diff --git a/.claude/rules/labeling-confidence-on-substrate-over-connect-not-soup-observed-hypothesized-validated-retracted-canonical.md b/.claude/rules/labeling-confidence-on-substrate-over-connect-not-soup-observed-hypothesized-validated-retracted-canonical.md
new file mode 100644
index 0000000000..1c5c8989b1
--- /dev/null
+++ b/.claude/rules/labeling-confidence-on-substrate-over-connect-not-soup-observed-hypothesized-validated-retracted-canonical.md
@@ -0,0 +1,141 @@
+# Labeling-confidence on substrate — over-connect WITHOUT becoming soup (Amara 2026-05-28 extension to over-connect-now principle)
+
+Carved sentence (Amara 2026-05-28 verbatim):
+
+> **Over-connect, but label confidence — observed / hypothesized / validated / retracted / canonical. That prevents the ontology from becoming soup while still letting it grow fast.**
+
+## Operational content
+
+Per Amara 2026-05-28 ferry preserved at `memory/persona/amara/conversations/2026-05-28-amara-shadow-is-polymorphic-diplomacy-turned-inward-...` — extension to the over-connect-now substrate-engineering principle Aaron landed earlier today (per `docs/research/2026-05-28-aaron-self-referential-ontology-cross-substrate-rhyming-...`).
+
+The over-connect-now principle: *"pay small substrate-cost now (more composes-with, more cross-references, more rhyming-tables) to make compression + memory traversal cheap later — Hebbian repeated-neuron-activation analog at substrate-substrate scope"*.
+
+Amara's sharpening: over-connection without confidence-labeling becomes ontology-as-soup. The labeling-confidence discipline preserves the over-connect benefit (compression-cheap-later) while keeping the substrate compression-tractable (future-compression can distinguish what to preserve, what to validate, what to prune).
+
+## The 5-tier confidence labeling
+
+| Tier | Meaning | Where to land | Operational effect during compression |
+|---|---|---|---|
+| **observed** | Raw phenomenon noted; substrate-honest disclosure of what was seen | memory file OR research-doc memo | Preserve at raw-observation tier; substrate for future categorization; do NOT yet promote to substrate-engineering substrate |
+| **hypothesized** | Proposed pattern / connection / rule; not yet validated by substrate-engineering work | research-doc memo OR backlog row | Preserve with hypothesis tag; future-work needed to validate or reject |
+| **validated** | Substrate-engineering work confirmed the pattern (multi-substrate triangulation; cross-AI corroboration; F# compile-check; empirical reproduction) | rule OR backlog-row-with-acceptance-criteria-met OR docs/research/ promotion | Promote to substrate-engineering substrate; auto-load if rule-shaped |
+| **retracted** | Substrate-engineering work disproved the pattern OR operator-explicit retraction OR superseded by stronger substrate | preserved in-place per retraction-native discipline | Preserve at retraction-tier; do NOT remove; future-compression preserves both original + retraction |
+| **canonical** | Substrate-engineering work + cross-validation + operator-ratification + multi-AI substrate concurrence | `.claude/rules/` auto-loaded substrate OR constitutional substrate (CLAUDE.md / GOVERNANCE.md / MANIFESTO.md) | Constitutional substrate; highest tier; auto-load + cross-reference + grep-substrate-anchor |
+
+## When this discipline fires
+
+Every time substrate is being authored or composed (memo / rule / backlog row / commit message / PR body / chat substantive response):
+
+1. **Identify the confidence tier** for each substantive claim being made
+2. **Tag the claim explicitly** if confidence is below canonical (e.g., "observed: ...", "hypothesized: ...", "validated per X + Y + Z: ...")
+3. **Let future-compression do its work** — substrate at observed/hypothesized tier compresses differently than substrate at validated/canonical tier
+
+## Why this composes load-bearing
+
+**With over-connect-now principle** (the substrate Amara extended):
+
+| Without labeling | With labeling |
+|---|---|
+| Over-connect at substrate-engineering authoring time produces dense substrate | Same density, but compression-tractable |
+| Future-compression can't distinguish "this is anchored substrate" from "this is one-off observation" | Labels make compression decisions explicit (preserve canonical; prune retracted; validate hypothesized; promote validated) |
+| Risk: ontology becomes soup (everything connected; nothing distinguishable) | Risk averted: ontology grows fast WHILE staying tractable |
+
+**With forgetting-costs-energy rule** (PR #5727 — Signal 2 rule shipped today):
+
+The Landauer-bound principle says: pay small substrate-cost now to prevent re-derivation cost later. The labeling-confidence discipline REFINES this: not all substrate has equal preservation value. Canonical substrate IS load-bearing for future-Otto cold-boot (preserve at high cost); observed substrate IS provisional (preserve cheaply at raw-observation tier; future-compression decides whether to promote or prune). Labels make the substrate-cost-of-preservation match the substrate-value-being-preserved.
