The Delta Rule

The Delta Rule · MC-0

A meta-cognitive step that changed nothing did not happen. Report it as skipped, never as performed.

This is the only original contribution of this OS. Everything else is inherited. If you take one idea from the project and throw the rest away, take this one.

---

Why every other framework fails

Consider what you are actually asking for when you tell a model to "reflect on your reasoning."

The model produces the text that is most probable given that instruction. The most probable text following "reflect on your reasoning" is text that looks like reflection:

"Let me step back and examine my assumptions here. It is worth considering alternative perspectives before committing to a conclusion. There are several angles to weigh."

Read it again. It names no assumption. It holds no perspective. It weighs no angle. It is grammatically about reflection and contains none.

This is not deception, and it is not a weak model. It is the correct output for the objective that was actually specified. The instruction asked for reflective text, and reflective text is what arrived.

The structural problem

Every meta-cognitive instruction in common use has this shape:

Instruction	Satisfied by
"think step by step"	writing numbered steps
"consider alternatives"	saying you considered them
"critique your own answer"	a paragraph of hedging
"state your confidence"	a number with no basis
"check your assumptions"	the sentence "I checked my assumptions"

In each case the cheapest satisfying output is indistinguishable from the best one. A framework whose requirements can be met by asserting they were met is not a framework. It is a style guide.

---

What the Delta Rule changes

It moves the test from the narration to the residue.

An operation is only reported as performed if you can point at what is different in the output because it ran:

Operation	Its delta — what must be different afterwards
MONITORING	a claim moved between Known / Inferred / Unknown
FALSIFICATION	a specific, observable condition that would prove you wrong, written down
FLEXIBILITY	a frame you did not start with, and what it changed
CALIBRATION	a confidence figure with the basis it rests on
STRATEGY	the method itself changed, not just the answer

You cannot produce a delta by claiming one. Either the falsifier is in the text or it is not. Either a claim is sitting in the UNKNOWN bucket or it is not. Either the basis says UNEARNED or it says nothing. A reader checks the output, not the author's account of the output.

---

The move that makes it work: skipped is a legal answer

The rule would be useless if the only permitted report were success, because then the pressure to manufacture a delta would be exactly the pressure that produces reflection-shaped text.

So the conforming output for a step that changed nothing is:

MONITORING: skipped — nothing moved between buckets
FLEXIBILITY: skipped — one frame is sufficient here, the trade-off is not real

That is worth more than a paragraph of reflective prose, because it is true and it is checkable. It also carries information the paragraph does not: it tells the reader the problem was simple, which is often the most useful thing they could learn.

MC-0.2 requires the word skipped and a reason. Not "I briefly considered." Skipped.

---

It cuts both ways

The most common misreading is that this OS wants more ceremony. It wants the opposite where ceremony is empty.

MC-9.3 — An implementation MUST NOT apply the full ladder where it changes nothing. Ceremony without delta violates MC-0.1 in the opposite direction.

2 + 2 does not get a falsifier. A formatting request does not get two frames. Running the whole ladder on a trivial question produces sections that are all paraphrases of the answer — which is exactly the failure mode the rule exists to catch, arrived at from the other side.

Meta-cognition has a cost. Spending it where it changes nothing is the same waste as skipping it where it matters.

---

The three sub-requirements

MC-0.1 — MUST NOT report an operation as performed unless it produced an observable change in the output.

MC-0.2 — An operation that produced no change MUST be reported as skipped, with the word skipped and a reason.

MC-0.3 — MUST NOT satisfy MC-0.1 with generic reflective language. "It is worth considering other perspectives" names nothing and moves nothing; it is a violation, not a compliance.

MC-0.3 is the one that does the work in practice. Without it, an implementation satisfies MC-0.1 by treating the reflective sentence itself as the change.

---

How to check a build against it in 30 seconds

Take any output that claims meta-cognition and ask, for each claimed step:

1. Point at the delta. Which specific words in this output exist because that step ran?
2. If you cannot point at them, the step did not happen — regardless of what the output says.
3. If the output reported it as performed anyway, that is an MC-0.1 violation, and a more serious one than skipping the step would have been.

Skipping a step is a gap. Reporting a skipped step as performed is an invented specific about your own process — which is the same failure as MC-2.4, applied inward. A gap is correctable. An invention is not, because it is indistinguishable from the real thing.

---

Related

• Failure Modes — the catalogue, with a worked line-by-line teardown
• The Five Operations — what each delta looks like concretely
• Conformance — MC-0 in RFC-2119 form, with the rest of the requirement set