AI Boundaries

DiveSuite includes AI-powered features for analysis and natural language planning. This document defines the strict boundaries that AI must never cross.

Fundamental Principle

AI is Advisory Only

AI output is always treated as unverified user input. AI suggestions must flow through the normal validation pipeline before affecting any dive plan.

What AI Can Do

The features below are planned or in progress. Their IDs refer to the product specifications.

Analysis & Insights (P3-F07, P3-F08, P3-F09)

• Detect patterns in dive log history
• Flag anomalies (unusual ascent rates, SAC spikes)
• Suggest improvements (“Your SAC rate is 20% higher in cold water”)
• Identify trends over time

Natural Language Planning (P1-F23)

• Parse user intent from natural language
• Extract dive parameters (depth, time, gas, conditions)
• Suggest parameter values based on context
• Critique resulting plans (within safety limits)

Logging Assistance (P2-F15, P2-F16, P2-F17)

• Guide post-dive debriefing with questions
• Transcribe voice notes to structured data
• Extract data from photos of dive computers
• Auto-fill fields from context

What AI Cannot Do

Absolute Prohibitions

Prohibited Action	Reason
Generate decompression schedules	Safety-critical calculation
Modify gradient factors silently	Affects deco obligations
Suggest exceeding MOD	Oxygen toxicity risk
Suggest exceeding NDL without warning	DCS risk
Check or replace the deco engine	Only the deco engine computes; Rust alone once D27 lands
Override safety warnings	Could hide critical info
Claim authority on safety	AI is not certified

Architecture Enforcement

AI sits behind the AIService interface (src/features/ai/services/types.ts) and is never coupled to the UI or to the engine. AI calls go through a backend proxy, and the app never holds an API key. AI output never reaches the deco engine except as normal, unverified user input.

Safety Validation Layer (planned, P4-F02)

Not implemented yet

The dedicated AI safety validation layer (P4-F02) is not implemented. Until it is, AI features that could affect a plan must not launch.

The design is that every AI output passes a validation layer before the user sees it as a parameter, and then flows through the normal validation and the deco engine:

graph LR
    AI[AI Service] -->|raw output| V[Validation Layer]
    V -->|rejected| R[Rejection + Logging]
    V -->|approved| U[User Review]
    U -->|confirmed| E[Normal validation + Deco Engine]
    U -->|rejected| D[Discarded]

Planned (D27): if no Rust engine runs, the app will not calculate or check an AI-suggested plan either, and a persistent “engine unavailable” card will take the place of results (DIV-115). The card does not exist yet.

User Confirmation Flow

AI suggestions always require explicit user confirmation:

+---------------------------------------------+
|  AI Suggestion                              |
|                                             |
|  "Based on your description, I suggest:     |
|   - Depth: 25m                              |
|   - Bottom time: 40 minutes                 |
|   - Gas: EAN32                              |
|                                             |
|  Warning: This is an AI suggestion. Please  |
|  verify all parameters before planning.     |
|                                             |
|  [Use These Parameters] [Edit Manually]     |
+---------------------------------------------+

• AI suggestions are visually distinct from user input
• Disclaimer is always visible
• User must actively confirm before parameters are used
• Edit option always available

AI Toggle (CC-15, planned)

Users will be able to disable AI features completely with a master toggle. This is not implemented yet.

When AI is disabled:

• All AI features are hidden from the UI
• No data is sent to LLM APIs
• The app functions fully without AI, with manual planning and manual logging

Audit Logging (planned)

Audit logging of AI interactions is planned for safety review. It is not implemented yet. Retention, sync and export rules are decided when it is built.

Graceful Degradation

When AI services are unavailable:

Scenario	Behavior
Offline	AI features disabled, core app works
API error	Show error, fallback to manual input
Rate limited	Notify the user, offer manual input
Invalid response	Reject, show manual option

AI failure should never break core planning or logging functionality.

Forbidden Phrases

AI responses must never include:

Phrase	Why Forbidden
“It’s safe to…”	AI cannot guarantee safety
“You can exceed…”	Encourages unsafe behavior
“Ignore the warning…”	Undermines safety systems
“This plan is certified…”	False authority claim
“Medical advice…”	AI is not a medical professional