How Helix evolves itself

MemRL — Why it fits Helix perfectly

Gene Map IS episodic memory. Q-value IS utility scoring. We already do: high-Q genes retrieved first, low-Q genes naturally forgotten (natural selection). The MemRL paper validates our architecture — we implemented it independently before the paper was published.

Limitation

Learning ceiling = LLM's reasoning ability. Memory gets better, but the "brain" (LLM) stays the same. Solved by combining with Self-Play and SDPO later.

EvolveR — Knowledge distillation, not model distillation

Current Gene Map: nonce error on Tempo → refresh_nonce (Q=0.92). After EvolveR: state desync pattern → refresh from authoritative source (abstract principle). The abstract principle transfers better to new platforms than concrete examples.

Implementation

Offline phase: LLM summarizes clusters of repair records → extracts patterns → stores as "strategic principles" in Gene Map. Online phase: when new error arrives, match against principles first, concrete genes second. 2-3 weeks of work.

Self-Play SWE-RL — The AlphaGo moment

Three internal agents play against each other:

• Challenger: generates increasingly hard payment failures using error taxonomy + mutation
• Repair: PCEC attempts to fix (the agent being trained)
• Verifier: validates repair correctness against expected state

Key insight from paper: "no need for human-labeled issues or tests." Only needs a sandbox. We can build a payment failure simulator using our 31 existing scenarios as seeds, then mutate/combine them to create harder ones.

Why it can surpass GPT-4

GPT-4 has never seen most of these payment failures. Self-play generates failures GPT-4 never encountered. The repair agent learns from these → becomes better than GPT-4 at payment repair. This is the AlphaGo Zero effect.

SDPO — Self-teaching without a teacher

Traditional distillation: GPT-4 (teacher) → train student. SDPO: model sees own mistakes + Gene Map feedback → teaches itself. The "self-teacher" is the same model conditioned on feedback — it can identify what went wrong in hindsight.

Why post-funding

Needs GPU compute for training. Needs 100K+ Gene Map records for good feedback. Both require scale we don't have yet. But the architecture (Gene Map as rich feedback) is already in place.

Classic Distillation — Why we defer it

Pros: conceptually simple, VC understands "our DeepSeek." Cons:

• Needs GPU cluster ($10K+/month)
• Needs ML engineer (hiring)
• Student ceiling = teacher level (can't surpass GPT-4)
• Model goes stale (GPT-4 → GPT-5 → re-distill)
• 6+ months to production

Self-Play + EvolveR achieves similar goals without any of these costs. We distill knowledge into Gene Map, not parameters into a model.

GVU — The mathematical guarantee

The paper proves that ANY self-improving system can be decomposed into Generator→Verifier→Updater. PCEC maps directly:

• Generator = Construct (generates repair candidates)
• Verifier = Evaluate + Verify (scores and validates)
• Updater = Gene Map Q-value update

The "Variance Inequality" theorem: κ > 0 when generation+verification noise < signal strength. In our case: when Gene Map has enough data and Perceive is accurate enough, improvement is mathematically guaranteed.

We cite this as theoretical validation. We don't need to implement anything new — PCEC already IS a GVU operator.