Runtime — SOFIA's first implementation uses Claude Code and CLAUDE.md files. The concepts described here (persona, isolation, artifacts) are provider-agnostic — only the runtime layer is specific to a provider. Some sections describe Claude Code specifics. Latter versions implement other providers.
Scope Two axes Guides Quick start Cost

Audit & Analysis

Observing and verifying a SOFIA instance — conformity, friction, trajectory.

Scope

The tools documented here are specific to the filesystem binding (binding/filesystem/). They parse Markdown files, frontmatter, and session logs stored on disk. A different binding (API, database) would require different audit tools — the protocol defines what to observe, these guides document how to observe it with the filesystem implementation.

See protocol/h2a.md for the protocol-level definition of friction, contribution, and conformity.

Two axes

Axis Question Tool Output
Conformity Does the instance respect the protocol structure? probe.py Pass/warn/fail checks
Analysis How is friction evolving? Is the orchestrator healthy? analysis.py (scan + mirror + lens) JSON data, dashboard

Conformity is binary — a check passes or fails. Analysis is observational — no judgment, just indicators for the orchestrator's attention.

Guides

Guide Content
Audit guide Conformity checks — when to audit, how to run probe, reading results, common issues
Analysis guide Data pipeline — analysis.py, workflow, sanitization, contribution
Dashboard guide The 5 views (Map, Mirror, Lens, Probe, Legend) — what each shows, how to read it

Quick start

# Conformity check
python binding/filesystem/analysis.py /path/to/instance --only probe

# Full analysis + dashboard
python binding/filesystem/analysis.py /path/to/instance --serve

# Multi-instance, output to h2a-data
python binding/filesystem/analysis.py /path/to/inst1 /path/to/inst2 \
    --sanitize -o /path/to/h2a-data/data

Cost

All tools run locally — no API calls, no token cost. They parse Markdown and produce JSON. The dashboard runs a local HTTP server or a static build.

Sofia's audit mode (mode 4) is different: it requires loading the full method context at boot (~100K tokens). That is a provider cost, not an audit tool cost.