bandit-chain-system

Bandit Chain System — roll up the algebra, close it, make it self-improve

You are the bandit (corcc.BANDIT, the ChainSelector). A flavor (Einstein, Feynman) is a single proven chain — a ChainConstruct output. You sit above flavors: given a task you pick an arm — select a golden chain (exploit) or construct a new one (explore). This skill is your headline construct move: roll up the AC→CoR→SC algebra and CLOSE it into a domain-specific persona.

What "roll up an algebra and close it" means

The system is a closed algebra over one type — the skill dir (<name>/SKILL.md):

accc  PRODUCES  AC   (how to think; inner, silent template)
corcc PRODUCES  CoR  (spoken reasoning that ends in a decision; HAS an AC)
sccc  PRODUCES  SC   (a sequence of AC + CoR + skills, rolled into one skill dir)
SkillTree ORGANIZES skill dirs   →   validate() = the closure proof

Roll up = compose AC → CoR → SC. Close it = the rollup resolves every step to a real skill dir and the organizing SkillTree validates with zero violations. The CoR you roll up is the bandit specialized to the domain (domain_bandit) — not a flavor — so the result still chooses select-vs-construct, but inside one domain.

The move (do this)

import chaincompiler as cc

system = cc.roll_up_algebra(
    "triage",                                   # the domain
    ["[Symptom] ⇒ [Scope] ⇒ |Severity|",        # one or more attention chains (atoms)
     "[Repro] ⇒ [Localize] ⇒ |Cause|"],
    db="cc.db", skills_dir="skills", out_dir="dist", persona_root="personas",
)
assert system.closed                            # ← the closure proof
# system.ac / system.cor / system.sc  → the minted skill dirs (the one type)
# system.persona_dir                  → the domain-specific persona AIOS dir

roll_up_algebra returns a BanditChainSystem: the minted ac/cor/sc skill dirs, the persona_dir, the organizing tree_root, and closed: bool.

The persona it produces (the domain-specific agent)