S7 SkyQUBi — Architecture (Public Overview)

Release status (as of 2026-04-14): the architecture below is the stable public-facing overview. Internal work on the CORE reframe (PRISM/GRID/WALL), the persona-internal council pattern, the immutable-fork rebuild ceremony, and voice calibration for the three personas is in active Chair-draft / Jamie-authorized review pending the household Chief of Covenant’s witness. Full public launch: July 7, 2026 · 7:00 AM Central. See README.md for release status and path.

This is the high-level architecture. Internal runbooks, deploy procedures, and infrastructure details live in the private repo and are not published.

The shape

  ┌────────────────────────────────────────────────────┐
  │                   You / Your App                    │
  └─────────────────────────┬──────────────────────────┘
                            │
                   (HTTP or direct CLI)
                            │
  ┌─────────────────────────▼──────────────────────────┐
  │             S7 Public Chat API / CLI                │
  │         (accepts a question, returns consensus)     │
  └─────────────────────────┬──────────────────────────┘
                            │
                  ┌─────────▼──────────┐
                  │    CWS Engine       │  ← the covenant layer
                  │   (classification,  │
                  │    circuit breaker) │
                  └─────────┬──────────┘
                            │
            ┌───────────────┼───────────────┐
            │               │               │
  ┌─────────▼─────┐ ┌───────▼──────┐ ┌─────▼─────────┐
  │  Witness 1    │ │  Witness 2   │ │  Witness N    │
  │  (qwen 3B)    │ │  (deepseek)  │ │  (qwen 0.6)   │
  └───────────────┘ └──────────────┘ └───────────────┘
            │               │               │
            └───────────────┼───────────────┘
                            │
                  ┌─────────▼──────────┐
                  │   MemPalace         │  ← shared memory
                  │   (Akashic index)   │
                  └─────────┬──────────┘
                            │
                  ┌─────────▼──────────┐
                  │  Memory Ledger      │  ← insert-only store
                  │  (PostgreSQL +      │
                  │   SQLite)           │
                  └────────────────────┘

The flow

  1. You send a question to the public chat API (or call the witness function directly).
  2. CWS Engine receives it, forwards it in parallel to every witness. No witness sees another’s answer.
  3. Each witness answers independently. Different model family, different training data, different cognitive plane.
  4. CWS scores agreement using embedding similarity (all-minilm) with Jaccard token fallback.
  5. Classification is assigned:
    • FERTILE — agreement above threshold, answer returned
    • AMBIGUOUS — partial agreement, returned with confidence
    • BABEL — disagreement beyond 70%, circuit breaker trips, no answer
    • UNVERIFIED — insufficient witness response, no answer
  6. MemPalace records the interaction in the Akashic language index for future recall, with PLAN location metadata and hallucination flags.
  7. Memory Ledger stores the bond vectors in INSERT-only form. Nothing that entered the system can be made to have never existed.

The covenant

The architecture embodies seven laws. See COVENANT.md for the full table.

The most important one: the circuit breaker. At 70% BABEL, the system refuses to answer. This is not a hyperparameter. It is the constant that makes disagreement a feature, not a failure.

The witnesses

The default install runs the lite 3+1 set — three small models from three different families plus the CWS reporter. The full set is 7+1 — seven models on seven cognitive planes, plus the deterministic reporter at the center. Both sets use the same consensus and classification machinery; the lite set is optimized for running on hardware you already own.

Lite set Full set (7 witnesses)
qwen2.5:3b LLaMA 3.2 3B (Sensory)
deepseek 1.3b Mistral 7B (Episodic)
qwen3:0.6b Gemma 2 9B (Semantic)
+ CWS Phi-4 (Associative)
  Qwen 32B (Abstract)
  DeepSeek R1 (Causal)
  BLOOM (Generative)
  + CWS

What the covenant protects against

What it does not protect against: all witnesses being wrong in the same way. If every witness is trained on the same biased data, they may agree on something untrue. The covenant catches disagreement, not error. That’s why the full set uses models from seven different training lineages.

Sovereignty properties

What’s out of scope for this document

Those live in the private repo’s docs/internal/ and are not published. If you’re integrating S7 in a way that needs them, contact omegaanswers@123tech.net.

The motto, again

Love is the architecture. Everything above is a consequence of taking that seriously.