Home

Meta-Cognition Agent OS

Teach any agent to think about its own thinking — and prove it did.

MC-0 Delta Rule · MC-1…5 Operations · MC-6 Gate · MC-7 Levels · MC-8 Self-development · MC-9 Scope

---

The problem this targets

Ask any model to reflect on its reasoning and it will reflect, fluently:

"Let me examine my assumptions and consider alternative perspectives."

That sentence names no assumption and holds no perspective. It costs nothing to produce and is indistinguishable from having done the work.

This is not a prompt-quality problem. It is structural. Every meta-cognition framework — every "think step by step, then critique yourself" instruction — can be satisfied by writing as if it had been followed. Which means none of them was ever enforcing anything. The model is not lying; it is producing the text most consistent with having reflected, and that text is cheap.

So this OS is built on one law:

A meta-cognitive step that changed nothing did not happen.

Not "reflect." Name what changed, or report the step skipped.

It is the only requirement here that fluency cannot fake, and every other rule is written to be an observable property of an output for the same reason. → The Delta Rule

---

What it is

A control layer. It governs how an agent thinks, not what it concludes. A domain skill answers the question; this decides whether the way the answer was reached can be trusted.

flowchart TB
    MC["META-COGNITION AGENT OS<br/><i>how the thinking is checked</i>"]
    FP["FIRST PRINCIPLE CODEX OS<br/><i>the epistemic base layer</i>"]
    DOM["your domain skill<br/><i>finance · logistics · code · medicine</i>"]
    RT["any instruction-following model"]
    MC --> FP --> DOM --> RT

Loading

One file. No tools, no runtime, no dependencies, no API calls. If a host reads a system prompt, it runs this. → Getting Started · Host Adapters

---

The shape of it

5 operations	STATE · MONITORING · FALSIFICATION · FLEXIBILITY · CALIBRATION	The Five Operations
5 levels	Automatic → Reflective → Analytical → Strategic → Architect	The Five Levels
1 gate	non-skippable, and permitted to refuse	The Five Operations
1 loop	journal → 3 instances promote · 1 counter-instance demotes	Self-Development

The levels are read from the output, never declared. Each has a tell a reader can verify without trusting the author — and overclaiming a level is itself a Level-1 act, because it is automatic, unexamined self-flattery.

---

Two ways in

You want to use it — 2 minutes. Copy SKILL.md into your agent's instructions. Then read Getting Started for what actually changes, and what does not.

You want to know whether it works — 10 minutes. Read Failure Modes first. It is the catalogue of reflection-shaped text: what a fluent, confident, entirely hollow meta-cognitive output looks like, and which requirement each part violates. If those patterns look unfamiliar, this OS will not seem necessary. If they look like every AI answer you have ever read, start with The Delta Rule.

---

The uncomfortable part

A conforming output is usually worse-looking than a non-conforming one. It is shorter on conclusions, longer on unknowns, more willing to refuse, and it reports a lower level.

That is not a side effect to be tuned away. Reflection-shaped text does not look like a failure — it looks like diligence, and it is longer and more confident than the correct answer would have been. If adopting this makes your agent sound less impressive, it is working.

→ FAQ

---

What it is made of

Not a framework assembled from reading. Every operation is an organ from a system that already ran, and each earned its place by catching a real defect. → Synthesis

Operation	From
Monitoring · Shadow Gate	First Principle Codex OS
Falsification	Genesis Protocol
Flexibility	Genesis Mind
State, affect as data	Genesis Consciousness OS
Authority limits	Genesis Governance OS
Calibration — EARNED / UNEARNED	Genesis Reality Grading
Self-development	SkynetClaw — the running proof

---

Verify rather than trust

Conformance lists MC-0 … MC-9 as RFC-2119 requirements, each phrased as a property of an output rather than of an intention — so a third party can audit a build without trusting it.

A build cannot certify itself by asserting the IDs. That is the whole point of writing them as observables.

python tools/validate_skill.py    # installability + integrity
python -m pytest -q               # the contract tests

---

Most minds are the piece being moved. Meta-cognition is seeing the board. This OS is the part that checks you are actually looking at it.