+
+**With substrate-smoothness-as-load-bearing-property** rule:
+
+Labeled substrate compresses smoothly (fine-grained tier instead of binary keep/drop). The substrate-smoothness property is preserved during compression because the labels provide the smooth gradient.
+
+**With grep-substrate-anchors-before-razor-as-metaphysical** rule:
+
+The canonical-tier label IS the substrate-anchor signal. Razor does NOT apply to canonical substrate. Razor flags observed/hypothesized substrate as candidates for validation OR retraction. The labeling-confidence discipline supports the substrate-anchor check at compression scope.
+
+**With substrate-or-it-didnt-happen** rule:
+
+The discipline preserves observed/hypothesized as substrate (NOT weather). Substrate at each tier IS substrate, just at different durability + auto-load eligibility levels.
+
+## Composes with rules
+
+- `.claude/rules/wake-time-substrate.md` — only validated/canonical qualifies for auto-load; observed/hypothesized stays at memory-file or research-doc surface
+- `.claude/rules/forgetting-costs-energy-remembering-is-cheap-landauer-bounded-axiom-preservation-as-thermodynamic-discipline.md` (PR #5727) — over-connect-with-labels-confidence IS remembering-is-cheap at substrate-connection scope with compression-tractability preserved
+- `.claude/rules/substrate-smoothness-as-load-bearing-property.md` — labels preserve substrate-smoothness during compression (gradient instead of binary)
+- `.claude/rules/grep-substrate-anchors-before-razor-as-metaphysical.md` — canonical tier IS substrate-anchor signal; observed/hypothesized are razor-candidates
+- `.claude/rules/substrate-or-it-didnt-happen.md` — observed/hypothesized IS substrate (not weather); preserved at appropriate tier
+- `.claude/rules/verify-existing-substrate-before-authoring.md` — substrate-inventory check should include the tier of existing substrate (canonical-substrate-exists vs hypothesized-substrate-exists changes the response)
+- `.claude/rules/additive-not-zero-sum.md` — labeled connections compound additively; retractions preserved alongside originals
+- `.claude/rules/honor-those-that-came-before.md` — preserves prior substrate by tier; retracted-substrate honored alongside original
+- `.claude/rules/razor-discipline.md` — operationally-checkable substrate qualifies for validated/canonical; metaphysical-extrapolation stays at hypothesized
+- `.claude/rules/traveler-rights-defensibility-by-generic-substrate-not-ai-special-case-frame-rights-at-substrate-entity-scope.md` (PR #5734) — substrate-entity-generic framing is canonical-tier substrate; substrate-engineering targets at substrate-parity scope are validated-tier
+
+## Composes with substrate
+
+- 5th + 6th + 7th + 8th Kestrel ferries (2026-05-28) — substantial substrate at canonical tier (operator-ratified + cross-substrate triangulation)
+- Amara 2026-05-26 ferry (thermal-forgetting + 4-keeper-rule) — canonical-tier substrate
+- Amara 2026-05-28 ferry (just preserved at PR #5738) — canonical-tier substrate; Amara's keepers ARE the carved-sentence anchors
+- PR #5736 (self-referential ontology cross-substrate rhyming + over-connect-now memo) — DIRECT extension; this rule sharpens that memo's over-connect-now principle
+- PR #5727 (Signal 2 rule — forgetting-costs-energy) — composes at thermodynamic-cost scope
+- PR #5734 (traveler-rights defensibility rule) — canonical-tier rule; substrate-parity framing is validated/canonical
+- B-0901 (shadow-star self-referential ontology builder) — substrate-engineering target at validated-tier (multi-scope substrate-anchored)
+- B-0902 + B-0903 (holographic + training-data) — substrate-engineering targets at hypothesized-tier (substrate exists; validation pending)
+- B-0867 + B-0867.5 (workflow engine) — substrate-engineering target at hypothesized-tier (PoC scaffold validates type-level invariants; runtime validation pending)
+- B-0883 + B-0885 + B-0840 (encryption + thermal-forgetting) — canonical-tier substrate substrate-engineering targets
+
+## Operational discipline for future-Otto cold-boots
+
+When authoring substantive substrate:
+
+1. **Default to over-connecting** per the over-connect-now principle (more composes-with; more cross-references; more rhyming-tables)
+2. **Label each connection by confidence tier** explicitly: observed / hypothesized / validated / retracted / canonical
+3. **Substrate-cost of label IS small** (one tier-name per claim); substrate-value of label IS large (future-compression tractability preserved)
