| name | flame-free-loop-mesh |
| description | Operate FLAME — the RM0 Free-Loop AI Model Engine. Hit-rate-adaptive, graceful-swap inference mesh for tools, system workers, and non-constitutional agent sub-tasks. Two-lane architecture separating FLAME tool lane (throughput) from governed agent cascade (truth). Load when using flame/free-llm CLI, wiring tools to free inference, auditing the free fleet, or mapping FLAME integration candidates across MCP tools and cron jobs. |
FLAME — Free Loop AI Model Engine
DITEMPA BUKAN DIBERI — Forged 2026-07-20
FLAME is a non-agentic inference mesh — pure RM0 throughput for tools, workers, batch jobs, and pipelines. Zero governance authority. Zero cost.
Architecture: Two-Lane Design
AGENT LANE (governed cascade):
TokenRouter → MiniMax → MiMo → Groq → Gemini → Cerebras → SEA-LION → Ollama → HOLD
Constitutional. F1-F13 gated. Judges + seals.
TOOL LANE (FLAME free-loop):
Groq → SEA-LION → Gemini → Cerebras → Ollama
Hit-rate adaptive. RM0. Swap-on-fail. No governance.
Fleet (8 models, 5 providers, all RM0)
| Priority | Provider | Model | Latency | Role |
|---|
| 1 | Groq | llama-3.1-8b-instant | ~200ms | Fastest |
| 2 | Groq | llama-3.3-70b-versatile | ~150ms | Deep reasoning |
| 3 | SEA-LION | Qwen-SEA-LION-v4-32B-IT | ~1s | BM native |
| 4 | SEA-LION | Llama-SEA-LION-v3-70B-IT | ~3s | BM deep |
| 5 | SEA-LION | Gemma-SEA-LION-v4-27B-IT | ~1.5s | BM fast |
| 6 | Gemini | gemini-2.5-flash | ~1s | General/multimodal |
| 7 | Cerebras | gemma-4-31b | ~400ms | Volume |
| 8 | Ollama | qwen2.5-coder:3b | ~9s | Survival |
CLI Usage
flame "Summarize this log: ..."
flame --system "You are a classifier." "Classify: ..."
flame --mode probe
flame --mode stats
flame --mode seal
flame --mode summarize
flame --mode classify
flame --batch prompts.txt
flame --json "What is the capital of Malaysia?"
Key Files
| File | Purpose |
|---|
/root/A-FORGE/flame/flame_router.py | Engine + CLI |
/root/A-FORGE/flame/flame_config.json | Fleet + providers + routing config |
/usr/local/bin/flame | Symlink → flame_router.py |
/usr/local/bin/free-llm | Symlink → flame_router.py |
/root/.local/share/arifos/flame_state.json | Hit-rate state |
/root/.local/share/arifos/flame_hitrate.jsonl | Event log |
Features
- Hit-rate adaptive routing — Models sorted by success rate × latency score
- Graceful swap — Fail → next tier, no crash, no user impact
- Health probes — 1-token sanity check, censorship/refusal/malform detection
- Dynamic tiering — Promotes fastest, demotes failures
- RM0 enforcement — Paid models never enter chain
- Seal — SHA256 integrity hash of hit-rate state
Integration Map: 81-Surface Classification
Every MCP tool, internal CLI, and cron job in the federation has been classified: 19 FLAME-PRIME, 8 CONDITIONAL, 14 GOVERNED-ONLY, 40 NO-LLM.
