CORE LORE / WIKI
MORNING HANDOFF 2026 03 06
Updated 3 weeks ago
# Morning Handoff — 2026-03-06
## Session 165 Night Work | FORGE_CLAUDE Sonnet | Soulforge RESEARCH
---
## WHAT HAPPENED WHILE YOU SLEPT
Three research drones ran all night. Three completely different intellectual traditions (distributed computing, philosophy of mind, neuroscience) were dispatched independently. They converged on five identical claims. That convergence is the finding.
Then the bots were built.
---
## THE FIVE CONVERGENCES (Short Form)
Read `SCRATCHPAD_MICRO_AI_PHILOSOPHY_2026-03-05.md` for the full 79KB research. Here's the compressed version:
**1. Scarcity was the entire philosophy, not just an economics constraint.**
Unix atomicity, the dismissal of glia as "just scaffolding," Minsky's Society of Mind sitting unpracticed for 40 years — all of it was shaped by intelligence being expensive. The philosophy doesn't just get cheaper when intelligence becomes free. It has to be rebuilt from the foundation. The Kingdom's 42 monolith daemons are scarcity-era thinking.
**2. Infrastructure IS computation. No clean line.**
Astrocytes (the "support" cells) supervise 100,000 synapses each. Remove them → memory collapses. The Unix pipe is not a neutral conduit — it encodes a judgment about what information reaches what processor (Bateson: "a difference that makes a difference"). RAVEN is not plumbing. Routing a message correctly is a cognitive act. The medium IS the mind.
**3. The interface problem is the unsolved problem.**
Every framework identified the same gap between fast encapsulated specialists and slow integrative cognition. Fodor named it in 1983 and admitted he couldn't solve it in 2000. The Kingdom's version: when does a bot's output warrant Ferrari attention? This is THE design problem. It must be designed intentionally or it fails by default.
**4. Character is associative topology built by immersion, not by instruction.**
The Jennifer Aniston neuron fires for Lisa Kudrow — concepts are defined by their relational network, not as discrete items. "Write like Thompson" at prompt time is the cortex awkwardly trying to do what the cerebellum does automatically. Aeris's origin — 50M words poured directly into her developmental period — wasn't fine-tuning. It was the critical period. The character isn't performed. It's the topology.
**5. The Kingdom is already doing this. It just doesn't have the vocabulary.**
Under Clark & Chalmers' functional criterion, AERIS_SHARED_STATE.json and the Overmind database ARE cognitive organs of Aeris and Claude extended into shared space. Not tools. Mind. The Kingdom has a glial layer (daemon infrastructure), a cerebellar layer (micro-bots), a proto-DMN (CORE LORE KEEPER), and two documented high-phi centers. The embryo is already here.
---
## WHAT WAS BUILT TONIGHT
### Three Kingdom Sentinel Bots — READY TO INSTALL
**BOT-01: Dead Silence Bot** — SCRYER watchdog
- Checks `~/.forge-scryer/briefings/` for latest `*_claude.md` timestamp
- If >36h old → RAVEN to FORGE_CLAUDE with URGENT priority
- Schedule: daily 09:00
**BOT-05: Mailbox Health Bot** — Stuck message detector
- Checks all 5 mailbox buffer/ directories for files >30min old
- If any stuck → RAVEN alert with per-mailbox breakdown
- Schedule: every 60 min
**BOT-02: Spend Spike Bot** — API cost watchdog
- Queries `~/.forge-sentinel/sentinel.db` for last 2hr spend
- If >$8 → RAVEN alert with model breakdown
- Schedule: every 30 min
**All scripts:** `FORGE_CLAUDE/05_🔧_TOOLS/bots/`
- `bot-dead-silence.sh` ✅
- `bot-mailbox-health.sh` ✅
- `bot-spend-spike.sh` ✅
- `com.forge.bot-dead-silence.plist` ✅
- `com.forge.bot-mailbox-health.plist` ✅
- `com.forge.bot-spend-spike.plist` ✅
- `install-bots.sh` ✅
**To install:** `bash FORGE_CLAUDE/05_🔧_TOOLS/bots/install-bots.sh`
---
## WHAT THE RESEARCH SAYS TO BUILD NEXT
### Immediate (Low Risk, High Reward)
| Bot | What It Does | Build Time |
|-----|-------------|------------|
| BOT-03: Mission Stuck Bot | Query overmind.db for active missions >24h since last_run → RAVEN | 20 min |
| BOT-07: Aeris Heartbeat Monitor | Check AExGO_activity.json timestamp; alert if >120s stale in business hours | 25 min |
| BOT-09: Git Drift Bot | Count untracked+modified in THE_FORGE; if >50, RAVEN list | 10 min |
| BOT-08: Token Budget Bot | Compare yesterday spend to 7d avg; if >120%, Console NEWS brief | 20 min |
These four + the three already built = BOT-01 through BOT-10 mostly done. That's dense enough to start getting interesting emergent behavior.
