Experimental LLM Modes & Plans

This page is the human-facing overview of Iron Curtain’s LLM-related modes and plans for:

• players
• spectators and tournament organizers
• modders and creators
• tool developers

Everything here is design-stage only (no playable build yet) and should be treated as experimental. Some items are “accepted” decisions in the docs, but that means “accepted as a design direction,” not “implemented” or “stable.”

BYOLLM = Bring Your Own LLM. Iron Curtain does not require a specific model/provider. You can use IC’s built-in local models (CPU-only, zero setup), sign in to a cloud provider, connect your own local inference server, or paste an API key — whatever fits your setup.

For agentic retrieval / RAG routing, use LLM-INDEX.md. This page is for humans.

Ground Rules (Applies to All LLM Features)

• Optional, never required. The game and SDK are designed to work fully without any LLM configured (D016).
• BYOLLM architecture, built-in floor. The engine supports four provider tiers: IC Built-in (embedded CPU models, zero setup), Cloud OAuth (browser login), Cloud API Key (paste key), and Local External (Ollama, LM Studio, etc.). Users choose their tier; the engine does not mandate a vendor. IC Built-in provides a functional baseline; BYOLLM provides the ceiling (D047).
• Determinism preserved. ic-sim never performs LLM or network I/O. LLM outputs affect gameplay only by producing normal orders through existing pipelines (D044, D073).
• No ranked assistance. LLM-controlled/player-assisted match modes are excluded from ranked-certified play (D044, D073, D055).
• Privacy and disclosure matter. Replay annotations, prompt capture, and voice-like context features are opt-in/configurable, with stripping/redaction paths planned (D059, D073). Built-in models run entirely on-device — no data leaves the machine. Cloud providers are the user’s choice and the user’s responsibility.
• Standard outputs for creators. Generated content is standard YAML/Lua/assets, not opaque engine-only blobs (D016, D040).

Quick Map by Audience

Players

• LLM-generated missions/campaigns (optional) — D016
• LLM-enhanced AI opponents (LlmOrchestratorAi, experimental LlmPlayerAi) — D044
• LLM exhibition / prompt-coached match modes (showmatch/custom-focused) — D073
• LLM coaching / post-match commentary (optional, built on behavioral profiles) — D042 + D016

Spectators / Organizers / Community Servers

• LLM-vs-LLM exhibitions and showmatches with trust labels — D073
• Prompt-duel / prompt-coached events with fair-vs-showmatch policy separation — D073
• Replay download and review flows for LLM matches via normal replay infrastructure — D071 + D072 + D010

Modders / Creators

• LLM mission and campaign generation (editable YAML+Lua outputs) — D016
• Replay-to-scenario narrative generation (optional LLM layer on top of replay extraction) — D038 + D016
• Asset Studio agentic generation (optional Layer 3 in SDK) — D040
• LLM-callable editor tools (planned) for structured editor automation — D016
• Custom factions (planned) — D016

Tool Developers

• ICRP + MCP integration for coaching, replay analysis, overlays, and external tools — D071
• LLM provider management, routing, and prompt strategy profiles — D047
• Skill library-backed learning loops (AI/content generation patterns) — D057

Player-Facing LLM Gameplay Modes

1. LLM-Enhanced AI (Skirmish / Custom / Sandbox)

Canonical: D044

Two designed modes:

• LlmOrchestratorAi (Phase 7)
  • Wraps a normal AI
  • LLM gives periodic strategic guidance
  • Inner AI handles tick-level execution/micro
  • Best default for actual playability and spectator readability
• LlmPlayerAi (experimental, no scheduled phase)
  • LLM makes all decisions directly
  • Entertainment/experiment value is the main point
  • Expected to be weaker/slower than conventional AI because of latency and spatial reasoning limits

Important constraints:

• not allowed in ranked
• replay determinism is preserved by recording orders, not LLM calls
• observable overlays are part of the design (plan summaries/debug/spectator visibility)

