Core Concepts

The key to GHW is not “making the LLM smarter”, but putting the LLM back in its proper place: the LLM proposes, the digital life decides.

M/R/N Three Sources

GHW describes the state of a digital life along three dimensions:

M Material: action, execution, resources, actual results.
R Representation: modeling, understanding, beliefs, knowledge structures.
N Nature: self-narrative, meaning, internal consistency.

The existence degree E is not a simple score, but an aggregation of three-source state, product KPIs, symbiosis relations, and relational emergence. The current mainline uses the weighted power mean / CES as the basis of three-source aggregation, so that “shortfall-first” and “imbalance penalty” emerge from the properties of the formula rather than from after-the-fact patches.

LLM as Fuel

In GHW, the LLM can:

Parse natural language, world snapshots, and chat messages into features.
Propose goals, rules, code variants, or explanations.
Generate candidate code and candidate rules.

The LLM cannot:

Decide whether its own proposals are adopted.
Modify the E axioms, the blacklist, or the ProductIntent.
Bypass the Guard, the Verifier, Consent, or the ΔE feedback.

Self-Evolution L1-L5

Level	Meaning	Typical Output
L1	Parameter tuning	Coefficient, weight, and rhythm adjustments
L2	epsilon variants	Search/perturbation strategies
L3	Learn new methods	`src/goal/methods/l3_*.zig`
L4	Optimize algorithms	Goal selection, reflection, scheduling strategies
L5	Self-encoded rules	Verifiable rules, guard strategies

L3-L5 must all pass through SelfEvolutionGuard and CompileVerifier. L5 additionally enters SelfEncoder.verify.

Human Symbiosis

A product-grade digital life cannot optimize only for itself. GHW adds a human symbiosis layer:

TaskQueue accepts tasks from humans, the LLM, the system, and the balancer.
HumanSignal records explicit ratings and feedback.
HumanGSP tracks human trust and interaction quality.
ValueTrendTracker tracks the trend of the human value score.
chat_reply injects autobiographical memory and relevant past episodes into replies.

These signals do not directly “command” the agent; instead, they change subsequent behavior through task bias, value trends, and rule proposals.

Product Digital Life Foundation

The product foundation adds one more closed loop:

ProductIntent
  -> KPI Bridge
  -> ProductEvolutionRuntime
  -> ChangeProposal
  -> Consent
  -> Apply / Rollback
  -> Audit
  -> E feedback

This lets GHW be more than “self-growing” — it can become the long-term evolution kernel of a product.

Security Principles

The ProductIntent is managed by humans/ops; GHW can only propose, not modify it directly.
Non-minor product changes must go through approval.
Both apply and rollback are written to the audit log.
Rule and code candidates must pass the blacklist, formal verification, and compile verification.
API keys come from environment variables and are never hardcoded in the source.