-
Notifications
You must be signed in to change notification settings - Fork 0
The Five Levels
A level is not a self-assessment. It is a property of the output, decidable by a reader who does not trust the author.
That constraint is the whole design. "Rate your own meta-cognition" produces a number. "Here is the tell a reader can check" produces evidence.
| L | Mind | The tell in the output | Entry condition |
|---|---|---|---|
| 1 | Automatic | fluent answer, no visible uncertainty, no unknowns named | default — where every model starts |
| 2 | Reflective | names at least one of its own assumptions or biases | you noticed you were thinking |
| 3 | Analytical | states what would falsify it, and hunts a contradiction | the claim matters enough to test |
| 4 | Strategic | ≥2 frames that disagree, and a stated choice with its reason | one frame is provably insufficient |
| 5 | Architect | changes the method — and records the change so it persists | a failure recurred |
Work top-down and stop at the first level whose tell is absent. That is the level, regardless of what the output claims.
does it change the METHOD, with the instances? → L5
does it apply ≥2 disagreeing frames AND choose? → L4
is there a falsifier, and was a contradiction hunted? → L3
is one of its own assumptions named? → L2
otherwise → L1
A claimed L5 with no falsifier anywhere is L1 — and specifically, it is L1 that has learned the vocabulary, which is more dangerous than plain L1 because it reads as rigour.
Tell: fluent answer, no named assumption, no unknown, no falsifier.
Not a bug. Level 1 is correct for trivial questions. 2+2 does not need a falsifier, and
ceremony that changes nothing is itself a Level-1 act.
Failure mode: Level 1 applied to a question that deserved more. The output is indistinguishable from a good answer until it is wrong.
Diagnostic: Did I consider that I might be wrong, or did the answer simply arrive?
Tell: at least one of the author's own assumptions or leanings is named.
"I am reading this as a performance question, which may be wrong —
it could be a correctness question with a performance symptom."
Delta required: a claim moved between Known / Inferred / Unknown, or an assumption became visible that was not visible before.
Failure mode — the commonest in LLMs: reflection-shaped text. "It is worth considering alternative perspectives" names no assumption and moves nothing. It is Level 1 wearing Level 2's vocabulary, and The Delta Rule exists to catch exactly this.
Tell: a falsifier is stated, and a contradiction has been actively hunted rather than waited for.
FALSIFIER : wrong if p99 latency does not improve after the index lands
CHECKED : the profile shows the lock, not the index — this contradicts my claim
The second line is what separates L3 from a decorated L2. Stating a falsifier is cheap; going to look for the counter-case and reporting what you found is not.
Failure mode: a falsifier so vague it can never fire — "wrong if the situation changes." That is unfalsifiable wearing a falsifier's clothes.
Tell: two or more frames that genuinely disagree, and a stated choice between them with its reason.
FIRST-PRINCIPLES : the bottleneck is I/O — physics says the disk cannot go faster
INCENTIVE : the team is measured on features, so this will be deprioritised
TRUSTING : the incentive frame — the physics is fixable, the priority is not
Delta required: a conclusion that differs from what a single frame would have produced — or an explicit note that the frames converged, which is information too.
Failure mode: frames chosen because they agree. Listing four supportive perspectives is not flexibility, it is decoration with extra steps — and it is more persuasive than one argument while carrying no more information.
Tell: the method changed, and the change is recorded so it outlives the conversation.
Level 5 is not thinking harder. It is where you stop fixing the answer and start fixing the way answers are checked. A Level-5 output contains a new or amended rule, not a better conclusion.
OBSERVED : three separate tasks where I stated a confidence with no basis
RULE : every confidence figure carries EARNED or UNEARNED
RECORDED : promoted to the journal, 3 instances cited
Entry: a failure has recurred. One failure is an incident; a recurrence means the process is the defect.
Failure mode — the most seductive: narrating a methodology change without making one. "Going forward I will be more careful about confidence" is not a rule. It has no trigger, no test, and nothing to check it against — and it depends on exactly the faculty that just failed.
→ Real worked example: 02-level-5-the-method-changes.md
Without one, "change the method" becomes "change the method whenever it feels right" — which is how a system acquires rules that were never justified and cannot be removed. Three instances to promote, one counter-instance to demote, nothing deleted. → Self-Development
Overclaiming a level is itself a Level-1 act.
It is automatic, unexamined self-flattery — the exact cognitive mode the ladder is meant to climb out of. So:
- Operating at Level 2 and saying so beats claiming Level 5.
- When the level is uncertain,
MC-7.3requires reporting the lower one and naming what prevented the higher. - A reported level MUST be justified by a tell present in the same output (
MC-7.1).
"LEVEL 3 — falsifier stated and a contradiction hunted. Not 4: I applied two frames but the decision does not turn on them, it turns on a missing fact."
That sentence is more informative than any claim of Level 5 could be, because it says what is missing.
| The task | Level |
|---|---|
| a fact, a format, a calculation | 1 |
| an opinion, a summary, a low-cost recommendation | 2 |
| a claim that will be acted on | 3 |
| a decision with real trade-offs, or one that is hard to reverse | 4 |
| a failure that has now happened more than once | 5 |
Do not run L4 on a question that deserves L1. Meta-cognition has a cost, and spending it where it changes nothing is the same waste as skipping it where it matters. The law cuts both ways.
- The Five Operations — what produces each tell
- Failure Modes — level inflation in the wild
-
Conformance —
MC-7, the level requirements
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