Chapter 20: Compiling to any LLM (Cross-provider Interoperability)

The Abstract Instruction Layer: Decoupling the Mold from the Model

The core promise of Dredge is Sovereignty. To achieve this, we must decouple the “Logic” (Mold) from the “Compute” (Model).

Dredge defines an Abstract Instruction Layer (AIL). A Mold is written in AIL—it describes the intent (Bridge, Gavel). The Compiler translates this AIL into the specific tokens required by the target model.

• Intent: “Extract JSON”
• Compile to OpenAI: response_format: { type: "json_object" }
• Compile to Llama: <|begin_of_text|>... Output raw JSON only.

This layer protects the developer from the “API Churn” of model providers. The Mold remains constant; only the Compiler’s driver updates.

Provider Shims: Adapting Molds for Gemini, OpenAI, and Llama

Different models have different “Cognitive Accents.”

• Gemini loves high-context system instructions.
• Claude loves XML tags (<thinking>...</thinking>).
• Llama loves explicit instruction tuning patterns.

The Provider Shims are the adapters. When you switch your backend from OpenAI to Anthropic, the Shim automatically rewrites the Mold’s anchors to match the target’s preferred dialect. A “Reasoning Trace” might be compiled as a trace field in JSON for GPT-4, but as a <trace> XML block for Claude. The developer doesn’t care; the logic holds.

Performance Profiles: Which models compile which Molds best?

Not all compilers are equal. The Dredge Compiler includes a Profiler. It runs your Mold against a standardized benchmark suite on every available model.

• Llama 8B: 92% Accuracy, $0.05 cost.
• Gemini Flash: 96% Accuracy, $0.10 cost.
• GPT-4o: 99% Accuracy, $5.00 cost.

The Profiler gives you the data to make the “Arbitrage Decision.” You might choose Llama for the “First Pass” and escalate to GPT-4 only for the hardest 1% of cases.

[IMAGE PROMPT: A ‘Profiler Dashboard’ in the IDE. A bar chart compares ‘Cost vs. Accuracy’ for 5 different models running the same ‘Medical Triage’ mold. Llama 8B is highlighted as the ‘Best Value’. GPT-4 is highlighted as ‘Best Accuracy’.]

Write Once, Dredge Everywhere

This leads to the “Java” moment for AI. Write Once, Run Anywhere. A business can build its entire operational logic in Dredge Molds. Today, they run on OpenAI. Tomorrow, they run on a local Llama 4 cluster in their basement. The logic—the intellectual property—is portable.

This breaks the vendor lock-in that currently defines the AI industry. The model becomes a commodity component, swappable like a hard drive. The Mold becomes the persistent asset. We have moved the value from the “Cloud” to the “Code.”