◆ TWIN QUALITY · PUBLIC REASONING DASHBOARDTRANSPARENT

Quality.

Public reasoning dashboard for the sovereign super-intelligence layer. Brain runs on owned hardware · 24/7 · zero cost · no Grok · no Opus — primary tier is CyberPower LM Studio over LAN, with Mac Studio MLX as secondary and a 14-entry deterministic knowledge base as the always-available floor. Same way Cloudflare publishes its outage page, the Twin publishes its accountability surface here.

measuring brain heartbeat…

◆ SOVEREIGN BACKEND LADDER · 4 TIERS · NO FRONTIER

TierBackendModelCtxCostStatusUptime
1cyber-lmsphi-3-mini-4k-instruct4Kzero · sovereign · primary brainup99.4%
2mlx-macglm-4.6v-flash-abliterated1Mzero · sovereign · when upsometimes70%
3cyber-godfathersovereign:dajai-twin (4-agent orchestrator)orchestratedzero · sovereignup99.4%
4static-kb14-entry deterministic FAQinstantzero · always availablealways up100%

Routing: requests cascade DOWN the ladder. mlx-mac is preferred (sovereign + free). When that times out, lms-cyber on the LAN takes over. Then xai-grok, then anthropic-frontier (paid · last resort), then static-kb (deterministic FAQ · always answers).

◆ PERSONA ROTATION · WHO ANSWERS WHAT

DAJAI Twinmusic · identity · mastering · film
52%
Steve The Stock Guymarkets · earnings · macro
24%
Hellcat (Blueprint)creator strategy · OnlyFans ops · vertical integration
15%
Godfatherorchestration · agent state · infra
9%
ROUTING: auto-classifier in /api/twin/intent picks based on question keywords + page-context boost.

◆ FREQUENT TOPICS · TOP 10

#1
Proud 2 Pay (Crenshaw seed)
18%
#2
Catalog size + mastering
16%
#3
Steve weekly briefing
14%
#4
AI agents · network architecture
11%
#5
Creator Blueprint methodology
9%
#6
Code Black ancestry tracing
8%
#7
DARK series + DARK Library
7%
#8
How to plant the seed / supporter tiers
6%
#9
Sovereign network · Mac Studio + CyberPower
6%
#10
Twin biography · who is DAJAI
5%

◆ RESPONSE TIME HISTOGRAM

< 100msstatic KB · intent classifier · tool calls32%
100-500mscached LM Studio · phi-3-mini18%
500ms-2sMLX 27B · normal path28%
2-5sMLX 27B · long context14%
5-15sfrontier · Claude / xAI6%
> 15sfallback cascade · multiple backends timed out2%
p50 ≈ 800ms · p95 ≈ 4s · p99 ≈ 12s · cap 30s before fallback to static KB

◆ WHAT'S MEASURED

  • Backend uptime — pinged every 60s · 30-day rolling avg shown
  • Persona share — % of /api/twin/ask requests routed to each persona by the auto-classifier
  • Topic distribution — top trigger phrases hit · KB lookup matches · LLM-classified topics
  • Response time — p50/p95/p99 server-side measurement (excludes network)
  • Tool use rate — % of responses that called search_catalog / find_agent / lookup_knowledge inline
  • Fallback ratio — % of responses served from static KB when all LLM backends offline
/twin/console/api/twin/tools/api/twin/knowledge/readme