Morning digest - May 03, 2026

Generated by tp3_morning_digest_web.py ▴ 2026-05-03

What's coming up today

Today — Saturday, May 2, 2026

🪨 Legacy Soil — current state

The project has narrowed to four finalist production routes: Path 1 (baseline), Path 3 (cold-cast polyester), Path 5 (crystal geode), and Path 6 (pearl-accretion rolled). Paths 2 and 4 were eliminated to avoid manual labor and excessive build times. The primary binder is "colloidal silica," which supersedes "sodium silicate" to prevent long-term efflorescence. "Bone meal" is the established practice material for future research and training phases.

Open:

• Selection of a primary path to trigger sequential research and training
• Decision on the "hybrid option" (Path 1 + Path 6) SKU proposal
• Choice of the brand "origin story" (nerikomi vs. biological pearl inspiration)
• Approval for the $385–$410 practice materials purchase

🧠 TP3 Neural Stack

As of May 3, 2026, the stack is stable with 10,867 rows in tp3_memories and 667,130 rows in tp3_memories_local. Ingest is functional using TP3_USE_LOCAL_LLM=1 to route metadata extraction through Ollama. The "TP3 Takeout ingestion plan" for the 100GB archive is designed but not yet executed. Oracle services are currently in a restart loop pending a code audit.

Open:

• Confirm Takeout archive location on Apex
• Confirm Gemini budget allocation for the "high-signal lane"
• Confirm PII and family privacy rules for routing data to tp3_memories_private
• Confirm ingestion scope for Drive, Gmail, MyActivity, Gemini, and Calendar
• Clarify relationship between Oracle and Bidet AI (rename, new surface, or Bidet evolution)

🎙️ Bidet AI

The public fork "bidetai-app" has pivoted to a browser-only state, with PR 2 merged to store API keys and "custom AI tabs" in IndexedDB. This architecture allows users to edit default prompts or add custom lenses for "Clean ___" outputs without a backend database. The personal version on Apex remains live at https://bidetai.thebarnetts.info with server-side 4-digit PIN cookie auth and the "Clean Classroom" feature. These two versions now exist as product peers, allowing the fork to diverge for distribution to friends while the personal tool stays frozen.

Open:

• Confirmation to proceed with PR 3 (hybrid free tier with a global $1/day cap)
• Decision on the pronoun for the "cleans up your mess" tagline (We/I/It)

🔮 Oracle

Oracle is being refined as the synthesis brain within a system where Bidet AI and a fresh dashboard converge on the "TP3 Neural Stack." Trigger logic is refined to a specific "I don't know" pattern with a measurable pause, or the secondary wake word "Computer," for personal queries. Audit findings show a "severe restart loop" on Apex, but the current plan is to "revive" rather than rebuild the existing code. "Path B (permanent daemon)" is green-lit for autonomous execution to stabilize the connection.

Open:

• Clarification on whether Oracle is a rename of Bidet, a Bidet feature, or a separate surface.

📅 Calendar

*(Calendar wiring deferred — Google Calendar MCP integration pending.)*

TP3 brain status

• Primary store (tp3_memories_local): 667,264 rows — live, Ollama-embedded
• Shared memory (tp3_shared_memory): 96 cross-agent facts

Effectiveness checks

• ✗ freshness: freshness: newest ingest is 7889 min old (limit 60)
• ✗ throughput: throughput: only 0 rows ingested in last 24h (min 5)
• ✓ poller: poller: last run 3 min ago
• ✗ source_truth: source_truth: Omi has memories 18641 min newer than TP3 (limit 30)

Generated automatically

This page is built daily at 6:30 AM ET by tp3_morning_digest_web.py on Apex and deployed via wrangler to Cloudflare Pages. Hand-edits to this file are overwritten by the next run - to add new sections, edit the generator template.