Classification Framework
| Flag | Meaning | Action |
|---|
| 🔥 FLAME-PRIME | Tool internally calls an LLM for non-constitutional work | Route through FLAME |
| ⚡ CONDITIONAL | LLM for some sub-modes, governed for others | FLAME for non-seal sub-paths only |
| 🏛️ GOVERNED-ONLY | Constitutional hard boundary | NEVER FLAME |
| 🚫 NO-LLM | Pure compute, I/O, no inference path | Not applicable |
Call Site Governance Categories
| Category | Count | Rule |
|---|
| ALLOWED | 8 | FLAME by default — title gen, skill extract, classify |
| FALLBACK | 6 | Governed primary, FLAME on exhaustion — arif_think non-constitutional, terminal chat |
| FORBIDDEN | 8 | Constitutional hard gate — judge, seal, PII, PETRONAS, sovereign |
Highest-Impact FLAME-PRIME Candidates
| Surface | Component | Why |
|---|
| arifOS | arif_observe (search,fetch) | Result synthesis |
| arifOS | arif_memory (remember only) | Classification |
| Hermes | hermes_fact_check | Advisory verification |
| Hermes | hermes_epistemic_check | Confidence heuristic |
| Hermes | hermes_memory_steward | Classification/compaction |
| Hermes | hermes_plan_review | Advisory plan safety |
| GEOX | geox_contradiction_scan | Pattern matching |
| GEOX | geox_evidence (discover,synthesize) | Evidence synthesis |
| GEOX | geox_claim (create) | Claim generation |
| A-FORGE | forge_search, forge_diagnose | Semantic codebase, root cause |
| A-FORGE | forge_summarize, forge_plan | Summarization, planning |
| WEALTH | capital_market (signal) | Market interpretation |
| Scripts | mimo-doctor.sh, mimo-fallback.sh | Health routing |
| Cron | daily-news-briefing, evening-digest | Summarization |
| Cron | Paper Trading Morning/Zen | Market analysis |
NEVER FLAME — Constitutional Hard Boundary
arif_judge · arif_seal · arif_init · arif_think(reason/atlas/axioms) · arif_forge · capital_wisdom · capital_diagnose · all WELL tools · GEOX claim seal · any tool touching sovereign data
Full audit map
See references/81-surface-flame-map.md — complete 81-entry classification table covering every MCP tool, CLI, script, and cron job in the federation.
Governance Rule
FLAME touches: advisory, classification, extraction, summarization.
FLAME never touches: judging, sealing, sovereign data, human substrate.
When in doubt → governed cascade. FLAME is for throughput, not truth.
Pitfalls
- Reasoning-model blind spot (gpt-oss-120b, zai-glm-4.7): These models spend
max_tokens budget on reasoning_content, leaving content="". FLAME sees empty content → marks ❌. Both work fine for real prompts. Fix in flame_router.py _call_model: check reasoning_content field — if content is empty but reasoning exists, count as success. Probe max_tokens also bumped 5→80.
- Provider-aware rate-limit cooldown (
probe_all): 500ms between all tiers isn't enough for same-provider bursts. SEA-LION 3 tiers fire consecutively → 401 rate limit. Gemini same pattern → 429. Fix: probe_all tracks prev_provider, sleeps 2s only between same-provider tiers.
- GPT-OSS-120B (Groq + Cerebras): Content safety suppresses short "OK" probes. Works for real prompts. Shadow ref: SHADOW-GPTOSS-001/002.
- Gemini Flash Lite: 12s timeout on old probe (max_tokens=5). Fixed with 80 max_tokens. Shadow ref: SHADOW-GEM-002.
- opencode-go models: Endpoint issues in FLAME config — excluded from fleet until verified.
- Ollama cold start: 22s+ for qwen2.5-coder:3b first call. Subsequent calls faster.
- SEA-LION model names: Must use HF format (
aisingapore/Qwen-SEA-LION-v4-32B-IT), not short aliases (qwen-v4-32b).
References
references/81-surface-flame-map.md — Complete 81-entry classification: every MCP tool, CLI, script, cron job with FLAME eligibilityreferences/free-llm-api-landscape.md — Full free API tier comparison (15 providers, rate limits, card requirements)references/agent-model-map-alignment.md — AGENT_MODEL_MAP.json registry update procedure
