Analysis Models

Analysis Models

All analysis models are optional, local adapters. They create evidence-bound review material rather than autonomous conclusions.

See Model Portfolio for the longer evaluation funnel, training candidates, and acceptance metrics.

Current Profiles

Module	Profile	Purpose
SignalForge	PyOD ECOD	Comment-outlier candidates from transparent text, link, mention, hashtag, punctuation, and duplication features.
GlyphWatch	PP-OCRv6 Small	OCR for explicitly supplied local image evidence.
GlyphWatch.Depth	Depth Anything V2 Small	Relative-depth image derivative for comparison of explicit local image evidence.

Operating Rules

• No package installation or model download happens silently.
• OCR requires an explicit model-download confirmation on its first local run.
• Depth analysis requires confirm_depth_analysis: true and produces a local derived PNG.
• A model result is stored with its profile, revision or runtime context, timestamp, and source evidence reference.
• The original local evidence remains the primary reference.

Local LLMs

MindForge is a separate, opt-in loopback bridge for local providers such as Ollama, llama.cpp, or vLLM. It is intended for source-bound drafts and structured responses. A response should keep evidence-backed facts distinct from its probabilistic model assessment.

Private Model Mirror

Approved model snapshots can be mirrored to the private Hugging Face bucket Lennox9898/sithassembly-model-store. The human-editable allowlist is config/model_mirror_registry.json; tools/Sync-ModelBucket.ps1 runs as a dry run unless called with -Apply.

The mirror must contain model snapshots and permitted evaluation artifacts only. It must not contain case evidence, credentials, databases, vaults, logs, or general source code.