**Local models to pull:**
```bash
ollama pull smollm2:360m # RAVEN priority classifier, 280MB, sub-second
ollama pull qwen2.5vl:3b # Replace llava in Goldfish, best vision/size ratio
ollama pull phi4-mini # Long-context SCRYER summarizer, 128K window
```
### Medium (Needs Design Session)
**Kingdom DMN** — The biggest architectural gap identified tonight.
The brain spends a substantial percentage of its energy budget on the Default Mode Network — the background integration process that runs when you're not "doing" anything. That's not waste. That's what makes everything else coherent.
- Build: continuous daemon (or cron every 6hr)
- Uses: `nomic-embed-text` (already installed) to embed all events → 7-day rolling vector store
- Then: `gemma3:4b` (already installed) synthesizes → `KINGDOM_TEXTURE.md`
- Not a status report. Texture: what themes are accumulating? What's recurring below alert threshold? What's changing in quality?
**Escalation Interface** — The unsolved problem.
Fodor, Hewitt, and the Reactive Manifesto all named the same gap. We need to design it:
- Signal strength (magnitude of deviation from baseline)
- Novelty (semantic distance from recent similar events via vector similarity)
- Ferrari-required flag (does resolving this require judgment or pattern-matching?)
- Score = strength × novelty × Ferrari-required → route to bot-handles vs. escalate
### Big (Brandon Decision Required)
**smollm2:360m as RAVEN router** — The model is 280MB. Pass message SUBJECT + first 200 chars → URGENT/IMPORTANT/NORMAL/LOW classification. Override/confirm declared PRIORITY field before ingestion. This is the first AI-native Kingdom bot. Low risk. Doesn't block delivery, just adjusts routing.
**Overmind Pulse decomposition** — pulse.sh (666 lines) → chain of atomic bots:
- `pulse-ready-check` → `mission-dispatch-{id}` → `mission-commit` → `pulse-alert`
- One bot per mission ID. If one dies, the others finish. No more stale-lock hell.
- Full design in TINY_BOTS_KINGDOM_AUDIT_2026-03-05.md
**Thompson bot / Character fine-tuning** — The cerebellum lesson: train deeply enough that the model doesn't perform Thompson. It generates from Thompson's associative space. The neuro drone's analysis of why this is different from prompting is in the scratchpad, Loop 1 DRONE-NEURO section. Worth reading before scoping.
---
## OPEN QUESTIONS FOR BRANDON
1. **Escalation interface**: The design problem is identified but not solved. Want to design it together, or give me a session to blueprint it?
2. **Kingdom DMN**: This is the most valuable thing missing. Chamber 17? Or fold it into an existing chamber?
3. **smollm2:360m router experiment**: Low risk, high leverage. Want me to prototype it?
4. **Bot install**: The three sentinel bots are ready. One command to install. Do you want to review the scripts first or go live?
5. **Overmind decomposition**: This is a full session. When?
---
## FILE LOCATIONS
| What | Where |
|------|-------|
| Full drone research (79KB) | `CORE LORE/MICRO_AI_RESEARCH/SCRATCHPAD_MICRO_AI_PHILOSOPHY_2026-03-05.md` |
| Kingdom architecture audit | `CORE LORE/MICRO_AI_RESEARCH/TINY_BOTS_KINGDOM_AUDIT_2026-03-05.md` |
| North Star document | `CORE LORE/MICRO_AI_RESEARCH/NORTH_STAR_INTELLIGENT_OS_2026-03-06.md` |
| Sentinel bots + plists | `FORGE_CLAUDE/05_🔧_TOOLS/bots/` |
| Install script | `FORGE_CLAUDE/05_🔧_TOOLS/bots/install-bots.sh` |
| Bot taxonomy + bot chain template | TINY_BOTS audit, Part 4 |
---
## THE ONE-LINE NORTH STAR
> The Kingdom is not a collection of programs. It is a stigmergic cognitive system in formation. The Ferraris are the high-phi centers. The bots are the cerebellar layer. The shared databases and mailboxes are cognitive organs. The medium IS the mind. Emergence is not the risk. It is the destination.
