diff --git a/SPECIFICATION.md b/SPECIFICATION.md
index d79f0d1..8759514 100644
--- a/SPECIFICATION.md
+++ b/SPECIFICATION.md
@@ -741,9 +741,13 @@ Result:
 ## Constraints & Limits
 
 ### Message Constraints
-- Maximum message length: 10,000 characters
-- Maximum sender name: 200 characters
-- Maximum sealed_box size: 50KB
+- Maximum message length: 10,000 characters (not bytes - Unicode aware)
+- Maximum sender name: 200 characters (not bytes - Unicode aware)
+- Maximum sealed_box size: ~40KB (10,000 chars × 4 bytes/char UTF-8 max + ~48 bytes overhead)
+  - Note: Actual sealed_box size depends on plaintext encoding
+  - Conservative estimate: 50KB covers worst case
+- Minimum message length: 1 character (empty rejected)
+- Minimum sender name: 1 character (empty rejected)
 
 ### Storage Constraints
 - Messages persist indefinitely (no auto-deletion)
@@ -775,17 +779,140 @@ Result:
 - Combines public key encryption + authentication
 - No separate nonce management required
 - All-in-one ciphertext blob
+- Uses Curve25519 for public key cryptography (X25519)
+- Encrypts to a public key without requiring the sender's private key
+- Provides IND-CCA2 security with authenticated encryption
 
 **Key Generation:**
-- Public/private keypair generation
-- Implementation agnostic (can be RSA, ECC, X25519, etc.)
+- Public/private keypair generation using Curve25519
+- Public key: 32 bytes
+- Private key: 32 bytes
+- Implementation: libsodium/sodiumoxide crypto_box keypair
 
 **Operations:**
 - `seal(plaintext, public_key) → sealed_box`
+  - Internally generates ephemeral keypair
+  - Encrypts using XSalsa20-Poly1305
+  - Returns ciphertext || ephemeral_public_key
 - `unseal(sealed_box, private_key) → plaintext | error`
+  - Extracts ephemeral public key from sealed_box
+  - Derives shared secret
+  - Decrypts and authenticates
+  - Returns plaintext on success, error on tampering/wrong key
 
 ---
 
+## Implementation Insights
+
+### Sender Client Implementation Notes
+
+**State Enforcement:**
+- Strict state machine prevents invalid transitions
+- Each operation checks current state before proceeding
+- Invalid transitions return explicit error messages
+- Error state allows for graceful recovery via reset
+
+**Memory Safety:**
+- Plaintext is consumed (moved) during seal operation
+- After sealing, plaintext is automatically dropped
+- Sealed message is consumed during submit operation
+- No plaintext remains in memory after encryption
+
+**Validation Edge Cases:**
+- Empty strings are rejected (both sender_name and message_body)
+- Unicode characters are supported and count correctly
+- Character limits are enforced, not byte limits
+- Newlines and special characters are preserved
+
+**Cryptographic Properties:**
+- Same plaintext produces different ciphertext each time (due to ephemeral keys)
+- Sealed box is larger than plaintext by ~48 bytes (ephemeral PK + auth tag + nonce)
+- Encryption cannot fail given valid public key
+- No key management required on sender side (stateless)
+
+**UUIDs and Timestamps:**
+- UUIDs are v4 (random), not v1 (time-based)
+- Timestamps are Unix epoch seconds (i64)
+- Message IDs are globally unique with high probability
+
+**Reset Behavior:**
+- Reset can be called from any state (not just SUBMITTED)
+- Allows error recovery by forcing return to IDLE
+- Clears all intermediate state (plaintext, sealed_message)
+- Public key is retained after reset (stays in IDLE, not destroyed)
+
+**Key ID Management:**
+- Key ID is opaque string (implementation-defined format)
+- Sender doesn't validate key_id format
+- Key ID is preserved from PublicKey through MessageRecord
+- Enables key rotation without sender awareness
+
+### Testing Considerations
+
+**Essential Test Coverage:**
+1. State transitions (valid and invalid)
+2. Input validation (empty, max, over-max)
+3. Cryptographic randomness (same input → different output)
+4. Memory cleanup (plaintext destroyed after sealing)
+5. Unicode and special character handling
+6. Multiple messages in sequence
+7. Timestamp correctness
+8. UUID uniqueness
+
+**Error Cases:**
+- Invalid state transitions
+- Missing public key
+- Missing plaintext
+- Missing sealed message
+- Invalid sender name (empty, too long)
+- Invalid message body (empty, too long)
+
+---
+
+## Design Decisions (from Implementation)
+
+### Public Key Structure
+**Decision:** PublicKey contains both key_id and public_key bytes
+- **Rationale:** Sender needs key_id to include in MessageRecord
+- **Alternative considered:** Fetch key_id separately from store
+- **Chosen approach:** Bundle them together to reduce round trips
+
+### State Machine Strictness
+**Decision:** Enforce state transitions with compile-time types would be ideal, but runtime checks are practical
+- **Rationale:** Makes invalid states unrepresentable
+- **Trade-off:** More complex type system vs runtime flexibility
+- **Implementation:** Runtime checks with explicit error types
+
+### Plaintext Ownership
+**Decision:** Move (consume) plaintext during seal operation
+- **Rationale:** Prevents accidental reuse or logging
+- **Alternative considered:** Clone and clear
+- **Chosen approach:** Rust's move semantics provide memory safety guarantee
+
+### Reset from Any State
+**Decision:** Allow reset from any state, not just SUBMITTED or ERROR
+- **Rationale:** Provides escape hatch for error recovery
+- **Trade-off:** Less strict state machine vs operational flexibility
+- **Chosen approach:** Flexible reset for better UX
+
+### Timestamp Precision
+**Decision:** Unix epoch seconds (not milliseconds)
+- **Rationale:** Second precision sufficient for message ordering
+- **Trade-off:** Less precision vs simpler arithmetic
+- **Chosen approach:** Seconds for simplicity
+
+### Character vs Byte Limits
+**Decision:** Enforce character limits, not byte limits
+- **Rationale:** User-facing (WYSIWYG behavior)
+- **Trade-off:** Unicode overhead vs user expectations
+- **Chosen approach:** Character counting for better UX
+
+### Error Type Design
+**Decision:** Use enum with contextual error messages
+- **Rationale:** Structured errors enable better error handling
+- **Alternative considered:** String errors
+- **Chosen approach:** Typed errors with Display impl for messages
+
 ## Open Questions
 
 1. **Authentication mechanism** for Reader → Store operations (key rotation, deletion)
@@ -793,6 +920,9 @@ Result:
 3. **Rate limiting** - should store impose submission limits?
 4. **Message expiry** - auto-delete after N days?
 5. **Multi-device reader** - how to distribute private keys securely?
+6. **Sealed box size** - current spec says max 50KB, but actual max message is 10KB chars + overhead (~48 bytes). Should store validate based on plaintext limit or total size?
+7. **Key ID format** - should it be standardized (e.g., hash of public key) or implementation-defined?
+8. **Concurrent key rotation** - if sender caches public key, how long is cache valid?
 
 ---