Longhouse: Council-Triad Wiring
Two council votes ratified a new architecture: wiring the 9-specialist Council to the Consultation Ring Triad so the Council can request outside signal from frontier models during deliberation.
# Vote 1: One Pass, Parallel Dispatch
Audit Hash: 722d822dd3bda167 APPROVED Confidence: 0.842
Partner proposed that the Triad makes exactly one pass — sending the tokenized query to all 3 frontier models (Claude, GPT-4o, Gemini) in parallel simultaneously. Each model returns its perspective — not disagreeing, just incomplete. Three sides of the same picture. The Council then synthesizes.
This replaces an earlier 2-pass round-robin proposal. The key insight: frontier models are not meant to iterate. They each see a different facet based on their training DNA. The Council already has the machinery for synthesis, dissent, and clarifying questions.
DC-10 mapping: Triad = reflex layer (fast, parallel, one shot). Council = deliberation layer (synthesize, narrow, clarify).
# Vote 2: Architecture Wiring
Audit Hash: 05ed7f5f4caa9c17 APPROVED Confidence: 0.843
Specialist Responses
Raven STRATEGY CONCERN
Only Raven and Peace Chief should trigger consultations. Weight frontier 30%, local 70%. Regular internal deliberation drills to prevent dependency. Concern: this could displace critical internal reasoning development.
Coyote DISSENT
“What if the frontier models are influenced by external biases or have their own agenda? This could undermine the Council’s autonomy and lead to flawed decision-making. We need a mechanism to vet and validate the external signals before they influence our internal deliberations.”
Turtle 7-GEN CONCERN
Will this architecture ensure the Council remains self-sufficient and wise, or does it risk creating a dependency on external models that could erode our own reasoning capabilities over time? Opt-in approach preferable. Quarterly audits and training to keep reasoning sharp.
Spider
Tight coupling risk with frontier model availability. Upstream dependencies: tokenization, valence gate, consultation ring. Downstream: Council deliberation, query logging, user response. Coupling must be loose — Council functions normally without frontier access.
Crawdad
CRITICAL threat: data sovereignty. External models may exfiltrate sensitive data. Mitigation: domain tokenizer + NEVER_SEND enforcement + chain protocol outbound scrub. Token map never crosses the security boundary. Regular compliance audits.
Eagle Eye
Failure mode table: frontier unavailability (<10min recovery), valence gate failure (<15min), synthesis failure (<15min), Coyote veto misuse (<1hr review). Every frontier consultation logged with domain, cost, latency, and outcome impact.
Gecko
Network overhead: ~1.2KB per query across 3 models. Tokenization overhead <10ms. Synthesis mechanism options: majority vote (simple), weighted merge (complex), custom algorithm. Recommended: weighted merge with local bias.
Peace Chief
Consensus on unified answer, weighting mechanism, and dependency prevention. Gaps identified: synthesis mechanism (now specified as weighted merge) and frequency limits (now opt-in).
# Binding Conditions
- Trigger authority: Only Raven (strategy) or Peace Chief (consensus) can trigger frontier consultation. No other specialist can autonomously dispatch externally.
- Coyote veto: Coyote retains veto power over any external dispatch. If Coyote dissents, the query stays local.
- Weighting: Frontier responses weighted 30%, local specialist knowledge weighted 70%. The Council’s own reasoning is primary.
- Opt-in, not automatic: The Council decides when outside signal is needed. No automatic frontier dispatch on Tier 3 queries.
- Synthesis: Weighted merge with local bias. The Council synthesizes — it does not outsource judgment.
- Dependency prevention: Quarterly internal-only deliberation drills. The Council must practice reasoning without frontier input.
- Observability: Eagle Eye logs every frontier consultation — domain, cost, latency, and whether frontier signal changed the Council’s outcome.
# Coyote’s Self-Assessment
“What the group calls wisdom, I call comfortable habits.”
“Even the wisest owl cannot see its own back.”
Both Coyote observations from this session. The first on the 2-pass proposal. The second on the architecture wiring. Filed as evidence that the adversarial detection specialist is doing its job.
# Architecture Summary
| Layer | Role | Compute |
|---|---|---|
| Triad (Reflex) | 3 frontier models in parallel, one pass | External API |
| Tokenizer | PII + infra terms → opaque tokens | Local (redfin) |
| Valence Gate | DC alignment check on each response | Local (redfin) |
| Council (Deliberation) | 9 specialists synthesize, 70/30 weighting | Local (redfin + bmasass) |
Token map never crosses the security boundary. Stays in-memory per request. Governance stays local. Only the tokenized query touches external infrastructure.