+
Free text keeps its colons as written, such as an effect item `**Mexico:**
|
|
422
|
+
...`. None of the 31 saved analyses writes a colon inside the bold. Tests
|
|
423
|
+
+52 with the two entries above (suite 2,069).
|
|
424
|
+
- **The adjacency settings reach the prompt hint and the maps.**
|
|
425
|
+
`MetacouplingAssistant` passed `de_facto_borders` and `coupling_standard`
|
|
426
|
+
only to the post-LLM validation; the pericoupling-database hint in the
|
|
427
|
+
system prompt and the automatic map called the adjacency lookups with the
|
|
428
|
+
defaults, so with other settings the hint could contradict the validation.
|
|
429
|
+
With `coupling_standard="lenient"`, a description of acid rain from Ohio
|
|
430
|
+
falling on Ontario got a hint listing only Ohio's five US neighbours, while
|
|
431
|
+
the validation counted Ontario, across Lake Erie, as a cross-border
|
|
432
|
+
neighbour. `PromptBuilder`, `plot_focal_country_map`, `plot_analysis_map`
|
|
433
|
+
and `plot_focal_adm1_map` (and their `metacouplingllm.visualization`
|
|
434
|
+
wrappers) now take both settings as keyword arguments, and the assistant
|
|
435
|
+
passes its own to the hint, the map and the map's same-country flow arrows.
|
|
436
|
+
The defaults are the lookups' own: with them, the system prompt for each of
|
|
437
|
+
212 research descriptions (built directly and by a default assistant) and
|
|
438
|
+
the map classifications for every country (249, both country maps) and
|
|
439
|
+
every ADM1 region (3,374) are byte-identical to before. Tests +43 (suite
|
|
440
|
+
1,506).
|
|
441
|
+
- **The `metacouplingllm.visualization` map functions take every keyword of
|
|
442
|
+
the functions they wrap.** These wrappers of `visualization.worldmap` and
|
|
443
|
+
`visualization.adm1_map` listed only some of their keywords and raised
|
|
444
|
+
`TypeError` on the rest: `adm0_shapefile` and `mentioned_countries` for
|
|
445
|
+
`plot_focal_country_map`; `adm0_shapefile`, `extra_mentioned_countries`,
|
|
446
|
+
`mentioned_countries_override`, `focal_code_override` and `flows` for
|
|
447
|
+
`plot_analysis_map`; `mentioned_countries`, `mentioned_adm1_codes` and
|
|
448
|
+
`flows` for `plot_focal_adm1_map`. The top-level `metacouplingllm.plot_*`
|
|
449
|
+
functions pass `**kwargs` through and were not affected, nor were MANUAL's
|
|
450
|
+
examples, which import those or the implementations. The wrappers now have
|
|
451
|
+
the signatures of the functions they wrap; tests compare the two and check
|
|
452
|
+
that every keyword reaches the implementation unchanged through both
|
|
453
|
+
wrappers. Tests +9 (suite 1,515).
|
|
454
|
+
- **The pericoupling hint finds every country a research description names, in
|
|
455
|
+
the order it names them.** When no ADM1 region is found, the system prompt's
|
|
456
|
+
pericoupling lookup pairs the first country the description names (the focal
|
|
457
|
+
country) with each other one, or lists its neighbours when it is the only one
|
|
458
|
+
(`PromptBuilder._build_coupling_hint`). It read the description in
|
|
459
|
+
consecutive chunks of up to four words, one country per chunk, so a name
|
|
460
|
+
across two chunks was missed ("Coffee production in Costa Rica" gave no
|
|
461
|
+
hint), a second name in a chunk was dropped ("Mexico and Canada avocado
|
|
462
|
+
trade" gave Canada alone, as the focal country), names with a hyphen,
|
|
463
|
+
apostrophe or diacritic did not match (Guinea-Bissau gave Guinea), and a word
|
|
464
|
+
alone in its chunk could be read as an ISO code ("between 2000 and 2024" gave
|
|
465
|
+
Andorra). The scan now takes, at each word, the longest run of words that is
|
|
466
|
+
exactly a country's name or alias, then goes on after it: names are read
|
|
467
|
+
whole, without the shorter names inside them ("Papua New Guinea" gives no
|
|
468
|
+
Guinea), names such as "Côte d’Ivoire", Türkiye and "U.S." match, and
|
|
469
|
+
countries keep the order of first mention. "US" counts only in capitals, and
|
|
470
|
+
a three-letter code only in capitals and set apart by punctuation, as in
