Chamber for Strategic Decisions

Multi-model arbitration layer для L4/L5 decisions. Не для routine work — expensive, применяется selectively.

Реализует принцип Manifesto #6 — “Отдельный судья превыше self-critique” — на strategic уровне, когда одного Agent-Judge недостаточно.

Зачем

Agent-Judge делает quality check одной моделью. Хорошо для L1-L2 routine.

Для L3+ (strategic, cross-functional, irreversible) одной модели мало:

• Model bias — одна модель имеет blind spots, Chamber с несколькими — diversified judgment
• Confidence calibration — agreement between models = higher confidence signal
• Divergence detection — если модели расходятся, это сигнал что задача нетривиальна → human escalation
• Audit trail — multi-model reasoning сложнее manipulate, better for Law 6 compliance

Когда активируется

• L4 decisions — automatic (см. Rules-Criticality)
• L5 decisions — mandatory (+ Founder + future Board)
• Confidence < critical threshold — agent unsure → Chamber как tie-breaker
• Cross-agent contradiction — 2+ agents дают противоречивые рекомендации
• Human explicit request — Founder запросил second opinion
• Post-failure review — scenario failed badly → Chamber analyzes

Когда НЕ активируется

• L1-L2 routine work — Agent-Judge достаточно
• Scenarios где cost важнее чем extra quality
• Когда есть clear playbook — следуй playbook, не запрашивай Chamber
• Deadlock workaround — Chamber не замена для resolving bugs

Chamber expensive. Используется как escalation path, не default workflow.

Архитектура

Panel (multi-model review)

3 разные модели делают independent assessment:

• Model A: Claude Opus (current generation)
• Model B: GPT (current top tier)
• Model C: Gemini (current top tier)

Каждая получает одинаковый context, задача, criteria. Выдают independent:

• Recommendation (accept / reject / modify / escalate)
• Reasoning
• Confidence
• Risks / concerns

Arbiter (synthesis)

Одна senior model (Opus-class) синтезирует panel outputs:

• Aggregate recommendations
• Highlight agreements / disagreements
• Final recommendation с reasoning
• Confidence adjusted (все согласны = higher, split = lower)

Divergence handler

Если panel disagrees significantly (например 2 vs 1, или все trois divergent):

• Expand panel to 5 models
• Add domain-specific model если applicable
• If still divergent → mandatory human escalation

Divergence — сигнал что задача outside consensus capability моделей. Не пытайся force-resolve через larger panel indefinitely.

Chamber request envelope

chamber_request_id: uuid
trace_id: uuid
requester: "agent-ceo" | "orchestrator" | "human-founder"
trigger_reason: "L4_decision" | "confidence_low" | "contradiction" | "explicit"
 
decision_subject:
  type: "strategic_bet" | "policy_change" | "irreversible_action" | ...
  description: "..."
  context: [links, evidence, trace]
  options:
    - option_a: {...}
    - option_b: {...}
 
criteria:
  must_satisfy: [constraint1, constraint2]
  preferred: [preference1, preference2]
  prohibited: [anti_constraint1]  # из Codex
 
budget:
  max_cost_usd: 5.00
  max_duration_min: 10
  model_tiers_allowed: [opus, gpt-top, gemini-top]

Chamber response envelope

chamber_response_id: uuid
chamber_request_id: uuid
timestamp: iso8601
 
panel_outputs:
  - model: "claude-opus"
    recommendation: "..."
    reasoning: "..."
    confidence: 0.82
    concerns: [...]
  - model: "gpt-top"
    ...
  - model: "gemini-top"
    ...
 
arbiter_output:
  recommendation: "..."
  reasoning: "..."
  confidence: 0.78
  agreement_level: "high" | "split" | "divergent"
 
divergence_escalated: true | false
human_escalation_required: true | false
 
cost_usd: 3.40
tokens_total: 45000
duration_sec: 380
 
audit_trail: [full_panel_transcripts, arbiter_synthesis]

Cost discipline

Chamber expensive. Budget guardrails:

Level	Max cost per session	Max per day	Max per month
L4 strategic	$3	10 sessions	$100
L5 existential	$10	no limit per emergency	monitored
Explicit human	$5	5 sessions	$50

Если лимит exceeded — fallback к Agent-Judge + mandatory human escalation, не skip review.

Chamber sessions tracked в Observability — cost QoQ должен снижаться per Manifesto metrics (LLM pricing падает + selective invocation).

Relation к Agent-Judge

• Agent-Judge — ежеtask quality check, single model, L1-L2
• Chamber — strategic review, multi-model, L4-L5

Не duplicate. Judge — quality filter. Chamber — wisdom filter.

Judge может trigger Chamber если quality assessment reveals high-stakes situation agent is under-equipped для.

Founder participation

Per Manifesto принцип #7 — human в петле для критических.

L4 sessions: Founder notified, может participate (add context, question panel), но не blocker.

L5 sessions: Founder approval required после Chamber review. Chamber даёт recommendation, Founder принимает final call. В future (Board stage) — Board vote добавляется.

Audit & learning

Все Chamber sessions logged immutable:

• Full panel transcripts (reasoning от каждой модели)
• Arbiter synthesis
• Final decision + outcome (outcome labeling через Process-OutcomeLabeling)

Learning loop:

• Retro analysis: какие Chamber decisions correlated с good outcomes vs bad
• Panel composition tuning: какие models perform best на какие типы decisions
• Criteria refinement: когда Chamber invocation был justified, когда избыточен

См. ADR-0022-learning-loops.

Open Questions

• Model selection per decision type — universal panel или domain-specific rotations?
• Founder participation modality — real-time vs async review?
• Chamber для post-mortems — только для failures, или также для significant successes?
• Cross-Chamber learning — sharing insights между sessions (meta-level RAG)?
• Fallback strategies если model API down в критический момент

Связанные документы

• Reference-Org-Blueprint — section §5 Layer 5 Chamber
• Agent-Judge — complementary quality check
• Rules-Criticality — L4/L5 definitions
• Rules-Budget — cost guardrails
• Policy-Layer — Chamber calls validated
• HITL-Gateway — Chamber может trigger human escalation
• Manifesto — принципы #6 (separate judge) и #7 (human в петле)
• Codex — red lines Chamber не может override
• Process-Escalation
• Process-OutcomeLabeling
• ADR-0020-reference-architecture-blueprint
• ADR-0022-learning-loops
• Decision Log: ADR-0016 Chamber v2 vision (precedent)