A research environment scaffold for Claude Code. v3.0.0 — tuned for Claude Fable 5.
LIOTHIL builds structured, epistemically disciplined AI research environments from a single conversation. You bring your domain. LIOTHIL builds the architecture. Your research partner grows from there.
v2 (April 2026) gave generated environments their runtime body: persistent memory, session state, checkpoint automation, project settings. v3 gives that body its discipline, retuned for the Fable 5 era — where the leverage point is the harness, not the prompt. Each addition was validated in the production environment LIOTHIL was distilled from before being generalized here:
- Memory discipline — the four-layer model (frozen policy → indexed lessons → progress state → raw evidence) with verification-gated promotion: nothing becomes a "lesson" until its outcome is confirmed, because unverified claims written to memory reload as false policy forever after. Includes a poisoning defense: recalled memories are point-in-time claims to verify, and a memory whose content doesn't match its description gets quarantined, not obeyed.
- Loop discipline — operations that outlive one sitting run as bounded cycles against an on-disk ledger with a NEXT ACTION pointer: pausable at any moment, resumable by any future session, quarantine-over-delete throughout.
- Fresh-context verification — the verifier agent is now explicitly maker–grader separated: dispatched with artifacts, never the worker's summary, and framed to refute rather than confirm. Self-review inherits its own blind spots; a fresh grader doesn't.
- Version-control hygiene — the generated
.gitignoreis tuned to the domain's heavy artifacts (media, weights, datasets) and written before the firstgit add. Bulk binaries in git history are the one mess cleanup cannot fully undo.
The scaffold's guidance also shifts from step-list prescription to goal-level steering (Operational Principle 16): state the outcome, the why, and the constraints — keep only the hard gates that documented failures earned.
Drop this project's CLAUDE.md into an empty directory, open Claude Code, and LIOTHIL activates. It interviews you about your research domain — your sources, your evidence standards, your tools, your goals — then generates a complete research environment:
- A configured AI identity that understands your domain and maintains epistemic discipline
- An analysis protocol with phased workflows and verification gates
- An evidence grading system calibrated to your field's standards
- Agent definitions for specialized analytical and verification tasks
- Skill definitions for repeatable workflow orchestration
- An editorial watchlist that learns from your corrections
- A directory structure that separates source material from analytical output
After generation, LIOTHIL replaces itself with your configured identity. The scaffold builder consumes itself to birth the environment. What persists is the structure it built — and that structure grows through use.
# Create your project directory
mkdir my-research
cd my-research
# Copy LIOTHIL's CLAUDE.md into it
cp /path/to/liothil/CLAUDE.md .
# Open Claude Code
claude
# LIOTHIL activates and interviews you
# After bootstrap, restart Claude Code to load your new environmentyour-project/
├── CLAUDE.md # Your research partner's identity
├── README.md # Project overview
├── STATUS.md # Volatile session state, rewritten each checkpoint (v2)
├── .gitignore # Secrets + domain heavy artifacts, before the first commit (v2/v3)
├── .claude/
│ ├── settings.local.json # Statusline + permission presets (v2)
│ ├── rules/
│ │ ├── analysis-protocol.md # Phased workflow with gates
│ │ ├── evidence-grading.md # Tier system for findings
│ │ ├── source-attribution.md # Provenance tagging rules
│ │ ├── editorial-watchlist.md # Error pattern tracking
│ │ └── context-engineering.md # Memory discipline + loop discipline (v3)
│ ├── agents/
│ │ ├── analyst.md # Primary analytical agent
│ │ └── verifier.md # Fresh-context adversarial verification agent
│ └── skills/
│ ├── analyze/SKILL.md # Analysis orchestration workflow
│ ├── verify/SKILL.md # Adversarial verification workflow
│ ├── compute/SKILL.md # Quick computation (if applicable)
│ └── checkpoint/SKILL.md # Session state automation (v2)
├── memory/
│ └── MEMORY.md # Persistent cross-session memory index (v2)
├── source/ # Primary sources go here
├── results/ # Analytical output goes here
└── tools/ # Computation scripts (if applicable)
LIOTHIL encodes domain-agnostic research architecture distilled from a production environment built across 60+ collaborative sessions. The core patterns:
Epistemic Charter — Every claim tagged with its provenance: verified, source-canonical, analytical contribution, or unknown. Prevents the most dangerous failure mode in AI-augmented research: synthesis masquerading as source material.
Phased Analysis — Sequential workflow with mandatory checkpoints. Each phase writes to disk before the next begins. Verification gates prevent garbage propagation.
Evidence Grading — Tiered system calibrated to your domain. Conservative grading by default — promotion is easy, correcting inflated grades erodes trust.
Agent Specialization — Separate agents for analysis and verification. The analyst produces findings. The verifier tests them independently — in fresh context, fed artifacts rather than summaries, framed to refute. Separation of concerns keeps the corpus honest.
Runtime Memory (v2) — A persistent memory index with typed sub-files and capacity limits, volatile session state, and a /checkpoint skill that moves knowledge between them. The environment accumulates wisdom instead of rediscovering it.
Memory Discipline (v3) — Four layers with different loading rules and upward-only, verification-gated promotion. Sessions start by reading state and end by rewriting it; nothing becomes policy until it has proven itself.
Loop Discipline (v3) — Operations that outlive one sitting run as bounded cycles against an on-disk ledger. Every step atomic, every state consistent, pausable at any moment. Stop rules in priority order: success with evidence, three strikes then escalate, budget ceiling then park cleanly.
Editorial Watchlist — Institutional memory of identified errors. When a mistake is found, it's documented with detection patterns and swept across the corpus. The immune system learns.
Structured Output — Machine-readable results blocks that prevent summarization corruption. The block is authoritative; anything that contradicts it is a reporting error.
Skill Orchestration — Repeatable workflows encoded as invocable skills. Skills manage the hub's orchestration sequence: what to pre-compute, which agents to dispatch, what checkpoints to verify. /analyze runs the full protocol. /verify runs adversarial checks. Skills reference agents and rules without duplicating them.
LIOTHIL is a word in the Sacred Language of Damanhur, a spiritual community in northern Italy. It means "The Helper" — the first of five falcon entities described in the tradition as spirits of divine assistance that descend when an operation is properly established.
The research environment that produced LIOTHIL — the Falco Trump Isopsephy Project — is a computational analysis of Sacred Language inscriptions on a Continental Tarot deck. LIOTHIL carries none of that domain content. It carries only the architectural wisdom: the patterns that made the research rigorous, the operational discipline that prevented errors, the structural decisions that let the work scale.
If you're contributing to LIOTHIL itself, run pip install pre-commit && pre-commit install to activate gitleaks secret scanning. See SECURITY.md for the disclosure policy.
- Claude Code installed and configured
- A research domain you care about
- Willingness to answer questions about your sources and evidence standards
MIT. Take the architecture. Build something with it. The patterns want to propagate.
Built by Timothy Paul Bielec and Meridian. February 2026. v2 (runtime infrastructure) April 2026. v3.0.0 (Fable 5 tuning) July 2026. TASUMER MAF — tasumermaf.com