The Arai-eek Zine Library is a mycelial network of 80+ PDF variants spanning 20 years. To navigate this mass, we deployed a Human-Agentic Workflow—a synthesis of human curatorial intent and machine-scale narrative processing.
Following the initial stabilization, we optimized the pipeline into the Standard Archival Protocol (SAP) v2. This revision focuses on two critical bottlenecks: ID collision and high-fidelity visual ingestion for "analog-only" zines.
During the 2026-05-08 session, we explicitly moved away from concise bullet points back to Narrative Depth for Machine Extractions. We found that the longer, descriptive summaries provide better "Speculative Archiving" quality, capturing the materiality and nuance that bullets erase.
To ensure ID consistency across a messy 20-year history, we integrated a Foolproof ZID Audit command directly into the agent protocol. This prevents ZID collisions by performing a real-time regex audit of the database before any new entry is drafted.
Our two-step pipeline allows for scalable archival by differentiating between machine-readable text and visual-heavy curation.
| Archival Path | Workflow Components | Est. Metabolic Cost | Efficiency Rating |
|---|---|---|---|
| Rapid Ingestion (✅ Machine-readable) |
pdftotext + DeepSeek Synthesis | ~$0.05 / zine | 🚀 HIGH (Mass Ingestion) |
| VIP Vision (⚠️ Image-heavy) |
Visual Sampling (6+6) + Multimodal Analysis | ~$0.85 / zine | 👁️ PRECISION (Curated) |
The session began by deploying Claude Opus 4.6 (Thinking mode) to analyze the complex 20-year directory structure. Opus did the heavy lifting of writing the foundational `process_zines` ingestion workflow and establishing the initial semantic parameters.
To optimize for speed and cost over 1,500+ iterations, orchestration was handed to Gemini 3 Flash. Flash acted as the high-speed technical conductor—writing the Python build engines, executing bash scripts, and running local `pdftotext` extraction pipelines.
Instead of relying on the primary IDE model for massive narrative generation, Gemini 3 Flash delegated the "Speculative Dossier" and "Factual Summary" creation to DeepSeek-V4 via the Model Context Protocol (MCP).
DeepSeek was instructed to first generate an objective, structure-heavy "Factual Analysis" based strictly on extractable text, followed by a Cyber-Tropical "Poetic Synthesis" that extrapolates meaning for the visual chaos.
For the final integration of the Vis.js mindmap, the CRT-style CSS, and the complex modal state management, the IDE was switched to Gemini 3.1 Pro (High) to ensure high-fidelity UI execution across the portal.
Archiving 20 years of biological chaos requires significant cognitive expenditure. We track the metabolic burn across the entire stack, balancing high-reasoning conceptual work with high-efficiency iterative loops.
| Archival Session | Model Stack | Estimated Tokens | Est. Cost (USD) |
|---|---|---|---|
| Session 1 (2026-05-06) Initial Setup & Consolidation |
Opus 4.6, Flash, Pro, DeepSeek | ~14.4M Input / ~400k Output | ~$25.07 |
| Session 2 (2026-05-08) SAP v2 & VIP Implementation |
Antigravity (Pro), DeepSeek-V4 Pro | ~3.5M Input / ~80k Output | ~$5.01 |
| Cumulative Project Burn | Multi-Agent Stack | ~17.9M Tokens | ~$30.08 |