guide_v3.5_corrected.md

Mr. Smith-Davinci v3.5 — Operator's Manual (Corrected)

Matches the live code in smith_davinci_v3.5/ and Space kirikir13/Mr-Smith-Davinci_Pattern-Weaver-Super-Agent.
Supersedes the informal Kimi guide in guide.pdf (which mixed v3.5 with older 6AG / v4.2 concepts).

---

Layout (what you see on screen)

Area	Contents
Left	Pattern Weaver Chat · Your Request · WEAVE · System Status · L.I.P.S. JSON · Neuraxon JSON
Right tabs	Settings · Mode · Patterns · Buckets · Harness Evolution · System Status
Below chat	Council Deliberation Log (when council is enabled)

---

Step 1 — API keys (Settings tab)

Set at least one backend key. On Hugging Face Spaces, you can use Space Secrets instead of typing keys in the UI:

Field	Space secret names (any one works)
OpenRouter	OPENROUTER_KEY or OPENROUTER_API_KEY
Kimi	KIMI_KEY or MOONSHOT_API_KEY
HuggingFace	HF_KEY or HF_TOKEN

Backends available: OpenRouter · Kimi · HuggingFace · Local

• OpenRouter — recommended gateway (many models via one key)
• Kimi — direct Moonshot API
• HuggingFace — HF Inference API
• Local — llama.cpp / local endpoint (not for HF Space runtime)

Model Override (optional): forces a single model id for direct (non-council) calls. Default OpenRouter model in code: davidau/qwen3.6-27b-heretic-uncensored. Leave empty to use backend defaults.

---

Step 2 — Mode tab (operational mode)

Active Mode (dropdown)

These are prompt prefixes, not separate apps:

Mode	Purpose
weave	Pattern recognition and creative synthesis (default)
quant	Therium / compression / quantization framing
memory	Neuraxon memory and engram framing
docupack	Documentation and knowledge packaging

Consult Council of Advisors (checkbox)

Setting	Behavior
ON (default)	Four advisors run in parallel → DavidAU chairman synthesizes → richer, slower
OFF	Single backend call only → faster, lighter

Neuraxon Override (optional)

Force memory gating state: HOT · IDLE · INHIBITED
(This is Neuraxon trinary, not L.I.P.S. — do not confuse the two.)

---

What L.I.P.S. actually is (automatic)

L.I.P.S. runs internally on every WEAVE. You do not pick “Dopamine mode” or “Serotonin mode” in the UI.

5 chemicals (0.0–1.0): Dopamine · Oxytocin · Serotonin · Endorphin · Adrenaline

6 phases (cycle): Resolution → Excitement → Plateau → Edge → Climax → Refractory → (repeat)

Watch the L.I.P.S. Neurochemical State JSON under the chat after each run.

---

What Neuraxon actually is

State	Meaning
HOT (+1)	Full tool access, deeper chains
IDLE (0)	Standard operation
INHIBITED (-1)	Restricted; memory consolidation

Shown in Neuraxon Trinary State JSON. Updated from L.I.P.S. + pattern quality after each run.

---

Council (when checkbox is ON)

Advisors (parallel, ~60s timeout each):

Advisor	Role	Default route
Qwen Coder	Ministerial Reasoner	OpenRouter
Gemini	Hermes Generalist	OpenRouter
Kimi	Phi Critic	Kimi API
Grok	Seed Creative	OpenRouter

Chairman: DavidAU / Qwen synthesis merges advisor outputs.

Not in v3.5: 6-agent voting, Devil’s Advocate rounds, “Revision 1/2”, or QUICKWEAVE mode. Those are design targets / older docs — not this build.

---

Step 3 — Patterns tab

Shows extracted patterns from past runs (SQLite patterns table).

• Filter by type: script · idea · concept · training · pattern
• Click Refresh Patterns

Not in v3.5: pre-built template library, click-to-fill patterns, or “Save as Pattern” button.

---

Step 4 — Buckets tab

Auto-filled insight buckets from responses and runs:

script · idea · concept · training · pattern · harness · lips_trace · council_log

---

Step 5 — Harness Evolution tab

Meta-Harness loop (Niklaus 2026 style scoring):

1. Record Episode from Recent Runs — batches unscored agent_runs into an episode
2. Get Harness Suggestions — reads episode history and suggests improvements

Promotion uses blended score (pooled success + pattern quality + synaptic weight − token penalty).

---

Step 6 — Type request and hit WEAVE

The model is prompted to emit structured sections when possible:

PATTERN INSIGHT: <cross-domain pattern>
SOLUTION: <approach>
GENERALIZABLE TOOL: <reusable component>
NEXT EVOLUTION: <what to explore next>

These are parsed and stored automatically.
Known gap: NEXT EVOLUTION is saved but not yet auto-fed as your next prompt — you still type the follow-up yourself.

Example requests

Architecture:

Design a FastAPI service that loads a Therium tensor chunk, runs PCA to 3D,
scores player centroid submissions by SNR, and writes THEORETICAL recipes to SQLite.

Council debate (with council ON):

Should we use SVD rotation vs spectral basis for Therium quantization?
Stress-test the weakest assumption and propose a simpler 80/20 alternative.

Quant mode:

MODE quant: Compare BF16 vs GGUF kernel tradeoffs for a 27B layer cluster.

---

Quick reference

You want…	Backend	Council	Mode
Full deliberation	OpenRouter	ON	weave
Fast single answer	OpenRouter	OFF	weave
Quant / compression	OpenRouter	ON/OFF	quant
Memory / engram work	OpenRouter	ON	memory
Docs / packaging	OpenRouter	ON	docupack
Own HF model	HuggingFace	OFF	weave

---

Troubleshooting

Symptom	Likely cause	Fix
[No API key configured]	Empty keys	Fill Settings or Space secrets
[Council error] / timeouts	Advisor API slow or key missing	Check OpenRouter/Kimi keys; try council OFF
LIPSPhase enum error	Old build	Pull latest Space (fixed in database.py)
Patterns empty	No successful runs yet	WEAVE once with structured output
Data gone after restart	HF ephemeral disk	Export from Buckets / System Status; use local run for persistence
Heavy model crash on free GPU	ZeroGPU limits	Use OpenRouter for large models; avoid Local/27B on Space

---

Space URL

https://huggingface.co/spaces/kirikir13/Mr-Smith-Davinci_Pattern-Weaver-Super-Agent

---

Local run

cd smith_davinci_v3.5
pip install -r requirements.txt
python app.py
# http://localhost:7860

---

DR Studios · Dave / kirikir13 · God willing and pure.