gnr8tr/falco-decision-models

Falco Sovereign Decision Engine (Experimental Prototype)

Falco is an experimental, non-autoregressive System 1 routing model exploring multi-query sequence packing over a shared context prefix. It evaluates multiple categorical classification queries concurrently in a single encoder forward pass without inter-query token interference.

It was designed to test whether 4D block-diagonal prefix attention masking and dynamic position ID resets can eliminate slot permutation drift in multi-query encoder setups while keeping computational footprints under 300 MB.

• Primary Backbones: Fine-tuned DistilBERT-base-uncased (66M) for English; INT8-quantized IndicBERT-v2 (278M) for 22 Indic scripts.
• Format: Exported directly to ONNX Runtime (CPU and DirectML/CUDA compatible).
• Scope: Experimental research artifact. Designed for structured triage, policy gating, and fast classification—not open-ended generation.

---

Intended Use & Boundaries

Recommended Use Cases

• High-Frequency Policy Pre-Routing: Screening inbound text (support tickets, logs, prompts) for binary or low-cardinality ($K \le 4$) routing categories before invoking expensive LLMs.
• Air-Gapped & Edge Deployments: Environments requiring local, CPU-based classification with deterministic sub-30ms execution times and zero marginal API fees.
• Cross-Query Invariance Research: Applications requiring mathematical guarantees that query order does not leak information between concurrent classification heads.

Out-of-Scope & Unsupported Uses

• Do NOT use for open-ended text generation: Falco is an encoder with discrete linear heads. It cannot generate rationales, summaries, or conversational prose.
• Do NOT use as a replacement for high-capacity language models: DistilBERT (66M) has limited linguistic capacity. It will fail on complex, multi-sentence reasoning chains and subtle legal/medical nuances.
• Do NOT use for high-cardinality classification ($K > 4$): The output heads are structurally bounded to 4 logits. Passing queries with 5 or more options will result in clipped classes.
• Do NOT deploy in high-stakes clinical or safety-critical decisions without human verification: The model's training distribution is bounded, and out-of-distribution performance degrades sharply.

---

How It Works: Architectural Mechanics

Contiguous Input Sequence:
[CLS] State Prefix Context Tokens... [SEP] Q1 Tokens... [SEP] Q2 Tokens... [SEP] Q3 Tokens... [SEP]
                                         │                   │                   │
4D Attention Mask:                       ▼                   ▼                   ▼
Each slot attends to:           [State Only + Q1]   [State Only + Q2]   [State Only + Q3]
Position IDs:                   [0 ... L_state-1]   [L_state ... +N1]   [L_state ... +N2]
                                         │                   │                   │
Head Pooling & Clamping:                 ▼                   ▼                   ▼
Logit Clamping:                 Clamp(K=2)          Clamp(K=3)          Clamp(K=4)