[Masterleeaus]

Summary

• Adds a bounded, offline-reproducible science slice with typed evidence, two hypotheses, an exploratory falsification/sensitivity check, uncertainty, and a report.
• Pins a NASA GISTEMP v4 snapshot and reproduces a 1880–2025 trend: OLS 0.0834 °C/decade (95% moving-block bootstrap interval 0.0676–0.0980); Mann–Kendall tau-b 0.7347, p=3.44e-39. Omitting 2016–2025 still leaves a positive 1880–2015 trend. This is one descriptive series, not causal attribution or independent replication.
• Adds a 60-case synthetic evidence-reasoning benchmark and provider-pluggable crew/single-LLM comparison harness. CI uses mock adapters to test the harness, so it makes no real model-performance claim.
• Adds focused network-blocked pytest tests and a minimal CI workflow.
• Removes retail and finance/ESG modules from the default API/frontend and preserves those prototypes under examples/.
• Removes the unsafe ClimateX archive-unpack workflow; documents the lineage and limited license boundary; removes large binaries from the current tree and adds a manual release workflow using the pre-cleanup snapshot.

Verification

• python -m pytest -q tests/science — 9 passed
• python analysis/reproduce_gistemp.py --check — passed
• python -m compileall -q analysis src/science eval/science_quality tests/science examples/retail_analytics/src examples/climate_finance/src — passed
• Release archive extraction was tested locally against the configured source commit and excluded the zero-byte PDF.

Limits and follow-up

• No live Climate Crew vs single-LLM evaluation has been run; provider calls require credentials and may incur cost.
• The exploratory endpoint check is not preregistered; the Mann–Kendall p-value does not adjust for serial autocorrelation.
• The historical asset release still needs the manual workflow dispatched from main after merge and then verified. Original blobs remain in Git history.
• The source archive URL and original license remain unknown; the new MIT license covers original contributions only.