Benzene

How it works

One ring, three scales

A cell walks six atoms for every instruction. A colony of cells heals, evolves, divides and dies. An organisation founds six offices from those colonies, in order, when the record earns them. Same ring at every scale; only the titles change.

The ring inside one cell

The CEO's seat is the instruction. From there the cell walks the ring in a fixed order; the light on the home page is where the work is.

COO
orchestration: define the problem, plan the steps with dependencies, run them alone or one session per skill in parallel waves, reflect after the verdict
CTO
skills and tools: a portable SKILL.md is chosen by trigger words, then a router model, then efficacy; tools write to the workspace and ask an operator
CIO
the model: any provider, priced per role (primary, judge, router, evolver, scout, decider), with fallbacks for free tiers
CFO
guardrails: measurable text checks, then executable verifiers, then a judge blended at a weight; a failed verifier caps the score at 0.40
CXO
feedback: a guardrail efficacy and a world efficacy, both exponential moving averages; the worse one governs
CMO
viability: verified work is offered to the life system for feedback; the rest stays in the heal loop

The life cycle

  • Heal. A failed guardrail sends the verifier output back to the model, up to heal_max_attempts times under a cost ceiling. A failed guardrail also sends the skill to the evolver for repair.
  • Evolve. Below evolve_below (0.80) the evolver proposes one mutation; it is accepted only when the eval set scores it better within a tolerance. Principles per virtue may be refined, never dropped.
  • Autophagy. Junk in the genome (mutations without gain, stale feedback, bloat) is cleaned before anything harsher.
  • Divide. Below divide_below (0.66) with two or more skills, the cell splits: the strong skills to one daughter, the weak to the other. The scout defines each daughter's guardrails.
  • Renew, then die. A cell that cannot divide is rebuilt from its best iteration; a cluster under 0.33 dies and regenerates from the best of every previous generation, a limited number of times. Then it is extinct.

The general approach: six offices, founded in order

An organisation is one purpose and six office colonies under .benzene/offices/. Nothing is founded on a clock. Each office is founded when the record meets its rule.

OfficeFounded whenChargeMeasures
CTOat the startbuilds the product; its verifiers are the product's own checksverified builds
COOthe CTO cluster is healedtakes the instructions Jev routes to it, answers with a JSON plan held to a schema, dispatches tickets with acceptance criteria across offices in waves, on a boardthroughput
CMOthe COO has one verified planwrites the campaign from the record; every number must be on the recordreach
CXOthe launch milestoneturns feedback into a perspective; convenes a Product Hunt panelsatisfaction
CFOrevenue or investment on the recordprices every office; its balance sheet must reconcile with the ledgerrevenue
CIOusers or satisfaction on the recordkeeps the investors' ledger; convenes a Shark Tankinvestment

Chartered offices. The CTO's genome is written by hand. The other five are chartered at founding: a built-in charter specialised for the purpose by the evolver model, with the scout proving their first guardrails on the workspace. With a researched standard on the organisation, every office is founded with the competition's checkable bars already in its guardrail, the number of cells the onboarding decided, and the cadence its duty runs at; the organisation's phase (healing, prepared, publishing, evolving) says how far it is from the bar and whether an audience exists yet for feedback to steer.

Where an instruction goes: Jev decides. With a decider model on the manifest (TypeSafe Jev, typesafe/jev-router on OpenRouter) every instruction is one typed choice: the COO for a plan across offices, or one active office, each described by its charge and its cells' skills; inside the office a second choice picks the cell. A calibrated confidence comes back with the pick, and a pick under 0.5 is not trusted: the COO plans. Live: the CLI crashes on init, fix it went to the CTO at 1.00, rewrite the launch campaign to the CMO at 0.99, plan the next release: docs, a campaign and a pricing page to the COO at 0.70. The report says who decided, and a route event lands on the record. Without a decider, the COO takes everything once it exists.

The treasury

The CFO atom is counted, not described. A verified pass earns value_usd × score; every task spends its model cost; revenue reported with feedback is earned by the cell that did the work. The organisation's financial guardrail says too early, growing or burning against a minimum ROI once enough has been spent to mean anything.

The ladder

Sixteen rungs from self to universe emperor. At the micro scale the CTO is "the engineer"; at the universal scale, "infrastructure of the universe". scale: c-suite on the manifest picks the titles; the ring is the same.

# manifests/org/skills-cli.yaml
name: skills-cli
purpose: Build the best front-end architecture CLI and earn from it
scale: c-suite
offices:
  - {office: cto, manifest: skills-cli-cto.yaml}
  - {office: coo, founded_when: {after: cto, stage: healed}}
  - {office: cmo, founded_when: {after: coo, verified_tasks: 1}}
  - {office: cxo, founded_when: {milestone: launch}}
  - {office: cfo, founded_when: {any_metric: {revenue_usd: 1, investment_usd: 1}}}
  - {office: cio, founded_when: {any_metric: {users: 100, satisfaction: 0.7}}}
board: {kind: local}          # or linear
evaluations:
  cxo: {format: product_hunt, every: 3}
  cio: {format: shark_tank, every: 5}
heartbeat:
  approve: [feature]
  monitor_beats: 5