carloscape/octorato/harmonization-over-accretion

Harmonization Over Accretion

When this fires

Any of these prompts:

• "X has N skills/agents/commands, why don't we add them all?"
• "Let's bulk-import the <competitor> catalog"
• "More neurons = smarter brain, right?"
• "We should at least mirror their skill count"

This skill is the canonical reply. It is NOT a no-by-reflex — it is a structured five-point cost analysis that lets you say "no" with evidence, or "yes selectively" with criteria.

The five costs of bulk-import (in any TF-IDF / cosine-similarity brain)

1. Activation dilution — top-K retrieval returns lower per-match signal. The brain hesitates on which skill to load.
2. Wrong-bucket activation — irrelevant skills share enough vocabulary (auth, deploy, security, test) to occasionally win retrieval. A Python deploy task could load perl-security.
3. Silent coverage duplication — the connectome knows closest, not best. Two good skills for the same job = worse than one.
4. Maintenance debt × N — every skill ages. Bulk-import = bulk-staling. Each staled skill is a regression-in-waiting.
5. Forgetting-curve broken — biological brains prune unused synapses. Feature, not bug. A hoarder brain is not an expert.

Vanity metric mapping

Real metric to optimize: per-task activation precision. When a real task arrives, does the brain return the 3 right skills, or 30 with-noise? Bulk-import always pushes the second.

The right alternative — passive cherry-pick via daily watch

1. `repo-watch` — daily monitor of a curated 7-repo watchlist (cap is hard at 7 per Trend Researcher rule). Logs HIGH-SIGNAL diffs to knowledge/repo-watch/triggers/<repo>-<sha>.trigger.
2. `/repo-deep-learn` — manual evaluation of a single trigger using the harmonization model: each pattern earns ADD / MERGE-WITH:<skill> / REPLACE:<skill> / EXTEND:<skill> / SKIP, with similarity gate (TF-IDF cosine) + beat-factor (default 2× incumbent maturity proxy) required for REPLACE.
3. `github-trending-curation` — autonomous breadth sibling. Same harmonization model. Cap ≤3 auto-promotes per day.

Cadence: weeks-to-months for each peer-brain pattern that actually graduates. By design.

Reusable analogy bank

• "Una orquesta de 60 músicos top toca mejor Mahler que una banda escolar de 600."
• "More books on the shelf ≠ smarter. More books read and indexed = smarter."
• "A search engine with 10× more pages isn't 10× better. It's 10× harder to rank."

See also

• [[github-trending-curation]] — the autonomous breadth half of harmonization
• [[repo-watch]] — the targeted daily-watch half
• [[repo-deep-learn]] — the manual one-at-a-time evaluator
• [[feedback-more-skills-not-smarter]] — memory version of this stance
• [[feedback-professional-not-famous]] — adjacent operator stance: substance > visibility