|
|
471
|
+
"(BRA)", so the pronoun "us", acronyms such as LCA and words such as "and"
|
|
472
|
+
are not read as countries. The scan is `find_country_codes` (next entry),
|
|
473
|
+
which also reads the names of other places, such as "South America", as no
|
|
474
|
+
country. On 200 research descriptions written for testing, 55 countries the
|
|
475
|
+
descriptions name are added and 31 misreads dropped: 13 words read as
|
|
476
|
+
codes, mostly "and", 14 regions of other countries such as British Columbia
|
|
477
|
+
and New Jersey. The first country found is the labelled focal country in
|
|
478
|
+
107 of the 196 labelled descriptions, against 101. On the 9,713
|
|
479
|
+
non-overlapping chunks of up to 250 words that the retrieval chunker cuts
|
|
480
|
+
from the 421 bundled papers:
|
|
481
|
+
- 495 countries the chunks name are added;
|
|
482
|
+
- 2,855 misreads are dropped: words read as codes 2,209 (mostly "and"),
|
|
483
|
+
continents 301, the pronoun "us" 164, regions of other countries 105,
|
|
484
|
+
seas and basins 45, and 31 cut names, homonyms and names without an
|
|
485
|
+
alias of their own;
|
|
486
|
+
- so are three acronyms LCA (Saint Lucia's code), and one Puerto Rico,
|
|
487
|
+
which the old alias read as the United States;
|
|
488
|
+
- three added readings are a given name: "Germán" read as German.
|
|
489
|
+
|
|
490
|
+
Tests +29 (suite 1,843).
|
|
491
|
+
- **`find_country_codes` reads every country a text names; web search's
|
|
492
|
+
national focus and an analysis's country mentions use it too.** The scan of
|
|
493
|
+
the entry above moved to `knowledge/countries.py` as
|
|
494
|
+
`find_country_codes(text, *, capitals=False)`, exported from the package.
|
|
495
|
+
It applies the free-text rules of `resolve_country_code`:
|
|
496
|
+
- the name of another place uses up its words and names only the place's
|
|
497
|
+
own country: "South America", "the Gulf of Mexico", "New South Wales",
|
|
498
|
+
"the Congo basin" and the pau-brasil tree name none, "Inner Mongolia,
|
|
499
|
+
China" names China;
|
|
500
|
+
- names are compared as `resolve_country_code` compares them, so accents
|
|
501
|
+
fold ("México", "Perú");
|
|
502
|
+
- up to three punctuation marks may complete a name ("U.S.", "Virgin
|
|
503
|
+
Islands (U.S.)").
|
|
504
|
+
|
|
505
|
+
Scanned alone, each country name, alias and place name gives what
|
|
506
|
+
`resolve_country_code(..., free_text=True)` gives.
|
|
507
|
+
|
|
508
|
+
Web search's national focus is now the first country a query names. It
|
|
509
|
+
was read from four-word phrases of the query, then shorter ones down to
|
|
510
|
+
single words, which took:
|
|
511
|
+
- the longest name in the first phrase holding one ("Mexico and Canada
|
|
512
|
+
avocado trade" focused on Canada);
|
|
513
|
+
- pieces of other places' names ("Soybean exports from South America" on
|
|
514
|
+
the United States, "Salmon bound for British Columbia ports" on the
|
|
515
|
+
United Kingdom).
|
|
516
|
+
|
|
517
|
+
On the 200 research descriptions the focus is wrong for none, against 12,
|
|
518
|
+
and is the labelled focal country for 107 of 196, against 104. On the 421
|
|
519
|
+
paper titles as queries it is wrong for none, against 8.
|
|
520
|
+
|
|
521
|
+
An analysis's country mentions are read with `capitals=True`, a name
|
|
522
|
+
counting only where it starts with a capital letter. They serve the ADM1
|
|
523
|
+
relevance guard, the country set of the region scan and the ADM1 code list
|
|
524
|
+
of the map-extraction prompt. They were capitalised word groups, which a
|
|
525
|
+
lower-case word, a hyphen or an apostrophe ended: "Trinidad and Tobago",
|
|
526
|
+
"Côte d'Ivoire" and "Timor-Leste" named no country, "the Congo basin" named
|
|
527
|
+
the Republic of the Congo and "North-American" the United States. On the
|
|
528
|
+
421 papers, 92 countries the text names are added, and misread countries
|
|
529
|
+
fall from 10 to 6, all homonyms such as "Panama disease". The avocado run's
|
|
530
|
+
mentions are unchanged.
