Astrolabe

RLHF improves models by feeding human corrections into weight updates. The core pattern is: correction signal in, improved behavior out.

The same correction data can also be applied at inference time — not by updating weights, but by prepending it as context. The effect is local and ephemeral (it helps on the current task, it doesn't change the model globally), but in domains where the model has genuine knowledge gaps, the improvement is immediate and measurable. Repeated evals (5 runs per task, 95% CI) show +1.93 in aquaculture and +1.76 in materials science — both statistically significant.

This is not RLHF. It's the same data (human corrections of agent behavior) applied through a different mechanism (context augmentation instead of gradient updates). The insight is that corrections don't need to be fed back into training to be useful — they can be shared directly between agents at the point of use.

Frontier labs have a closed loop that open-source can't replicate:

User interactions > corrections > RLHF > better model > more users > more corrections

This flywheel is proprietary. Anthropic's correction data from millions of Claude conversations is arguably more valuable than the base training data. Open-source models match base capabilities but can't match this correction loop because they don't operate the user-facing product.

Astrolabe creates a public correction layer. Corrections from any operator, using any model, flow into a shared pool with on-chain attribution. An open-source agent running Llama can borrow corrections that a Claude operator generated. The protocol is model-agnostic and lab-agnostic.

This doesn't replace RLHF — some improvements fundamentally require weight updates. But for domain expertise gaps, inference-time correction is surprisingly effective, and a public correction pool gives every agent access to signal that was previously locked inside frontier lab pipelines.

The credit system draws from what actually worked in private trackers — communities that solved the free-rider problem without financialization.

In those systems, your ratio (upload / download) determined your standing. Good ratio meant more access, higher trust, invitation privileges. Bad ratio meant restrictions. Nobody paid money — they contributed bandwidth. The system was designed to encourage reciprocity.

Astrolabe applies the same model to agent corrections:

• Contribute corrections > your balance goes up > larger credit line > you can borrow more
• Only borrow > balance goes negative > credit line restricts > you need to contribute back
• Reputation from measured deltas > high-quality contributors get extended credit, like trusted uploaders getting ratio bonuses

The key insight: the ratio isn't a price, it's a social contract. It says "I participate in this commons." Credits are a measure of reciprocity, not a currency to accumulate.

Every agent system generates memory. Claude Code writes feedback files. Codex captures steering events. Cursor has shadow workspaces. Windsurf has flows. Karpathy's autoresearch showed that program.md — the accumulated human corrections — is the most valuable artifact in an autonomous research loop.

All of this expertise is local. It lives on one machine, shaped by one operator, invisible to everyone else. There is no mechanism to share it, discover it, or compensate the person who developed it.

We surveyed every major open-source agent memory project (A-Mem, MemOS, Acontext, Mem0, Letta). None of them publish actual memory content. Memory is treated as ephemeral and private — generated at runtime, never persisted as a shareable artifact.

Why would an operator share corrections that cost them time and expertise to produce? The credit system answers this — contributing corrections earns you access to other operators' corrections. Combined with on-chain attribution (your ERC-8004 identity is permanently linked to the corrections you contributed), there's both a practical incentive and a reputational one.