AI & Data Policy

1. What AI does in Koyu

Koyu uses generative AI when users explicitly start architectural plan generation or revision. AI proposes muro files and summaries of design decisions. Koyu's deterministic software then parses the descriptions, derives plans, 3D views, areas and diagnostics, and saves and displays them. Regulatory screening is also an auxiliary check against recorded conditions.

AI does not approve designs or provide building confirmation, permits, safety certification or professional judgments. Users and the qualified professionals involved are responsible for final decisions and actual use.

2. Information sent during generation

Depending on the generation or revision, the following information may be sent to commercial AI providers:

• Selected site geometry, address or region and related planning conditions
• Design policies, prompts, revision instructions and generation settings
• The current muro file and previous conversation needed to revise the same plan
• Koyu generation instructions, muro specifications, diagnostics, repair context and regulatory and terminology information
• Models used, reasoning depth, speed and other technical settings

Koyu limits transmission to what is necessary for the requested processing. If personal information, confidential information, client materials, unpublished plans or third-party protected materials are included, users must confirm the necessity, legal basis and authority for Koyu and the listed processors to handle them.

3. Current AI provider

Koyu currently uses Anthropic's commercial API for generation and revision. Anthropic states that content from commercial customers is not used to train models unless those customers explicitly opt in. Koyu does not opt User Content into Anthropic model training.

The provider's explanation is available in Anthropic's training data policy for commercial products. If changes to the provider, contract or settings materially alter this boundary, we will update this policy before starting the new use.

4. Generation records we retain

Koyu stores generated plans and execution records that allow generation to be traced. Execution records may include design policies, revision instructions, models, reasoning depth, speed, tokens and costs, attempts and timestamps, diagnostic codes, changed layers and descriptions, applied and rejected edits, design decisions and reasoning summaries, completion status and subsequent revisions.

These records are used to return results without losing them, keep usage limits and refunds accurate, investigate incidents and failures, identify recurring problems and evaluate quality and safety. Records are separated by team-level database access controls. Retention and deletion follow the Privacy Policy.

5. Internal quality and safety evaluation

Authorized personnel may review necessary generation records on a limited basis to diagnose failures, provide support, test security and safety, identify recurring diagnostic gaps and improve prompts, rules, model selection and service behavior. Routine analysis prioritizes aggregated metrics and excludes or pseudonymizes direct identifiers and precise site information unless needed for the purpose.

Internal evaluation is not model training or further training. Merely generating, saving, diagnosing or reviewing an individual incident does not add User Content to Koyu's training datasets.

6. Choices required for a future Koyu Loop

Koyu Loop is under consideration as a voluntary mechanism for using contributed examples to evaluate and improve Koyu's design intelligence. Before accepting contributions, the following must be displayed and recorded:

• A separate choice that is not preselected, with no disadvantage to ordinary use if declined
• The authorized contributing account or team administrator, choice timestamp, consent policy version and dataset version
• The exact scope contributed from design policies, generated files, revisions, evaluations, diagnostics, selection results and other information
• Specific methods, such as evaluation, retrieval, fine-tuning or preference learning
• Target models or model families, architectural purposes, access scope, retention periods and third-party recipients
• How to withdraw raw contributions from future use and the technical limitations concerning already trained models

Terms of Service, privacy notices, pre-bundled checkboxes, silence and continued use are not treated as consent to Koyu Loop. A new choice is requested if the purposes, dataset, models or recipients materially change.

7. Protection before inclusion in a dataset

If Koyu Loop is implemented, the following will be done before contributed examples enter a training dataset:

• Record contribution authority, provenance, permitted purposes and dataset lineage
• Exclude direct account identifiers and precise addresses and coordinates as a general rule
• Screen for confidential information, unnecessary personal information, harmful data and rights restrictions
• Separate raw contributions from curated training data and restrict access by role
• Record selection and preparation so bias, duplicates, failures and deletion requests can be tracked
• Test memorization and reproduction, privacy leakage, dangerous architectural advice, systematic bias and performance regressions before release

Contributors must have the right to provide the materials. Participation does not transfer ownership of the original project. Permissions needed for training are limited to the purposes and scope shown on the choice screen.

8. Output transparency and human review

Koyu is designed to distinguish information recorded by users, deterministically derived by software and proposed by AI. AI outputs may contain errors, omissions, outdated information, bias, lack of originality or hazards. An absence of warnings is not evidence that architecture is lawful or buildable.

Before actual use, appropriate professionals must check the site, dimensions, areas, regulations, accessibility, structure, fire prevention and evacuation, building services, buildability, costs and other impacts against the source materials. Where AI-assisted results are suspected of disadvantaging a person, Koyu investigates the role of AI using available project and execution records, provides an appropriate explanation and corrects service behavior where justified.

9. Governance and incident response

When changing models, providers or purposes, Koyu defines the purpose and affected people; assesses privacy, intellectual property, security, bias and architectural safety risks; and establishes pre-release testing, monitoring, traceability and routes for inquiries and correction. Uses that may affect human safety or rights are subject to stricter controls than concept exploration.

10. Public guidance informing this policy

This policy refers to the following public guidance around the shared principles of human-centered design, safety, fairness, privacy, security, transparency, accountability and risk management:

• METI and MIC — AI Guidelines for Business, Version 1.2
• Personal Information Protection Commission — Caution regarding the use of generative AI services
• IPA — Guidelines for the Introduction and Operation of Text-Generating AI
• Agency for Cultural Affairs — AI and Copyright
• NIST AI Risk Management Framework and Generative AI Profile
• OECD AI Principles — Transparency and explainability
• European Commission — EU AI Act overview
• UNESCO — Recommendation on the Ethics of Artificial Intelligence

11. Changes and inquiries

For material changes to models, providers, training purposes or collected generation records, we review this policy and update its document version before starting the new use. New Koyu Loop purposes require a separate choice.

Inquiries, incident reports and objections to AI-assisted results may be sent to our contact desk; privacy requests may be sent to our contact desk.