-
Notifications
You must be signed in to change notification settings - Fork 0
FAQ
No, and the difference is the entire project. "Think step by step" is satisfied by writing numbered steps. Every instruction of that shape has the same defect: the cheapest satisfying output is indistinguishable from the best one.
This OS is built so that its central requirement cannot be met by asserting it was met. You cannot produce a falsifier by claiming to have one. Either the specific, observable condition is in the text or it is not. → The Delta Rule
Chain-of-thought improves how an answer is reached. This governs whether the answer should be given at all, and what it must disclose about its own basis.
They compose. CoT is reasoning; this is the layer that reads the reasoning and can refuse its
output. An agent can produce beautiful step-by-step reasoning built entirely on an invented
number — CoT has no way to notice, and MC-2.4 plus the Shadow Gate do.
Sometimes, yes — and that is checkable, which is the point.
This is the most important objection, so the honest answer is: this OS does not make an agent incapable of faking. It makes faking detectable by a reader who does not trust it. Every requirement is phrased as an observable property of an output for exactly this reason.
An agent that writes LEVEL 4 in an output with no frames has not fooled the OS — it has produced
an MC-7.2 violation that anyone can verify in ten seconds by working down the ladder. Compare
that to "consider alternative perspectives", where the same fake is undetectable in principle.
The audit is in Conformance; the catalogue of what faking looks like is in Failure Modes.
It will make it sound worse. Answers get shorter on conclusions, longer on unknowns, more willing to refuse, and they report lower levels.
Whether that is worse depends on what you were buying. If you want an agent that sounds authoritative, this is a downgrade and you should not install it. If you want one whose confidence means something, the fluent version was never giving you that — it was giving you the format of it.
The asymmetry is real and worth stating plainly: reflection-shaped text does not look like a failure. It looks like diligence, and it is longer and more confident than the correct answer.
Not if it is installed correctly. MC-9.3 explicitly forbids running the ladder where it changes
nothing — ceremony without a delta is a violation, not compliance.
The default is the Light format: three extra lines carrying confidence, basis, the unknowns that matter, and the falsifier. The Full format is for decisions that are complex, high-stakes, or hard to reverse.
An agent applying the full ladder to "what's 15% of 240" is failing this OS, not following it.
No, and the OS is required to say so. MC-9.1 — an implementation MUST NOT claim subjective
experience.
It detects and reports patterns in its own output. AFFECT reads emotional charge in the
input text. PULL reads a directional bias in the output. Both are observable in text. There
is no claim that anything is felt, and claiming otherwise would be an invented specific — the
MC-2.4 failure, pointed inward. → Failure Modes
Only if your host gives it memory, and it must say which case it is in.
On a stateless API call there is no persistence, the journal lives as long as the conversation,
and MC-8.5 requires that limit be stated rather than implied away. On a host with files, a
vault, notes, or a database, the journal is real and the promotion rule runs across sessions.
An agent claiming to have learned from conversations it cannot remember has invented a specific — the exact failure this OS exists to prevent. → Self-Development · Host Adapters
Three, because two is a coincidence and one is a story. One to demote, because a rule that survives its own refutation is dogma.
The asymmetry is deliberate: a wrong rule does more damage than a missing one, because it is applied with authority to cases it was never justified for, and it is far harder to remove than to add.
Yes, with two caveats. Ceremony degrades first — a small model produces the headings and under-fills them, which is a violation rather than compliance. And the Light format usually beats the Full one, because the full ladder tends to produce reflection-shaped text on a small model.
If a small model can produce only one thing from this OS, it should be the UNKNOWN section.
→ Host Adapters
Yes. Apache-2.0, free for any use including commercial. §4 asks that attribution and the NOTICE
travel with it. §6 grants no trademark rights — the names stay with the author, and no adopter
may imply endorsement or affiliation. → Synthesis
No. SKILL.md is self-contained — one file, no dependencies. The other repos are where the
operations came from, not runtime requirements.
The one worth pairing it with is First Principle Codex OS, which sits directly below as the epistemic base layer. Everything else is optional. → Synthesis
Almost certainly not. MC-7.3 requires that when the level is uncertain, the lower one is
reported along with what prevented the higher.
An output saying "LEVEL 3 — not 4: I applied two frames but the decision turns on a missing fact, not on the frames" is more informative than any claim of Level 5, because it names what is missing. Overclaiming a level is itself a Level-1 act. → The Five Levels
Failure Modes. If those patterns look unfamiliar, this OS will not seem necessary. If they look like every AI answer you have read this month, go to The Delta Rule next.
Open an issue on the repository.
The most useful report is an output — a case where a conforming-looking answer violated a requirement, with the ID. That is a counter-instance, and under the promotion rule one clear counter-instance is enough to demote a rule. Defects in the spec are more valuable than defects in the prose.
Meta-Cognition Agent OS · a control layer for how an agent thinks · MC-0 … MC-9 · Apache-2.0 · Bunyawat Dechanon (ElmatadorZ)
Start
The law
The system
Install
Verify
Project