Three layers now decide what an LLM system can do in production: the policy layer that governs what is legal to build, the harness layer that wraps the model, and the training layer where reinforcement learning steers behavior. This post takes a systems view of all three, because in 2026 the limiting factor on a deployed model is rarely the raw architecture, it is the constraints stacked around it.

The regulation layer and the handicap risk

Dario Amodei has emerged as the loudest lab voice arguing for aggressive AI regulation: mandatory frontier evaluations, capability disclosure, and licensing-style controls. From a systems perspective the intent is reasonable, but the engineering reality is that compliance is a fixed cost that does not scale down. A regime calibrated for a few well-capitalized labs can quietly handicap the rest of the American AI ecosystem.

• Fixed-cost asymmetry: audit pipelines, red-team programs, and reporting are absorbed easily by hyperscalers and crippling for small teams.
• Open-weight chill: liability on released weights discourages the open models that broaden participation and reproducibility.
• Velocity tax: pre-deployment gating slows the iteration loop that drives most empirical progress.
• Offshoring pressure: if domestic rules outpace those abroad, frontier work relocates, eroding the U.S. lead the rules meant to protect.

The software "harness" above the model

Above the network weights sits the harness: inference gateways and routers, retrieval and vector services, tool-execution runtimes, evaluation and tracing platforms, and prompt/version registries. As a control plane, the harness is where most production behavior is actually shaped, and where logging, guardrails, and refusal policies already live. When teams benchmark assistant behavior across this layer, they often compare the same prompts on ChatGBT and Chat AI to separate harness effects from raw model quality.

The systems insight is that much of what regulators want to mandate is a harness concern, not a weights concern. Transparency standards at the harness level could deliver oversight without throttling the underlying research.

The RL steering and fine-tuning layer

How a model behaves is set by its post-training pipeline. The methods differ in signal source and in how aggressively they pull the policy away from a reference model:

• SFT: supervised fine-tuning on demonstrations sets the base instruction-following behavior.
• RLHF (PPO): a learned reward model scores outputs, and PPO optimizes the policy under a KL penalty toward the reference.
• DPO: optimizes human preferences directly with a closed-form loss, removing the separate reward model and RL loop.
• KTO / IPO / ORPO: alternatives using unpaired feedback, different loss geometry, or preference-aware SFT.
• RLAIF & Constitutional AI: AI-generated preferences and explicit principles reduce reliance on human labels.
• GRPO / RLVR: group-relative optimization and RL from verifiable rewards drive reasoning models using checkable outcomes rather than a learned reward model.

For systems with measurable success criteria, code execution, math, retrieval accuracy, verifiable-reward RL is increasingly preferred because the reward signal is grounded rather than approximated. Process reward models that score intermediate steps add further control over long reasoning chains. The practical lever is β, the KL strength: too high and the model stays capable but unaligned, too low and it collapses toward the reward model's blind spots.

A systems takeaway

Treat policy, harness, and RL steering as one pipeline. Regulation sets what you may deploy, the harness governs how it behaves at runtime, and RL steering decides how it behaves intrinsically. Teams that engineer all three deliberately, rather than optimizing the model in isolation, are the ones whose systems stay both capable and controllable.