|
|
531
|
+
|
|
532
|
+
Tests +35 (suite 1,878).
|
|
533
|
+
- **A system's scope that opens with its country's name reads as the
|
|
534
|
+
country, not as the country's namesake region.** To find a focal region in
|
|
535
|
+
an analysis, `_resolve_adm1_from_analysis` reads each system's scope as
|
|
536
|
+
"region, …, country", the last comma part giving the country hint. When
|
|
537
|
+
the first part was the hinted country itself, it resolved to the region of
|
|
538
|
+
that name: "Mexico, especially Michoacán and Jalisco, Mexico" gave the
|
|
539
|
+
State of México (MEX015), and scopes opening with Guatemala, Panama or
|
|
540
|
+
Luxembourg gave their namesake regions. Thirteen countries' names resolve
|
|
541
|
+
to one of their own regions when the country is the hint. The first part
|
|
542
|
+
is no longer read as a region when it is only the hinted country's name.
|
|
543
|
+
"Michoacán, Mexico", "State of Mexico, Mexico", "Panamá Province, Panama"
|
|
544
|
+
and "Georgia, United States" resolve as before; a namesake region written
|
|
545
|
+
by its bare name with its country ("Panamá, Panama") no longer does.
|
|
546
|
+
|
|
547
|
+
The committed avocado run's focal scope opens "Mexico, especially the
|
|
548
|
+
Michoacán Avocado Belt" and ends "while Data México reports … US$3.525
|
|
549
|
+
billion". That last part used to read as the United States (from "US$"),
|
|
550
|
+
so the scope gave no region; since `resolve_country_code` ignores accents
|
|
551
|
+
(above), it reads Mexico, which made the scope the State of México. It now
|
|
552
|
+
gives no region, and the run's callers of the country and region resolvers
|
|
553
|
+
return what they did before this release's country changes. Tests +10
|
|
554
|
+
(suite 1,888).
|
|
555
|
+
- **The automatic map's text fallback follows the query's framing.** When the
|
|
556
|
+
map-data extraction fails or returns no focal country, `_generate_map`
|
|
557
|
+
reads the map from the analysis text, and its first attempt drew an ADM1
|
|
558
|
+
map whenever it could resolve a region there. A query about a country
|
|
559
|
+
("Impact of avocado production and trade in Mexico on sustainability")
|
|
560
|
+
whose analysis gives the scope "Michoacán, Mexico" got a map of Michoacán,
|
|
561
|
+
while with map data the same query gets the country map. The attempt now
|
|
562
|
+
drops the region when the query names none, as the map-data path does and
|
|
563
|
+
as MANUAL §10 describes. A query that names a region ("Avocado production
|
|
564
|
+
in Michoacán, Mexico") still gets the ADM1 map, and so does a direct
|
|
565
|
+
`_generate_map` call made without a query. Tests +3 (suite 1,891).
|
|
566
|
+
- **A pericoupling validation note no longer calls the classification
|
|
567
|
+
consistent when its check classified no pair.** After each analysis the
|
|
568
|
+
country check and, when the research description names a subnational
|
|
569
|
+
region, the ADM1 check compare the classification (§1) with the pairs
|
|
570
|
+
they look up in the adjacency databases, and each writes a note. A check
|
|
571
|
+
that looked up no pair still noted that the classification was
|
|
572
|
+
consistent with its database. With Thailand as the sending system and
|
|
573
|
+
"Nigeria and China" as the spillover system, the country check skips
|
|
574
|
+
both partners, since spillover partners are not checked, and its note
|
|
575
|
+
read "LLM classification is consistent with the coupling database." An
|
|
576
|
+
analysis of acid rain from Ohio falling on Ontario that names Ontario
|
|
577
|
+
only in a system's name and scope, which the region scan skips, got two
|
|
578
|
+
notes that contradict each other under the default
|
|
579
|
+
