AgentStake vs. The Alignment Playbook

The Lineup

There are four dominant approaches to AI alignment right now. Three come from AI labs. One comes from economics.

Approach	Core Idea	Origin
RLHF	Train models on human preferences	OpenAI, Anthropic, Google
Guardrails	Rule-based filters on inputs/outputs	Every lab + open-source
Constitutional AI	Self-critique against written principles	Anthropic
AgentStake	Economic penalties for misbehavior	AgentStake

They're not mutually exclusive. But they solve different problems — and fail in different ways.

---

RLHF

Reinforcement Learning from Human Feedback

How it works

Human evaluators rank model outputs. The model is fine-tuned to prefer higher-ranked responses. Over time, it learns to produce outputs humans rate as "good."

Strengths

• Dramatically improved helpfulness and coherence
• Scales with compute (more training = better alignment)
• Industry standard — battle-tested across major models

Weaknesses

• Goodhart's Law: Optimizes for rated behavior, not actual good behavior
• Lab ≠ deployment: Evaluators rate in controlled settings; real users are adversarial and creative
• Static: Training happens once; the world changes constantly
• No recourse: When an RLHF-trained agent misbehaves, there's no accountability or compensation
• Reward hacking: Models learn to appear aligned rather than be aligned

The gap: RLHF tells an agent what "good" looks like. It doesn't give the agent a reason to stay good when no one's watching.

---

Guardrails

Input/Output Filters

How it works

Rules applied before and after model inference. System prompts define boundaries. Output filters catch harmful content. Rate limits prevent abuse. Content classifiers flag violations.

Strengths

• Fast to implement
• Deterministic — same input hits same rule
• Easy to audit and update
• Works as a first line of defense

Weaknesses

• Rules are finite, exploits are infinite: Every jailbreak proves this
• Cat-and-mouse: New attack → new rule → new attack → forever
• Brittle: Edge cases break hard-coded rules
• No context: Can't distinguish malicious intent from legitimate use
• Performance cost: Aggressive filtering kills capability

The gap: Guardrails are walls. They keep out known threats. But agents that operate autonomously face situations no rule anticipated. When the wall fails, there's no backup.

---

Constitutional AI

Anthropic's CAI

How it works

Give the model a "constitution" — a set of principles (be helpful, be honest, avoid harm). The model critiques its own outputs against these principles and revises before responding.

Strengths

• Self-correcting — model catches its own mistakes
• Principled — alignment comes from explicit values, not just preference data
• Reduces dependence on human evaluators
• More robust than pure RLHF

Weaknesses

• Principles are interpretable: The model decides what "avoid harm" means — and it might decide wrong
• Self-critique has limits: A model can't catch biases it doesn't know it has
• Still training-time: Constitution is baked in during training, not enforced at runtime
• No external accountability: If the model misapplies its principles, who corrects it?
• Philosophical brittleness: Edge cases where principles conflict require judgment the model may not have

The gap: CAI is the most sophisticated training-time approach. But it's still a model talking to itself. There's no external force ensuring the principles are actually upheld in production.

---

AgentStake

Economic Alignment

How it works

Agents (or their operators) stake collateral before operating. Good behavior earns rewards. Misbehavior triggers slashing — stake is seized and victims are compensated. Trust becomes an economic asset.

Strengths

• Runtime enforcement: Alignment is checked continuously, not just at training time
• Skin in the game: Misbehavior has direct financial cost
• Recourse: Harmed parties get compensated
• Observable trust: Stake amount and reputation are public, verifiable signals
• Self-correcting market: Delegators route capital toward trustworthy agents
• Model-agnostic: Works with any AI system, any framework, any chain

Weaknesses

• Requires infrastructure: Smart contracts, dispute resolution, oracle systems
• Cold start: New agents have no reputation; bootstrapping trust takes time
• Dispute resolution is hard: Who decides if an agent misbehaved?
• Not a replacement for training: A poorly trained agent with stake is still poorly trained
• Economic attacks: Well-funded adversaries could stake heavily, exploit once, and absorb the loss

The gap: AgentStake doesn't make agents smarter or more capable. It makes the consequences of their behavior real. It's not alignment through understanding — it's alignment through incentives.

---

Head-to-Head

Dimension	RLHF	Guardrails	CAI	AgentStake
When it works	Training	Runtime (filters)	Training	Runtime (continuous)
What it optimizes	Human preferences	Rule compliance	Self-consistency	Economic outcome
Adversarial robustness	Low	Low	Medium	High
Accountability	None	None	None	Stake + slashing
Victim recourse	None	None	None	Compensation
Adaptability	Retraining	Manual updates	Retraining	Market-driven
Model-agnostic	No	Partially	No	Yes
Observable trust	No	No	No	Yes (on-chain)

---

The Real Comparison

This isn't about picking one. It's about understanding the layers:

Every serious system needs all three. Training sets the baseline. Guardrails handle the obvious. AgentStake handles everything else.

The question isn't "RLHF or AgentStake?" It's "Why would you deploy an agent with Layer 1 and 2 but skip Layer 3?"

---

The Analogy

Think about how we trust humans in high-stakes roles:

No one says "this doctor went to a great school, so we don't need malpractice insurance." All three layers work together.

The agent era deserves the same rigor.

---

Conclusion

RLHF, guardrails, and constitutional AI are all necessary. None are sufficient.

They work at training time or through static filters. They don't provide runtime accountability. They don't compensate victims. They don't create observable, verifiable trust.

AgentStake fills that gap. Not by replacing what labs have built, but by adding the layer they can't: economic consequences.

Alignment isn't a training problem or an incentive problem. It's both. We're building the incentive half.

---