jlacour-git

## DigestLearnings-gist.md

      
              2 files
            
          
              0 forks
            
          
                0 comments
              
            
              0 stars
            
          
                jlacour-git
                / DigestLearnings-gist.md
            
            
              Last active
              March 16, 2026 08:46
            
              
                DigestLearnings — Periodic Review of Learning Signals (PAI Skill)
              
          
    DigestLearnings Workflow

Review unprocessed learning signals and extract improvement proposals.
Phase 1 — DIGEST (scan and extract)

Step 1: Find the watermark

Read ~/.claude/MEMORY/LEARNING/DIGEST-LOG.jsonl. Find the most recent entry's timestamp field. This is the watermark — only process files created AFTER this timestamp.

  
## README.md

      
              3 files
            
          
              0 forks
            
          
                0 comments
              
            
              0 stars
            
          
                jlacour-git
                / README.md
            
            
              Created
              March 12, 2026 18:24
            
              
                Rule Adherence Interventions for PAI — Three concrete changes that improved AI rule-following behavior
              
          
    Rule Adherence Interventions for PAI

Three concrete changes that improved AI rule-following behavior in PAI. Found through a structured investigation (Science → Council → RedTeam) documented in Discussion #945.
Background

PAI's AI assistant systematically underused its own quality mechanisms despite explicit rules. Across 43 tracked sessions: effort levels skewed low (35% Standard), capability invocation averaged 0.4/session against stated minimums of 1-2, and weekly ratings trended down.
The root cause: the model finds rationalization paths to skip rules. An "Efficiency" section in steering rules gave explicit textual permission to shortcut. "Standard = DEFAULT" anchored effort estimation downward. And without PlanMode enforcement, complex tasks got no structured planning.

  
## AISTEERINGRULES.md

      
              4 files
            
          
              0 forks
            
          
                0 comments
              
            
              0 stars
            
          
                jlacour-git
                / AISTEERINGRULES.md
            
            
              Last active
              March 16, 2026 09:03
            
              
                Rule Adherence Investigation — Science → Council → RedTeam findings (PAI)
              
          
    AI Steering Rules — Personal

Personal behavioral rules for {PRINCIPAL.NAME}. These extend and override SYSTEM/AISTEERINGRULES.md.
Priority: Trust > Correctness > Quality

When rules conflict, higher level wins. [T]=Trust [C]=Correctness [Q]=Quality.
Trust rules NEVER yield.


## 01-LOCAL_PATCHES-TEMPLATE.md

      
              5 files
            
          
              0 forks
            
          
                0 comments
              
            
              1 star
            
          
                jlacour-git
                / 01-LOCAL_PATCHES-TEMPLATE.md
            
            
              Last active
              March 18, 2026 20:57
            
              
                PAI Local Patch Tracking + Upstream Monitoring System
              
          
    Local Patches to PAI SYSTEM Files

Tracks modifications to SYSTEM-tier files that would be overwritten by a PAI update.
Check this file before and after any git pull + reinstall.

How to Use


Before PAI update: Run the upgrade check tool first:


## BuildSemanticIndex.ts
#!/usr/bin/env bun
/**
 * BuildSemanticIndex.ts — Builds/refreshes the semantic search vector index
 *
 * Embeds files from MEMORY/, auto-memory/, and ~/AI-Projects/ CONCEPT.md files
 * using Ollama mxbai-embed-large and stores vectors in LanceDB.
 *
 * Sub-file chunking: splits markdown at heading boundaries, splits oversized
 * sections (>4000 chars) at paragraph breaks, skips empty-heading chunks.
 * Content-hash dedup: SHA-256 per chunk, skips re-embedding unchanged content.

## Algorithm-Complexity-Gate.md

      
              4 files
            
          
              0 forks
            
          
                1 comment
              
            
              0 stars
            
          
                jlacour-git
                / Algorithm-Complexity-Gate.md
            
            
              Last active
              March 12, 2026 19:59
            
              
                Fix: Algorithm Mode Skipping in PAI v4.0 — root cause analysis + implementation
              
          
    STEP 1 — COMPLEXITY GATE (before reading the Algorithm file):
Default: ALGORITHM. Only downshift if the task is genuinely trivial.
Evaluate: "Can I complete this with ≤2 tool calls, no external APIs, no file creation/deletion, and no verification needed?"

NO (or uncertain) → continue to Step 2 (read the Algorithm file)

Voice: bun ~/.claude/PAI/Tools/voice.ts notify algorithm-confirmed


YES → downshift to NATIVE mode only if NONE of these disqualifiers apply:

Involves GitHub, Forgejo, or any external API
	#!/usr/bin/env bun
	/**
	* BuildSemanticIndex.ts — Builds/refreshes the semantic search vector index
	*
	* Embeds files from MEMORY/, auto-memory/, and ~/AI-Projects/ CONCEPT.md files
	* using Ollama mxbai-embed-large and stores vectors in LanceDB.
	*
	* Sub-file chunking: splits markdown at heading boundaries, splits oversized
	* sections (>4000 chars) at paragraph breaks, skips empty-heading chunks.
	* Content-hash dedup: SHA-256 per chunk, skips re-embedding unchanged content.