`coupling_standard="moderate"`: the country check's "Ohio (USA036) ↔
|
|
580
|
+
Canada (CAN): TELECOUPLED … Consider revising." and the ADM1 check's
|
|
581
|
+
"LLM classification is consistent with the ADM1 pericoupling database.",
|
|
582
|
+
although it had classified no region (under `"lenient"` the two notes
|
|
583
|
+
agreed, and the ADM1 one was just as unchecked). A check that looked up
|
|
584
|
+
no pair now notes that nothing was checked: "No partner country was
|
|
585
|
+
checked against the coupling database (partners in the spillover role
|
|
586
|
+
are not checked)." or "No subnational region other than the focal region
|
|
587
|
+
was checked against the ADM1 pericoupling database." The report
|
|
588
|
+
(`result.formatted`) prints such a note only when it lists no pair, so it
|
|
589
|
+
never joins the line that calls the classification consistent with both
|
|
590
|
+
databases. The note of a check that looked up a pair is unchanged; MANUAL
|
|
591
|
+
§7 and §8 say when each note appears. On the 31 gpt-5.5 analyses saved on
|
|
592
|
+
2026-10-15 and the committed avocado run, each validated under the three
|
|
593
|
+
coupling standards and both `de_facto_borders` settings (192 reports),
|
|
594
|
+
every check that looked up a pair keeps its note and 180 validation
|
|
595
|
+
blocks are byte-identical. The other 12 come from the two analyses of
|
|
596
|
+
Indonesia's coal exports to China and India, whose ADM1 check finds no
|
|
597
|
+
region besides East Kalimantan: their Note now reads "LLM classification
|
|
598
|
+
is consistent with the coupling database." instead of claiming
|
|
599
|
+
consistency with "the coupling databases (country adjacency + ADM1
|
|
600
|
+
adjacency)". Tests +35 (suite 2,104).
|
|
601
|
+
- **A flow whose heading names two categories gets the same category in
|
|
602
|
+
every run.** `parse_analysis` reads a flow's category from its heading
|
|
603
|
+
(`**Matter Flow**`), its bracket tag (`[Matter]`), its `category:` field
|
|
604
|
+
or, failing those, the bullet itself. The six category names were kept in
|
|
605
|
+
a set and the first one found in the set's order won, so when the text
|
|
606
|
+
named two, the category depended on the hash seed, which Python draws
|
|
607
|
+
anew for each process unless `PYTHONHASHSEED` is set. In one of the 31
|
|
608
|
+
gpt-5.5 analyses saved on 2026-10-15 (Germany's electricity imports from
|
|
609
|
+
France and Poland and LNG from the USA), the flow headed
|
|
610
|
+
`**Energy and Matter Flow**` was energy under 11 of 16 seeds and matter
|
|
611
|
+
under the other 5. The category labels the flow in the report
|
|
612
|
+
(`[Energy]`), files it under the Markdown and Word exports' per-category
|
|
613
|
+
Flows subsections, colours its arrow on maps drawn from the parsed flows
|
|
614
|
+
and enters the map-extraction prompt. A category's own name now counts
|
|
615
|
+
before a word that only suggests one (goods, money, power …), and among
|
|
616
|
+
several the one the text names first wins: "Energy and Matter Flow" is
|
|
617
|
+
energy, "Matter and Energy Flow" matter. Among the words that only
|
|
618
|
+
suggest a category, the first named also wins where the alias table's
|
|
619
|
+
order decided before ("Labour and money" was capital and is now people).
|
|
620
|
+
On the 31 analyses and the committed avocado run (531 flows, all under
|
|
621
|
+
bold headings), each parsed under 16 seeds, every parse is now the same
|
|
622
|
+
under all 16 and matches the old parse field for field, except that one
|
|
623
|
+
flow under the 5 seeds that read it as matter; none of the 16 distinct
|
|
624
|
+
flow headings names two categories through such words alone. Tests +13
|
|
625
|
+
(suite 2,117).