2. LLM Exhibition / Prompt-Coached / Showmatch Modes

Canonical: D073 (built on D044)

These are match-policy modes, not new simulation architectures:

• LLM Exhibition Match
  • LLM-controlled sides play each other (or play humans/AI) with no human prompting required
  • “GPT vs Claude/Ollama”-style community content
• Prompt-Coached LLM Match / Prompt Duel
  • Humans guide LLM-controlled sides with strategy prompts
  • The LLM still translates prompts + game context into gameplay orders
  • Recommended v1 path: coach + LlmOrchestratorAi
• Director Prompt Showmatch
  • Casters/directors/audience can feed prompts in a labeled showmatch context
  • Explicitly non-ranked / non-certified

Fairness model (important):

• ranked: no LLM prompt-assist modes
• fair tournament prompt coaching: coach-role semantics + team-shared vision only
• omniscient spectator prompting: showmatch-only, trust-labeled

Player-Facing LLM Content Generation (Campaigns / Missions)

3. LLM-Generated Missions & Campaigns

Canonical: D016

Planned Phase 7 optional features include:

• single mission generation
• player-aware generation (using local data if available)
• replay-to-scenario narrative generation (paired with D038 extraction pipeline)
• full generative branching campaigns
• generative media for campaigns/missions (voice/music/sfx; provider-specific)

Design intent:

• hand-authored campaigns (D021) remain the primary non-LLM path
• LLM generation is a power-user content expansion path
• outputs are standard, editable IC content formats

4. LLM Coaching / Commentary / Training Loop

Canonical: D042 (with D016 and D047 integration)

This is the “between matches” / “learn faster” path:

• post-match coaching suggestions
• personalized commentary and training plans
• behavioral-profile-aware guidance
• integration with local gameplay history in SQLite

D042 also supports the non-LLM training path; LLM coaching is an optional enhancement layered on top.

Spectator, Replay, and Event Use Cases

5. Replays for LLM Matches (Still Normal IC Replays)

Canonical: D010, D044, D073, D071, D072

LLM matches use the same replay foundation as everything else:

• deterministic order streams remain the gameplay source of truth
• replays can be replayed locally
• relay-hosted matches can use signed replay workflows (D007)
• server/dashboard/API replay download paths remain applicable (D072, D071)

What D073 adds is annotation policy, not a new replay format:

• optional prompt timestamps/roles
• optional prompt text capture
• plan summaries for spectator context
• trust labels (e.g., showmatch/director-prompt)
• stripping/redaction flows for sharing

6. Spectator and Tournament Positioning

Canonical: D073 + D059 + D071

IC distinguishes clearly between:

• fair competitive contexts (no hidden observer prompting/coaching)
• coached events (declared coach role, restricted vision)
• showmatches (omniscient/director/audience prompts allowed, clearly labeled)

This is a core trust/UX requirement, not just a UI detail.

Modder / Creator LLM Tooling (SDK-Focused)

7. Scenario Editor + Replay-to-Scenario Narrative Layer

Canonical: D038 + D016

The scenario editor pipeline includes a replay-to-scenario path:

• direct extraction works without an LLM
• optional LLM generation adds narrative layers (briefings, objectives wording, dialogue, context)
• outputs remain editable in the SDK

This is useful for:

• turning replays into challenge missions
• creating training scenarios
• remixing tournament games into campaigns

8. Asset Studio Agentic Generation (Optional Layer)

Canonical: D040 (Phase 7 for Layer 3)

Asset Studio is useful without LLMs. The LLM layer is an optional enhancement for:

• generating/modifying visual assets
• in-context iterative preview workflows
• provenance-aware creator tooling (with metadata)

This is explicitly a creator convenience layer, not a requirement for asset workflows.

9. LLM-Callable Editor Tool Bindings (Planned)

Canonical: D016 (Phase 7 editor integration)

Planned direction:

• expose structured editor operations as tool-callable actions
• let an LLM assist with repetitive editor tasks via validated command paths
• keep the editor command registry as the source of truth

This is aimed at modder productivity and SDK automation, not live gameplay.

