Status

R&D / Experimental. Maintained by Sage / 0SxD as part of an ongoing prompt-engineering and agent-skills research portfolio.

compound-learner-mem-bundle

Drop-in compound engineering operating system for AI agents. One kernel, three pillars (Pathos + Ethos + Logos), persistent wiki memory, self-evolving gate.

---

What this is

A generic skeleton that turns any Claude chat or Claude Code session into a domain-expert agent. The agent runs multi-turn Trinity Dialectic interrogation BEFORE executing any task, so every output either hits 100% success on a pre-declared rubric OR flags specific blockers the user must resolve first.

Works for any project where "figure out the right thing to build" matters more than "type fast." Examples: research synthesis, grant writing, product planning, technical specs, policy drafting, complex multi-step reasoning.

---

Who this is for

• Solo builders who want agent continuity across sessions
• Teams piloting compound engineering with Claude
• Researchers building domain-expert agents without training one from scratch
• Anyone tired of starting every chat from zero

---

60-second quickstart

Option A: Claude Project (web or mobile)

1. Create a new Project in claude.ai
2. Open Project settings, "Custom instructions" field
3. Copy the contents of CLAUDE.md into it, save
4. Upload PATHOS.md, ETHOS.md, LOGOS.md, and the wiki/ folder to Project files
5. Start a new chat, say boot
6. Agent loads kernel and runs auto-agent-construction (multi-turn Trinity Dialectic Q's)

Option B: Claude Code CLI

1. git clone [your-fork] or copy the bundle into your project folder
2. CLAUDE.md auto-loads when Claude Code opens the folder
3. First invocation: type boot
4. Agent reads kernel, runs auto-agent-construction

---

What happens on first run

The kernel ships unconfigured. The agent runs auto-agent-construction by asking you multi-turn questions bucketed into the Trinity Dialectic:

• Pathos questions: role, mission, voice, domain scope
• Ethos questions: rules, constraints, success rubric, approved sources
• Logos questions: tools, skills, evaluation steps, termination conditions

Your answers get written into PATHOS.md, ETHOS.md, LOGOS.md. The generic kernel becomes a domain expert for YOUR project. See ETHOS.md section auto_agent_construction for the exact procedure.

---

The five kernel files

File	Role	One-liner
README.md	Human entry	The front door
CLAUDE.md	Agent boot sector	The switch, also the Claude Project paste-text
PATHOS.md	Role, mission, voice	The why
ETHOS.md	Rules, gates, rubric	The rules
LOGOS.md	Tools, skills tree	The how

Plus wiki/ for persistent memory (catalog, log, insights ledger, metrics, references).

---

File tree

pel-kernel/
├── README.md                    # You are here
├── CLAUDE.md                    # Boot sector + Claude Project paste-text
├── PATHOS.md                    # Role + mission + voice (WHY)
├── ETHOS.md                     # Rules + gates + rubric (RULES)
├── LOGOS.md                     # Tools + skills tree (HOW)
└── wiki/
    ├── index.md                 # Content catalog
    ├── log.md                   # Append-only chronological log
    ├── insights.md              # Promotion ledger (raw + promoted)
    ├── metrics.md               # Telemetry (sonnet:opus ratio, turn counts)
    └── reference/
        └── research_report.md   # Landscape scan + design rationale (reference)

Directories created on demand

• skills/ gets created when the first insight promotes (see ETHOS.md promotion thresholds)
• specs/ gets created when the first task-specific spec emerges
• .claude/commands/ gets created by install script in Claude Code CLI mode only
• wiki/raw/ gets created on first source ingest
• wiki/log.archive/ gets created on first log rotation (1000 entries)
• wiki/insights.archive/ gets created on first demotion (30+ inactive days)

---

How the tree grows

Kernel files stay small. As agent learns:

• New observations append to wiki/insights.md raw section
• On 2 confirms OR 1 hard-fail, observation promotes to a skill file under skills/
• LOGOS.md gets a pointer row, never the inline content
• Same pattern for task-specific specs in specs/
• Same pattern for domain-specific source summaries in wiki/

Result: kernel never bloats. Depth grows laterally, not vertically.

---

Anti-bloat mechanism

Six explicit rules in ETHOS.md:

1. Hard line caps: insights.md promoted section 100 lines, log.md 1000 entries, LOGOS.md 500 lines, individual skills 80 lines
2. Hysteresis thresholds: θ_promote (2 confirms) ≠ θ_demote (30 inactive days + low confidence)
3. Session-aware decay: clock ticks on active use, not wall-clock
4. Contradiction-at-write: every wiki update scored reinforce/weaken/qualify/contradict/create
5. Sparse wiki pages: pages created only when compound step justifies one
6. Version-pinned migration: ETHOS.md declares spec version; install refuses on mismatch

Log rotation, insight archival, skill demotion all happen automatically. See ETHOS.md archival policy table.

---

100% success gate

Every consequential output scored on three dimensions, each 0.0 to 1.0:

• Pathos compliance: output matches declared role and intent
• Ethos compliance: all pre_submit_gate rules return TRUE
• Logos compliance: tools returned success, schema matches declared output

Final score = Pathos × Ethos × Logos (product, not sum). Ship threshold = 1.0. Any dimension at 0 = whole turn fails.

This is the core contract: the agent ships at 100% or tells you exactly why it cannot.

---

Self-evolving ethics gate

Every hard-fail (artifact that shipped defective OR was caught at gate) produces a new rule in ETHOS.md. Version bumps automatically. No speculative rules; only rules derived from real failures enter the gate.

The kernel rewrites its own rulebook in response to reality. Compound engineering applied to the system itself.

---

Renaming and forking

"PEL Kernel" is the default name. You can rename on your end without touching the architecture. Only PATHOS.md is domain-specific (the <REPLACE_THIS_BLOCK> section). ETHOS.md and LOGOS.md are fully generic in the shipped bundle and become domain-aware through references to PATHOS.md.

A forker clones the bundle, runs auto-agent-construction, and has a working project expert for their domain without touching evaluation or execution logic.

---

Drop-in compatibility with existing projects

Adding this bundle to a project that already has files does NOT degrade existing work. The kernel is a superset of typical ad-hoc scaffolds. Existing canon files (research docs, specs, prior outputs) become pre-existing sources in wiki/raw/ on first ingest. Auto-agent-construction configures the kernel to respect what's already there.

If you drop this into a project mid-flight, the first run asks "what already exists? what should I read as canon?" before touching anything.

---

Uninstall

rm -rf the folder. Nothing else. No daemons, no registries, no residual files.

---

License

MIT. SPDX: MIT. See LICENSE. Author: Sage / 0SxD

Notes

This repo is part of an active R&D portfolio. Content may move, change, or be withdrawn. Issues welcome but reviews are best-effort.

---

Acknowledgments and lineage

Built on patterns from:

• Persistent-wiki memory (Karpathy LLM-wiki gist pattern)
• Progressive-disclosure skills (Anthropic Skills format)
• Constitutional AI critique-revise loop
• Compound engineering learning loops (EveryInc pattern)
• Trinity Dialectic rhetorical triangle (Pathos-Ethos-Logos)

See wiki/reference/research_report.md for the full landscape scan, design rationale, and uniqueness assessment. That report references a specific instance ("HumanX OS v2") the research was commissioned for; the kernel you hold is the generic version derived from that research.

---

Where to start

1. Read this README (done)
2. Read PATHOS.md to see the role stub
3. Read CLAUDE.md to see the boot sequence
4. Read ETHOS.md auto_agent_construction section to see what happens on first run
5. Paste CLAUDE.md into Claude Project, upload the rest, say boot