DECISION ARCHIVE / SELECTED INDEX

FIELD NOTES.

Ein öffentliches Archiv von Arbeitsregeln, offenen Fragen und dokumentierten Rejections. Vollständige Einträge trennen Kontext, Beobachtung, Evidenz, Bewertung, Unsicherheit und nächsten Proof. Die Nummern sind Ordnung, kein Ranking.

FIELD NOTES.

A demo is not proof.

DONE is not a filename.

Truth before narrative.

Measure before optimize.

Reduce until the decision becomes visible.

No assumption gets promoted to fact.

Uncertainty must stay visible.

A system must be able to say no.

A documented rejection is a result.

Reality gets the final vote.

Authority must be explicit.

Permission and capability are not the same.

Execution is different from intention.

Real state matters more than process completion.

Independent verification closes the loop.

A frontend is a view, not truth.

Read models are not authority.

Version numbers do not create truth.

BLOCKED / UNKNOWN beats invented green.

Traceability beats memory.

Market opinion is not market science.

An edge must survive data.

An edge must survive costs.

An edge must survive execution.

Forward walk matters.

Almost proven is still unproven.

12,083 pairs. Required 242. Observed 237. Shortfall 5.

When proof fails, permission stays off.

Intrabar damage and close confirmation are different events.

Origin structure and current structure are different references.

Session rules and signal rules should be tested separately when possible.

Can the existing structure logic survive multi-day holding periods?

Can free proxy data reject a thesis before expensive data is needed?

Fees can kill a beautiful backtest.

No live permission from research aesthetics.

AI is an execution layer, not a truth source.

Prompting is not system architecture.

Tools without rules can increase chaos.

AI agents need explicit authority.

An agent needs permission boundaries.

Context should reduce repetition, not remove control.

Human judgment remains accountable for consequential decisions.

Automation should remove cognitive load.

The best AI workflow has a visible handoff.

Execution needs proof after the agent says done.

How much work can an agent own without becoming authority?

Can personal context become useful without becoming invasive?

Codex is strongest with a defined target and repository context.

Chat is for thinking. Execution belongs close to the code.

Can an AI-native personal operating system remain understandable at scale?

The problem is often fragmentation, not missing features.

A new interface does not create a new category.

Real differentiation must be structural.

Old categories do not become new because the landing page says so.

An MVP must demonstrate the mechanism.

Do not build the SaaS monster before the wedge works.

One painful workflow can be a stronger product than twenty features.

Empty states should tell the truth.

No fake customers. No fake metrics. No fake readiness.

A product should earn complexity.

Can Role OS reduce the coordination tax around ownership and authority?

Can property administration become a measurable trust system?

Can legal complexity become operationally understandable?

Can cultural visibility become useful infrastructure?

Can a child feel authorship before learning the technology behind it?

Trust should be inspectable where possible.

Can trust become infrastructure?

Can security proofs travel without raw data?

Network effects are worthless if the shared signal is weak.

Privacy is an architecture decision, not footer copy.

Can Liechtenstein become a useful proving ground for trust systems?

A small network does not remove the need for proof.

Collective detection is only useful if correlation is explainable.

Transparency without boundaries can reduce security.

Can proof itself become a shareable protocol?

Output > Talk.

One next executable action.

No wild refactors.

Do not overwrite what already works.