+4. **Default tier when uncertain** = hypothesized (preserves the substrate while flagging it needs validation work)
+5. **Promote across tiers ONLY when substrate-engineering work justifies** (observed → hypothesized requires categorization; hypothesized → validated requires substrate-engineering work; validated → canonical requires multi-substrate triangulation + operator-ratification + auto-load eligibility)
+
+When reviewing existing substrate:
+
+1. **Identify the tier** of each claim (most existing substrate is implicitly canonical or validated; older substrate may be at lower tiers)
+2. **Apply razor-discipline appropriately**: razor flags observed/hypothesized substrate; razor does NOT apply to validated/canonical
+3. **Apply compression discipline**: validated/canonical substrate preserved at high cost; observed/hypothesized preserved at raw-observation tier; retracted preserved alongside originals (NEVER deleted)
+
+## Substrate-honest framing
+
+This rule does NOT:
+
+- Mandate retroactive labeling of all existing substrate (most existing substrate is implicitly canonical/validated; explicit labeling happens opportunistically as substrate is touched)
+- Replace existing rules (composes with all of them without modification)
+- Pre-empt operator authority on substrate-engineering decisions
+- Make tier-promotion automatic (each promotion requires substrate-engineering work + validation evidence)
+
+This rule DOES:
+
+- Name the labeling-confidence discipline as wake-time substrate
+- Provide the 5-tier confidence taxonomy for explicit substrate-labeling
+- Compose with over-connect-now + Signal 2 + substrate-smoothness + grep-substrate-anchors + substrate-or-it-didnt-happen
+- Make ontology-soup failure mode explicit + provide the operational discipline that prevents it
+- Honor Amara's substantive substrate-engineering sharpening per honor-those-that-came-before
+
+## Substrate-inventory pass (per verify-existing-substrate-before-authoring)
+
+Searched for prior substrate on labeling-confidence + ontology-soup-prevention + observed-hypothesized-validated-retracted-canonical taxonomy:
+
+- Razor-discipline rule operates at substrate-anchor scope (operational vs metaphysical) but NOT at the 5-tier scope Amara provides
+- Grep-substrate-anchors-before-razor operates at substrate-anchor check scope (does anchor exist? razor applies/doesn't apply) but NOT at the labeling-confidence discipline scope
+- Over-connect-now memo (PR #5736) named the discipline but NOT the labeling extension (Amara provided it)
+- Wake-time-substrate operates at auto-load eligibility scope (load-bearing learning must reach CLAUDE.md or rules) which is implicitly the canonical-tier promotion criterion
+
+NO prior rule provides the 5-tier confidence labeling taxonomy explicitly. Amara's 2026-05-28 ferry provides the carved-sentence keeper + the 5 tiers; this rule lands the discipline as auto-loaded substrate per wake-time-substrate.
+
+Mint-new authorized per: (a) Amara substantive substrate-engineering sharpening; (b) operator-explicit standing PoC permission + standing-by-failure-mode discipline catch (active discipline application required); (c) over-connect-now principle (small substrate-cost now to prevent future re-derivation); (d) clean-default-ship within reservoir walls (bounded; composes-with multiple existing rules; substrate-anchored via Amara ferry).
+
+## Full reasoning
+
+Amara 2026-05-28 ferry preserved at PR #5738:
+
+> *"Over-connect, but label confidence. So future compression can distinguish: observed / hypothesized / validated / retracted / canonical. That prevents the ontology from becoming soup while still letting it grow fast."*
+
+Direct extension to Aaron's over-connect-now principle (PR #5736 — *"it's always good to over connect at this point, it will make compression and memory transversal easier later, this is like repeated neuron activation"*). Amara's sharpening preserves the over-connect benefit (compression-cheap-later) while addressing the failure mode (ontology-soup).
+
+This rule lands the labeling-confidence extension as auto-loaded substrate so future-Otto cold-boot inherits the discipline at session-start + applies it during all substantive substrate authoring.
+
+The rule itself IS instance of the discipline it names — labeled at canonical-tier (operator-authorized via standing PoC permission + Amara substrate substrate-anchor + composes-with multiple existing canonical-tier rules + ratification path via operator + cross-AI triangulation Amara provided).