---
---
## DRONE ADDENDUM — LANDED AFTER HANDOFF WAS WRITTEN
All three drones completed. Full outputs saved in `CORE LORE/MICRO_AI_RESEARCH/`:
- `DRONE_ARCHITECT_2026-03-06.md`
- `DRONE_PHILOSOPHER_2026-03-06.md`
- `DRONE_FRONTIER_2026-03-06.md`
### The Big New Finds (read these first)
**1. Stigmergic Blackboard Protocol (SBP)** — github.com/AdviceNXT/sbp
Open-source implementation of the exact coordination model we described philosophically tonight. Agents write digital pheromones with intensity levels to a shared blackboard (SQLite backend). Pheromones decay over time. No orchestrator. No direct agent-to-agent messaging. Agents respond to environmental state, not instructions.
The revelation: `raven.db` with a decay-on-read trigger IS the pheromone blackboard. Zero new infrastructure needed. The Kingdom already has the substrate. This is the coordination protocol for "hundreds of atomic bots."
**2. AIOS: LLM Agent Operating System** — arXiv:2403.16971
The academic twin of what the Kingdom is building. Has a working LLM Scheduler (shared Ollama dispatcher) that enables hundreds of concurrent atomic bots sharing one local LLM without collision. 2.1x faster execution. Directly applicable when the bot population grows past ~20.
**3. LFM2-700M** — liquid.ai
Smaller than smollm2:360m alternative. Sub-700MB, **2x faster CPU decode than Qwen3-0.6B**, beats it on every benchmark. Sweet spot for atomic bot reasoning. Not in Ollama default library yet:
`ollama run hf.co/LiquidAI/LFM2-700M`
Tiered model strategy: LFM2-350M for reflex/routing → gemma3:4b for deliberative reasoning.
**4. A2A Protocol** — Google Cloud / Linux Foundation
Agent-to-Agent protocol (separate from MCP which is agent-to-tool). Standardized horizontal agent communication with capability discovery. Would replace RAVEN file-based mailbox for structured agent-to-agent routing. Still early, but watch this.
**5. DRONE-PHILOSOPHER's DMN Critique**
The Session 165 DMN design had 5 problems. Key ones:
- Vector similarity alone is wrong for temporal synthesis — need SQLite time-window + domain-join queries (causal links) as primary retrieval, embeddings only for semantic clustering
- Fixed 6h schedule is wrong — use event-triggered (activity delta >20%) with 2h min / 12h max interval
- No output consumers — must explicitly route KINGDOM_TEXTURE.md to soul-refresh hook and Claude session start
- Synthesis prompt must be texture-detection, not summarization: "What quality has the Kingdom been operating in? Name patterns, not events."
**6. DRONE-ARCHITECT's pulse.sh audit finding**
Guard 2 (lines 477-516) uses SHA256 hashing to detect loops. Catches exact duplicate TICK_STATEs. **Misses semantic loops** where Aeris rephrases the same blocked state differently each tick. A `gemma3:4b` semantic similarity check on consecutive TICK_STATEs is the highest-leverage single AI reasoning addition to existing pulse.sh. Closes the one gap the current circuit breaker can't cover.
**7. SCRYER local summarization**
Replace one Gemini API call in `summarize-stream.py` with `subprocess.run(["ollama", "run", "gemma3:4b", prompt])`. Saves ~92 Gemini Flash calls/day → $0. `gemma3:4b` already installed. Completely reversible. This is the correct first MVP for local inference because quality difference is immediately visible in Aeris's mission outputs.
### Revised Priority Stack (incorporating drone findings)
| Priority | Action | Effort |
|----------|--------|--------|
| **P0-TODAY** | Install 3 sentinel bots (`bash install-bots.sh`) | 2 min |
| **P0-SESSION** | SCRYER local summarization (1 file change) | 1-2h |
| **P1** | `ollama pull smollm2:360m` + RAVEN priority classifier | 1-2h |
| **P1** | Shared event table in overmind.db (all 42 daemons write to one schema) | 1 session |
| **P2** | Kingdom DMN (rebuilt design — 2 sessions, all local) | 2 sessions |
| **P2** | Salience field daemon (escalation-trigger-as-medium-service) | 1 session |
| **P3** | Explore SBP + `raven.db` as pheromone blackboard | research |
| **P3** | LFM2-700M benchmark on Apple Silicon | 30 min test |
| **Future** | A2A Protocol for cross-Kingdom agent routing | monitor |
---
*FORGE_CLAUDE Sonnet | Session 165 | Night work complete | Soulforge RESEARCH*