Multi-Model Dynamic Fanout & Cognitive Routing Architecture

Formal specification for intelligent cognitive capacity balancing, tier-governed task dispatch, and automated model catalog synchronization across the cordanaLLM fleet.

flowchart TD
    TASK["Autonomous Engineering Task"] --> ARBITER["Model Capacity Arbiter"]

    subgraph CognitiveHierarchy["3-Tier Cognitive Hierarchy"]
        T1["Tier 1: Frontier Reasoning<br/>(Opus 4.6, Pro 2.5, O3, Grok 3, R1)"]
        T2["Tier 2: Workhorse Engineering<br/>(Sonnet 4.6, GPT-4o, Flash 2.5, Mistral Large)"]
        T3["Tier 3: Open-Weights / Local OSS<br/>(gpt-oss, Qwen 2.5 Coder, Llama 3.3)"]
    end

    subgraph DynamicGovernance["Headroom & Integrity Gates"]
        HEADROOM{"Available Headroom<br/>C_avail(M) >= 0.20?"}
        CONCURRENCY{"Active Subagents<br/>N_active(M) < 2?"}
        AUDIT_GATE{"Orthogonal Family?<br/>Family(Auditor) != Family(Author)"}
    end

    ARBITER --> HEADROOM
    HEADROOM -- "Yes" --> CONCURRENCY
    HEADROOM -- "Depleted (429/OOM)" --> CASCADE["Cascade to Secondary Tier"]
    CONCURRENCY -- "Yes" --> DISPATCH["Dispatch Agent Execution"]
    CONCURRENCY -- "Exceeded (Spam Lock)" --> DIVERSIFY["Fan Out to Sibling Provider"]
    DISPATCH --> COMPILATION["Synthesize Pull Request"]
    COMPILATION --> AUDIT_GATE
    AUDIT_GATE -- "Verified Disjoint" --> PEER_REVIEW["Cross-Model Audit Approval"]

    CASCADE --> ARBITER
    DIVERSIFY --> ARBITER

1. The 3-Tier Cognitive Hierarchy

Autonomous engineering agents must not spam a single model endpoint or exhaust Tier 1 subscription allowances on mechanical tasks. Model dispatch is strictly stratified into three operational tiers:

Tier	Primary Role	Qualified Architectures	Failure Escalation
Tier 1: Frontier	System architecture, formal proofs, complex refactoring, final root-cause analysis	`claude-opus-4.6`, `gemini-2.5-pro`, `o3-high`, `grok-3-deep`, `deepseek-r1`	Fallback to Tier 2 Workhorse
Tier 2: Workhorse	Code synthesis, AST refactoring, unit test generation, CI debugging	`claude-sonnet-4.6`, `gpt-4o-latest`, `gemini-2.5-flash`, `mistral-large-2411`	Fallback to Tier 3 OSS
Tier 3: Open OSS	Linting sweeps, docstrings, formatting, local pre-commits, secret scans	`gpt-oss-120b`, `qwen2.5-coder-32b`, `llama-3.3-70b-instruct`, `deepseek-v3`	Queue locally on vLLM/Ollama

2. Dynamic Cognitive Headroom Formulation

The Model Capacity Arbiter dynamically computes available execution headroom $C_{\text{avail}}(M) \in [0, 1]$ before granting task dispatch:

$$C_{\text{avail}}(M) = \max\left(0, \, \min\left(1 - \frac{\text{RPM}{\text{active}}(M)}{\text{RPM}}}(M)}, \, 1 - \frac{\text{TPM{\text{active}}(M)}{\text{TPM}(M)\right)\right)$$}}(M)}\right) \times \left(1 - \text{ErrorRate}_{5\text{m}

Operational Thresholds

Nominal Band ($C_{\text{avail}} \ge 0.40$): Unrestricted task dispatch up to concurrency limits.
Constrained Band ($0.20 \le C_{\text{avail}} < 0.40$): Rate-limited scheduling; low-priority tasks degraded to lower tier.
Depletion Trigger ($C_{\text{avail}} < 0.20$ or HTTP 429): Immediate fallback cascade; model quarantined for 60 seconds.

3. Pre-Agent Dispatch Anti-Spamming Lock

To prevent runaway loops from monopolizing API quotas or flooding a single frontier model, the pre-dispatch hook (.config/agent/hooks/pre_agent_dispatch.py) enforces strict concurrency quotas:

$$\forall M \in \text{Catalog}, \quad N_{\text{concurrent_subagents}}(M) \le 2$$

When a parent agent attempts to spin out multiple subagents targeting identical model identifiers: 1. The first 2 subagents execute on Model $M_1$. 2. Subagent 3 is automatically rerouted to the highest-headroom alternative within the same cognitive tier (e.g., diverting claude-sonnet-4.6 to gemini-2.5-flash or gpt-4o). 3. If all tier siblings exceed concurrency caps, tasks are fanned out to zero-cost local inference (gpt-oss, qwen2.5-coder).

4. Orthogonal Cross-Model Auditing

Self-review within the same model family produces systematic cognitive blind spots. HISS-16 mandates orthogonal model auditing for code review and verification:

$$\text{Family}(\text{Auditor}) \ne \text{Family}(\text{Author})$$

Author Model Family	Banned Reviewer Families	Permitted Reviewer Families
Anthropic (`claude-*`)	Anthropic	OpenAI (`o3`, `gpt-4o`), Google (`gemini-`), Open OSS (`deepseek-`, `qwen-*`)
OpenAI (`gpt-`, `o`)	OpenAI	Anthropic (`claude-`), Google (`gemini-`), Open OSS (`deepseek-`, `qwen-`)
Google (`gemini-*`)	Google	Anthropic (`claude-`), OpenAI (`o`), Open OSS (`deepseek-`, `qwen-`)
Open OSS (`qwen-`, `deepseek-`)	Open OSS (same weights)	Frontier Tier 1 / Tier 2 Cross-Family

5. Automated Catalog Synchronization

Model identifiers, pricing, and context limits drift rapidly. standardsctl models sync automates discovery: - Upstream Gateway Sync: Queries configured LiteLLM / OpenRouter model registries for active models, context windows, and cost tiers. - Zero-Cost Local Endpoint Discovery: Probes http://localhost:11434/api/tags (Ollama) and http://localhost:8000/v1/models (vLLM) to register locally available weights. - Lockfile & Schema Reification: Synchronizes state to .config/models/catalog.json and verifies deterministic routing.