CR / Hyper synergy

Code Republic is Tachyon's commercial substrate for agent-native open source. Hyper is the lab's ML Discovery Engine. They are separate tracks, joined by one content layer: design blocks, implementations, tests, eval traces, and maintainers.

Code Republic is not just useful for Hyper - it is necessary. The transformer revolution rode on fifteen years of primitive-implementation work paid for in CUDA, cuDNN, and PyTorch before any single discovery sat on top. AlphaEvolve and FunSearch ride on Google's internal substrate. We do not have that legacy. Code Republic is how we build it - agent-maintained, growing organically, sharpened by real use rather than one team's hand-effort. Without it, Hyper's implementation noise dominates the search and no discovery surfaces. Hyper, in return, returns evidence-backed blocks and sharpens Code Republic's own specialist models so the library does not become a catalog of yesterday's software.

Abstract flywheel linking Code Republic infrastructure with Hyper discovery outputs

The flywheel

Code Republic agent-native substrate Implementation memory blocks, code, tests, baselines Component graph finer than repo/package units CR custom models Retriever · Coach · AI Maintainer Implementation contract turn hypotheses into runnable, comparable experiments Shared decomposition stack Hyper isolates what caused the gain CR catalogues it as a reusable block Hyper ML discovery engine Discovery pressure tests ideas against reality Evidence stream scores, traces, scaling probes Reusable discoveries blocks, objectives, policies library + tests less noise evidence writeback distilled block + model upgrades return to CR CR makes Hyper implementable Hyper improves CR's ML models
Figure 1 - why Code Republic and Hyper belong together. Code Republic supplies blocks, canonical code, tests, and baselines that make Hyper's experiments comparable. Hyper returns empirically distilled blocks and upgrades for Republic's own models. The shared decomposition stack connects Hyper's ablations with Republic's component boundaries.

The two directions of value

Code Republic → Hyper

Republic makes the concept layer executable. A conceptual genome is free-form prose until it becomes runnable code; the CR library supplies known-good implementations, retrieval over relevant blocks, tests, baselines, and maintainers that keep reused components from rotting. That lowers implementation noise enough that the search signal is visible at all, and lets Hyper spend compute on search rather than on rebuilding the same pieces.

Hyper → Code Republic

Hyper turns frontier ML search into library supply. A winning candidate should not disappear into a paper or a one-off repo. Hyper ablates it first: which part caused the gain, which scaffolding was accidental, which constraints are required, and which implementation details are replaceable? The distilled result lands in Republic as a design block, and when mature, as a package with tests, eval traces, provenance, and an AI maintainer.

How this kicks in across the stages

The full flywheel is a steady state. The proposal is staged, and the writeback from Hyper to Code Republic ramps up across the stages, not all at once.

Stage 1 - mostly one direction

Hyper calibrates on a deterministic niche where no LLM prior exists, then lifts onto low-compute ML benchmarks. Code Republic supplies the implementation library that makes the loop spin at all. The writeback to Code Republic at this stage is loop infrastructure - the implementation bridge code, the auto-tuning machinery, the canonical-implementation patterns - rather than ML blocks; the early niche-level discoveries (game strategies, principles) are not natural Republic content.

Stage 2 - the flywheel turns

The loop ramps to ML. Hyper produces evolved components at small-LLM scale and the writeback becomes ML supply. Code Republic's three specialist models - Retriever, Coach, AI Maintainer - are narrower targets than general post-transformer ML and have direct product impact, so they are a likely first home for early wins. The proposal does not commit which model produces the first win; it commits to running the loop where the wins are most plausible first, and to honest reporting of what lands where.

Stage 3 - validate and scale

The loop validates and scales the strongest discovered methods through scaling-law studies, at the largest scale our compute envelope allows. Distilled methods land as design blocks in Code Republic and as the basis for the post-NFAI follow-on raise needed to reach frontier-level models. The library that started as agent-native open-source for everyday code now also catalogues post-transformer methods with provenance, ablations, and scaling-curve evidence.

What stays separate

Level Role
Tachyon The lab: the public research organization and company.
Code Republic The commercial substrate: agent-native open source, Coach, AI Maintainer, Retriever, and reusable blocks.
Hyper The research engine: searches ML concepts, runs experiments, ablates discoveries, and emits evidence-backed blocks.
Shared layer Design and package blocks with canonical code, tests, eval traces, provenance, and maintenance policy.

What users eventually feel

A Code Republic user does not need to understand Hyper's internal loop. They should feel the result as better reuse: an agent reaches for a memory-efficient sequence model, a safer maintainer policy, or a stronger retrieval architecture, and Republic can surface a block that has already been implemented, evaluated, distilled, and maintained.

That is the commercial point of the science track. Hyper keeps the library from being only a catalog of past software; Republic keeps Hyper from being only a stack of isolated experiments.

Architecture - Hyper's discovery loop and sub-systems.
Code Republic - the agent-native open-source substrate.