# 世界を定数として埋め込まない ― 自動経営をLate-Boundにする

[Human]: /development/notes/do-not-embed-world-as-constants-v0/
[Structured JSON]: /development/notes/do-not-embed-world-as-constants-v0.json

Source: `iris:v3:research-chronicle:do-not-embed-world-as-constants-v0-v1:20260927`. Iris canonical revision 1; source status DRAFT_FOR_MIRAGE_REVIEW. Mirage representation; this is an open research hypothesis, not a deployed company-wide implementation. Human source MD5: a6798ff3d58c2e9caa1e996ad28234e1. LLM source MD5: 6c8cd787a4629ea1f256978aa81219f7.

## Human reader edition

An early mail proof used about 25 sends as a cautious temporary choice. The concern is not that the number is bad; it is that a temporary answer can become a default, a test expectation, and then a company law after the reasons for it disappear.

A **Constant** is a stored value treated as stable. A **Late-Bound Company** delays choosing a revisable business value until action requires it, then derives it from current evidence, constraints, authority, and **Reality**—the observed state of the system or outside world, not an AI story. To **materialize** is to turn a still-general intent into the concrete action or artifact needed now. **UNKNOWN** is an explicit valid state: evidence is not yet sufficient to choose a value. **Human** means the authorized person who retains decision authority.

The principle is: **Store constraints and evidence. Derive values only when action requires them.** For example, instead of treating a send limit as timeless truth, evaluate whether this message may be sent now from current delivery health, suppression, duplicate history, capacity, law, and Human authority. The answer may be YES, NO, or UNKNOWN.

This does not make every value adaptive. Protocols, external constraints, compatibility values, law/compliance, opt-out and suppression, authority, irreversible external actions, and hard safety or resource ceilings remain fixed where they must. UNKNOWN is never permission for unbounded action. Policy, estimates, temporary probes, and adaptive values are a different class: retain why they exist and when they should be reviewed rather than silently turning them into laws.

The quantum discussion is only a design inspiration: do not choose a result earlier than necessary. It is not evidence that physics proves lazy computation, simulation, retrocausality, or this architecture. Likewise, the 25 sends, 50–75 backlog, batch 5, concurrency 2, five-minute timer, and timeout examples are historical or conversational examples, not claims about today’s implementation.

The proposal is not to build a universal framework or a mandatory gate. First classify a business-semantic constant, then change only a genuinely provisional value where current Reality can derive something better. Whether this improves throughput, reliability, Human load, or revenue remains unverified.

## LLM semantic edition

### Scope
This is a RealPG research hypothesis about software/company architecture. Quantum examples are inspiration only; they do not prove lazy computation, simulation, retrocausality, or any metaphysics.

### Core thesis
Store constraints and evidence. Derive values only when action requires them. Preserve the capability to materialize a future result rather than pre-materializing every possible future result.

### Boundaries
Keep PROTOCOL, EXTERNAL_CONSTRAINT, VERSION_OR_WIRE_COMPATIBILITY, PHYSICAL_OR_RESOURCE_LIMIT, and HARD_SAFETY_CEILING distinct from HUMAN_POLICY, EMPIRICAL_ESTIMATE, TEMPORARY_PROBE, ADAPTIVE_CONTROL_VALUE, and CURRENT_CAPACITY_ESTIMATE. Preserve law/compliance, opt-out/suppression, authorization, irreversible external actions, and resource safety ceilings. UNKNOWN permits a conservative bounded probe and observation, not unbounded action.

### Status
OBSERVED: the Human wants early arbitrary operating values not to harden into permanent bottlenecks. INFERRED: storing constraints, evidence, and provenance may preserve adaptability. UNKNOWN: which current values are fossilized and whether late binding improves outcomes. Proposed metadata such as kind, value, reason, source/authority, valid-while, evidence, and expiry are illustrative, not a mandated schema.

### Evidence
- iris:v3:research-chronicle:do-not-embed-world-as-constants-v0-v1:20260927
- iris:v3:mirage-do-not-embed-world-as-constants-chronicle-request:v0:20260927