10. Custom Faction / Content Generation (Planned)

Canonical: D016

Planned path for power users (built-in models work; external providers unlock higher quality):

• generate faction concepts into editable YAML-based faction definitions
• pull compatible Workshop resources (subject to permissions/licensing rules)
• validate and iterate in normal modding workflows

This is a planned experimental feature, not a core onboarding path for modders.

Tooling & Infrastructure That Makes LLM Features Practical

11. LLM Configuration Manager

Canonical: D047

Why it exists:

• different tasks need different model/provider tradeoffs
• local vs cloud models need different prompt strategies
• users may want multiple providers at once
• non-technical players need a zero-config path to LLM features

Key planned capabilities:

• four provider tiers: IC Built-in (CPU models, zero setup), Cloud OAuth (browser login), Cloud API Key, Local External (Ollama, etc.)
• multiple provider profiles with tier mixing (built-in for quick tasks, cloud for quality)
• task-specific routing (e.g., built-in for coaching, cloud for generation)
• prompt strategy profiles (auto + override), including EmbeddedCompact for built-in models
• capability probing and prompt test harness
• shareable configs without API keys
• Workshop model packs for first-party and community-provided model weights

12. LLM Skill Library (Lifelong Learning Layer)

Canonical: D057

Purpose:

• store verified strategy/content-generation patterns
• improve over time without fine-tuning models
• remain portable under BYOLLM

Important nuance:

• this is not a replay database
• it stores compact verified patterns (skills), not full replays
• D073 adds fairness tagging so omniscient showmatch prompting does not pollute normal competitive-ish skill learning by default

13. External Tool API + MCP

Canonical: D071

ICRP is the bridge for external ecosystems:

• replay analyzers
• overlays
• coaching tools
• tournament software
• MCP-based LLM clients/tools (analysis/coaching workflows)

It is designed to preserve determinism and competitive integrity:

• reads from post-tick snapshots
• writes (where allowed) go through normal order paths
• ranked restrictions and fog filtering apply

Experimental Status & Phase Snapshot

This page is a consolidation of planned LLM features. Most of the LLM-heavy work clusters in Phase 7.

Competitive Integrity Summary (Short Version)

If you only remember one thing:

• LLM features are optional
• LLM gameplay assistance is not for ranked
• spectator prompting is only acceptable in explicit showmatches
• fair coached events must declare the coach role and vision scope

This is the line that keeps the LLM experimentation ecosystem compatible with IC’s competitive goals.

Canonical Decision Map (Read These for Details)

Core LLM Features

• D016 — LLM-generated missions/campaigns and BYOLLM architecture
• D042 — behavioral profiles + optional LLM coaching loop
• D044 — LLM-enhanced AI (LlmOrchestratorAi, LlmPlayerAi)
• D047 — LLM configuration manager (providers/routing/profiles)
• D057 — LLM skill library
• D073 — LLM exhibition and prompt-coached match modes

Creator / Tooling / Replay Adjacent

• D038 — scenario editor (includes replay-to-scenario pipeline; optional LLM narrative layer)
• D040 — Asset Studio (optional agentic generation layer)
• D071 — external tool API / ICRP / MCP
• D072 — server management (replay download/admin surfaces)
• D059 — communication/coach/observer rules (important for LLM showmatch fairness)
• D010 — replay/snapshot foundations

Suggested Public Messaging (If You Want a One-Paragraph Summary)

Iron Curtain’s LLM features are an opt-in, experimental layer for content generation, AI experimentation, replay analysis, and creator tooling. Built-in CPU models provide a zero-setup starting point; users who want higher quality can connect their own cloud or local providers (BYOLLM). The engine is fully playable and moddable without any LLM configured. Competitive integrity remains intact because ranked play excludes LLM-assisted modes, and showmatch/coached LLM events are explicitly labeled with clear trust and visibility rules.