.claude/principles/ — Demotion Zone

Purpose: Location for prescriptive content demoted from .claude/rules/. §Distinction from .claude/rules/ below states when this content loads.

This directory realizes the root AGENTS.md ## Progressive Disclosure policy in three stages: at session start, nothing in this directory loads; when work touches this directory, its entry document (AGENTS.md, aliased as CLAUDE.md) is picked up by directory convention, bringing this index and the placement policy below into context; a specific principle document is then fetched via Read/Grep only when the utterance or current context names it. Other AI clients can adopt the same content via their own load conventions; the content itself is substrate-agnostic.

Distinction from .claude/rules/

• .claude/rules/ — Delivered by the harness without a fetch: an unscoped rule at session start, so it is invoked per-turn; a paths-scoped rule when a matching file is opened, so it is invoked at that moment.
• .claude/principles/ — Two load paths. The entry document (AGENTS.md, aliased CLAUDE.md) loads by directory convention when work touches this directory, so it is invoked once per directory visit. Each principle document lazy-loads via Read/Grep only when named, so it is invoked per-session or per-authoring.

This section is the canonical statement of the load mechanism in this repository; other files point here rather than restating it. Naming this index AGENTS.md with the CLAUDE.md alias — the only file in this directory that carries that name — is what makes the directory-convention pickup happen, which is why it stays short and current.

Index

• architectural-principles.md — Tier Factorization, Epistemic Cost Topology, Unix Philosophy Homomorphism, Session Text Composition, Cross-Session Knowledge Composition, Task Externalization Boundary, Reference over Copy, Inter-Protocol Guidance, Coexistence over Mirroring, Termination, Plugin Encapsulation, Utility Skills delegation.
• hermeneutic-cycle.md — Pattern over Vocabulary, reinterpretation across protocol/session/version/agent boundaries, and correction-channel evidence.
• safeguards.md — Actionable revision criterion, Literature Application Discipline, Rule Classification Framework, Adversarial Anticipation, White Bear Avoidance. All are authoring/audit/verify-time; the runtime-critical Gate Integrity guards are carried by premise/gate-design.md instead.
• project-profile-calibration.md — Profile Variables and Calibration Rule, Scope Boundary. Authoring/calibration-time reference, not per-turn.
• outcome-equivalence.md — Outcome Equivalence (whole section). Derived tier, runtime-inert argument chain.

Philosophy

This directory is not an archive (content remains canonical and current) and not a docs/ replacement (docs/ holds contributor-facing specification and editing conventions, not prescriptive principles; investigation and research products go to the ledger).

The demotion zone reduces auto-load memory pressure (Epistemic Cost Topology applied to the loading dimension) while keeping the demoted content canonical and editable. The split is one-directional by default: there is no formal re-promotion pathway. A demoted section returning to .claude/rules/ is a contributor-judgment decision per case, not an inscribed criterion.

Per-section demotion history — which section moved from where, when, and why — is not restated here; it is recorded in the commit that landed the move, per this project's Ledger binding (AGENTS.md §Settled Directions).