Atlas Heritage Systems
Field Guides
Operational reference · Evergreen · Not protocol documents
Field guides are operational references that don't belong inside a protocol. They describe how to work with language models in practice — what prompts actually do, what context injection actually does, what failure modes look like from inside a session. These documents are not tied to a specific instrument. They apply across the suite.
Technician's Guide
Step-by-step operational guide for running FVE-1 sessions — what to do, in what order, every time. Covers the full session checklist, Tier A requirements, common mistakes, and lab tool reference.
V2.0 — updated for HTML tool pipeline and three-layer protocol structure.
The Technician's Read
The human operator function that lives at the seam between Layer 1 and Layer 3 of every Atlas instrument protocol. Read #0 before the session. Read #1 before logging. The gap between them is the data.
The timestamp is part of the data. The Read is not analysis. It is perception.
Human Coder Protocol
Governs the human coding pass for all FVE-1 instrument sessions — what the coder receives, what they produce, and what gates the output must clear before the session is eligible for downstream processing.
V1.0 — codebook dependency: ECM Resolution Code Coder Guide V4.0.
Prompt Framing Best Practices
Eight sections covering fundamentals, do's, don'ts, context injection, register fidelity, and adversarial architecture. Developed from 27+ revision cycles across multiple model families.
Living document — updates as practice develops.
Context Integrity
Three failure modes documented during Atlas development: Context Saturation Drift, Context Compression Bias, and Frequency-Weighted Distortion. Not hypothetical — encountered repeatedly and countermeasures developed in response.
Operational controls included for each failure mode.
Gaming the Domain Manager
A field note on negative-space literature pulls — how to redirect a model away from the consensus shelf and into the actual structure of a field. Includes verification discipline and ghost citation failure mode.
v1.1 — April 2026. The model is the instrument. You're still the one reading the gauge.