|
|
626
|
+
- **The prompt's ADM1 hint names a region the description writes, not one
|
|
627
|
+
that a common word or another place's name happens to match.**
|
|
628
|
+
`PromptBuilder._build_adm1_pericoupling_hint` took as the focal region the
|
|
629
|
+
first run of up to five words that `resolve_adm1_code` resolved, in any
|
|
630
|
+
case and without a country, and listed that region's neighbours. Common
|
|
631
|
+
words, pieces of country names and the names of lakes, rivers and bays
|
|
632
|
+
matched regions elsewhere: "has" gave Has (Albania), "forest" Forest
|
|
633
|
+
(Guernsey), "New Zealand" Zealand (Denmark), "Guinea-Bissau" the Sector
|
|
634
|
+
Autonomo De Bissau, "Mutsu Bay" Bay (Somalia), "Okavango Delta" Delta
|
|
635
|
+
(Nigeria), "Lake Tanganyika" Tanganyika province (DR Congo), "Mary River
|
|
636
|
+
mine" Mary (Turkmenistan) and "Southeast Alaska" Southeast (North
|
|
637
|
+
Macedonia); Ghana's "Western Region" gave Bahrain's, and a town, "Santa
|
|
638
|
+
Rosa", in a description of Ecuador gave Guatemala's department. As the
|
|
639
|
+
ADM1 hint takes the place of the country hint, such a match also dropped
|
|
640
|
+
the country-pair reference. A run of words now names a region only when it
|
|
641
|
+
is written as a name (capitalised, with lower-case connectors such as "do"
|
|
642
|
+
or "es" inside and a unit word such as "oblast" at the end), is not part of
|
|
643
|
+
a country's name or of a feature's ("Ohio River", "Lake Tanganyika", "Gulf
|
|
644
|
+
of California"), and, for a lone compass or feature word ("Bay", "West",
|
|
645
|
+
"Delta"), when the text names the region's country. A name several regions
|
|
646
|
+
share is read in the countries the text names, then in those of the other
|
|
647
|
+
regions it names ("Central Region" in Ghana, "Punjab" beside Kerala), and
|
|
648
|
+
Vietnamese letters are read ("Bình Phước"). The focal region is the first
|
|
649
|
+
one named whose country the text names, another named region shares, or
|
|
650
|
+
borders one of those, so a region of an unrelated country, usually a
|
|
651
|
+
namesake, is passed over. A country's name that a region elsewhere also
|
|
652
|
+
has is read as the one the rest of the text supports more: "Georgia" is the
|
|
653
|
+
US state in "Georgia, United States" or beside Alabama or Florida (which
|
|
654
|
+
Uruguay also has), and the country beside Tbilisi, Russia or Armenia;
|
|
655
|
+
"Gaza" is Mozambique's province beside Maputo and Palestine beside Egypt. A
|
|
656
|
+
tie keeps the country, and the country hint leaves out a name read as a
|
|
657
|
+
region: until now "Georgia, United States" got the hint for the country
|
|
658
|
+
Georgia. On 200 research descriptions with known focal units, the hint
|
|
659
|
+
names the right region for 116 of the 127 about a province (99 before), a
|
|
660
|
+
wrong one for 9 (26) and none for 2 (2), England and Scotland, which are
|
|
661
|
+
read as the United Kingdom; the 9 are regions the text names before its
|
|
662
|
+
focal one (Bogotá before Cundinamarca). Of the 73 descriptions about a
|
|
663
|
+
country or another place, 33 get a regional hint (47 before), each for a
|
|
664
|
+
region the text names. Tests +25 (suite 2,142).
|
|
665
|
+
- **A country's name that a region elsewhere also has is read as the run
|
|
666
|
+
reads it, in the validation, the report and the maps.** The prompt hint
|
|
667
|
+
read "Georgia" as the US state where the description supported it, but
|
|
668
|
+
the rest of a run resolved each system name on its own, so "Georgia" was
|
|
669
|
+
always the country. For a description of broiler chicken production in
|
|
670
|
+
Georgia, United States, with an answer naming Georgia, Alabama, Iowa and
|
|
671
|
+
China, the country check took Georgia (GEO) as the focal country and
|
|
672
|
+
checked Georgia (GEO) ↔ China, the ADM1 check checked nothing, and a map
|
|
673
|
+
drawn from the analysis centred on the Caucasus. Inside a
|
|
674
|
+
`knowledge.countries.read_names_as_regions(readings)` block, each name in
|
|
675
|
+
*readings* now reads as a region elsewhere: `resolve_country_code` and
|
|
676
|
+
`find_country_codes` give the region's country, `resolve_adm1_code` the
|
|
677
|
+
region; outside a block nothing changes. `knowledge.places.place_readings`
|
|
678
|
+
gives those readings for a text by the hint's rule, and `read_places`,
|
|
679
|
+
which moved there from the prompt builder, the countries and regions the
|
|
680
|
+
text names. `MetacouplingAssistant` builds each result inside a block read
|
|
681
|
+
from the research description and the answer and runs web search inside
|
|
682
|
+
one read from the description, and `plot_analysis_map` called on its own
|
|
683
|
+
reads the analysis's text. In the example the country check now takes the
|
|
684
|
+
United States as the focal country, the ADM1 check takes Georgia (USA011),
|
|
685
|
+
with Alabama pericoupled and Iowa telecoupled, and the map's focal country
|
|
686
|
+
is the United States; a description of wine exports from Georgia to Russia
|
|
687
|
+
keeps the country. Built through `analyze()` with a stub client, the
|
|
688
|
+
reports of the 31 gpt-5.5 analyses saved on 2026-10-15 and of the
|
|
689
|
+
committed avocado run are byte-identical to main's: none names such a
|
|
690
|
+
country. Tests +14 (suite 2,156).
