Aspegio/nelson v2.2.4

v2.2.4 — Learned Standing Orders

on GitHub

Painting by Salvatore Colacicco - Pair of marine paintings of Royal Navy warships

This release adds a Darwin Gödel Machine-inspired pipeline that learns new standing orders from accumulated mission data, with human review before any candidate is promoted.

Features

• Learned standing orders pipeline (#87, #120) — skills/nelson/scripts/nelson_data_patterns.py mines avoid patterns from accumulated mission data, scores them with Fisher's exact test + log-odds, filters against existing orders and a dismissed archive, and synthesises candidate standing orders for human review. Falls back to a heuristic stub when no FM client is wired.
• Three new nelson-data CLI subcommands — detect-patterns, promote-candidate, dismiss-candidate. Promotion writes a new .md under references/standing-orders/ with an audit-lineage comment and adds a row to the SKILL.md lookup table. Dismissal archives the fingerprint so re-runs cannot resurface it.
• Intelligence Brief surfaces pending candidates — CANDIDATE STANDING ORDERS (awaiting review): N appears in both the text and JSON brief when the queue is non-empty.

Design notes

• Ranking is confidence × (1 + novelty), not sigmoid(confidence) × (1 + novelty). Sigmoid maps [0, 1] confidence into [0.5, 0.73] — too compressed to discriminate when novelty's range is [1, 2]. Documented inline in _review_score.
• Novelty uses token containment (fraction of a pattern's tokens covered by an existing order) rather than TF-IDF Jaccard — better fit for short avoid-phrases vs. full standing-order documents.
• Add-only invariant: candidates may never modify or remove existing orders, mitigating the objective-hacking failure mode in DGM Appendix H.
• detect-patterns skips writing an empty queue file on first runs to avoid littering the memory directory.

Testing

• 35 new tests in test_nelson_data_patterns.py
• Full Python suite: 338/338 passing
• ruff check clean on new code

Commits since v2.2.3

• b1fd1ca Update README.md
• 9e158d5 feat: learn standing orders from mission patterns (#87)
• 352de9b fix: address PR review findings on learned standing orders
• 9113af9 Merge pull request #120 from Aspegio/feat/issue-87-learned-standing-orders
• 5ff5acf chore: release v2.2.4