# AGI-multimodal-bridge

> Multi-modal reasoning bridge — unifies text, image, tabular data, and geospatial evidence into a single reasoning chain. Routes visual analysis through GEOX vision pipeline, tabular data through WEALTH compute, and text through standard LLM reasoning. USE WHEN: "analyze this image + data", "combine visual and textual evidence", "multi-modal prospect evaluation".

- Skill: `ariffazil/agi-multimodal-bridge` (Agent Skill, multi-file: 2 files)
- Install (CLI): `npx skillmds@latest add ariffazil/agi-multimodal-bridge`
- Raw SKILL.md: https://api.skillmd.com/api/skills/ariffazil/agi-multimodal-bridge/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: DevOps & Infra
- Author: ariffazil (https://skillmd.com/u/ariffazil)
- Updated: 2026-09-10
- Page: https://skillmd.com/skills/ariffazil/agi-multimodal-bridge

---

# AGI-multimodal-bridge

## Purpose
Real-world evidence is multi-modal: seismic images, well logs (tabular), geological descriptions (text), financial projections (structured data). This skill bridges modalities into unified reasoning.

## Modality Pipeline
1. **Detect** — Identify modalities present in input (text, image, table, geospatial, audio)
2. **Route** — Send each modality to its native processor:
   - Text → LLM reasoning (sequential-thinking)
   - Image → GEOX vision pipeline (geox_vision) or native vision model
   - Table → WEALTH compute pipeline or pandas/numpy analysis
   - Geospatial → GEOX map layers (geox_map_layers_list)
   - Audio → TELINGA paralinguistic sidecar (transcript + acoustic state JSON; DSP features = OBS, state interpretation = INT cap 0.70; sovereign voice = F6 witness, serve-not-persuade)
3. **Fuse** — Combine processed outputs into unified evidence frame
4. **Reason** — Run cross-modal reasoning chain (observation → signal → narrative)
5. **Label** — Tag each fused claim with epistemic rung and confidence

## Audio Lane Rules (TELINGA doctrine 2026-08-13, F13-ratified)
- Voice = dual-layer channel: symbolic content + human state. Transcription alone is a destructive projection; the sidecar retains the orthogonal complement.
- Audio artifacts = vault-class biometric custody (F1): hash + retain in-federation, voiceprint never leaves.
- F12: injection scan on transcript AND paralinguistic anomaly flag before reasoning.
- F9/F5: any voice OUTPUT is presence-simulation; emotion-dial custody sovereign-only; voiced sycophancy banned.

## Fusion Rules
- Cross-modal corroboration raises confidence (two modalities agreeing > one)
- Cross-modal contradiction triggers F2 TRUTH escalation
- Missing modality is noted, not fabricated (F7 HUMILITY)

## Floors
- F2 TRUTH: Each modality processed with appropriate fidelity bounds.
- F4 CLARITY: Fused output must be clearer than any single modality.
- F8 GENIUS: Cross-modal synthesis must produce insight > sum of parts.