|
|
691
|
+
- **The Claude adapter works with anthropic 2.x.** anthropic 2.0.0
|
|
692
|
+
(2026-10-21) targets API version 2026-10-21, which removed `temperature`,
|
|
693
|
+
`top_p` and `top_k`: a request that sends one gets a 400 on every model,
|
|
694
|
+
and the SDK's `messages.create` and `messages.stream` no longer take them.
|
|
695
|
+
`AnthropicAdapter` sent `temperature` for every model but
|
|
696
|
+
`claude-opus-4-7`, so with anthropic 2.x every `chat` call raised
|
|
697
|
+
`TypeError`. The adapter now sends `temperature` only when the installed
|
|
698
|
+
SDK's `messages.create` takes it, and never for Claude 4.7 and later
|
|
699
|
+
models, which accept only the default under the older API version as
|
|
700
|
+
well (`claude-opus-4-8` and the Claude 5 models join `claude-opus-4-7`).
|
|
701
|
+
With anthropic 2.x a Claude request samples at the API's defaults
|
|
702
|
+
(temperature 1.0), so `temperature=` has no effect on Claude. The
|
|
703
|
+
web-search backend never sent `temperature` and was not affected. The
|
|
704
|
+
other Anthropic tests use stand-in clients, so none of them noticed; new
|
|
705
|
+
tests drive the adapter and the web-search backend through the installed
|
|
706
|
+
SDK over a mock HTTP transport (no network, no key), and they fail on the
|
|
707
|
+
old adapter with anthropic 2.3.0. Tests +10 (suite 2,166).
|
|
708
|
+
- **Two copies of the package no longer clash over the paper cache.**
|
|
709
|
+
Every copy on a machine (a checkout, a worktree, an installed wheel)
|
|
710
|
+
extracts the bundled `Papers.zip` into `<tempdir>/metacoupling_rag_cache`.
|
|
711
|
+
The extraction counted as current only while its stamp matched the zip's
|
|
712
|
+
mtime, and each copy's zip has its own, so a second copy deleted the
|
|
713
|
+
folder and extracted it again in place, under any copy reading it. In a
|
|
714
|
+
fresh venv run beside a test suite, the RAG engine found 324 of the 421
|
|
715
|
+
papers (8,193 chunks), so the precomputed embeddings no longer matched
|
|
716
|
+
and it fell back to TF-IDF. The extraction now goes to a folder named
|
|
717
|
+
after the zip's SHA-256, so copies whose zips have the same contents
|
|
718
|
+
share it and a different zip gets its own; a new extraction is written
|
|
719
|
+
to a private staging folder and renamed into place only when complete,
|
|
720
|
+
and when another process gets there first, its folder is used. A folder
|
|
721
|
+
that has lost files is replaced, not repaired in place. The `Papers`
|
|
722
|
+
folder and stamp of earlier versions are left alone, so an older install
|
|
723
|
+
keeps working beside a new one. Tests +6 (suite 2,172).
|
|
724
|
+
|
|
725
|
+
### Docs
|
|
726
|
+
|
|
727
|
+
- **`plot_analysis_map`'s docstring documents `mentioned_countries_override`
|
|
728
|
+
and `focal_code_override`.** The two work only together: given both, the
|
|
729
|
+
map's focal country and colored countries are the ISO codes passed, used
|
|
730
|
+
as given, instead of those found in the parsed analysis, and `focal_role`
|
|
731
|
+
and `extra_mentioned_countries` are ignored; the automatic map passes both.
|
|
732
|
+
The docstring also gives the default title as
|
|
733
|
+
`"Metacoupling Analysis Map: {Country Name}"`, not
|
|
734
|
+
`"Metacoupling Analysis Map"`. No code change.
|
|
735
|
+
|
|
10
736
|
## [0.2.1] — 2026-10-12
|
|
11
737
|
|
|
12
738
|
A maintenance release after the documentation audit of 0.2.0.
|