When the honest answer is “I refuse, and here is why”

Every commercial solver returns a number. This is the design for making a principled refusal an output with a reason and a cost, rather than an absence.

DRAFT — the ladder is real, the product is not built

Every commercial solver returns a number. Always. A tool whose distinctive output is a principled refusal, with its reason and its cost, does not exist commercially. This page is the design for making refusal an output rather than an absence. Levels 0–1 exist in this tree today; 2–4 do not.

The distinction everything rests on already held

From the existing refusal-frontier probe, and it is the sentence the whole idea turns on:

“the METHOD refuses here” is NOT “no solution exists here”

every refusal is a statement about the admissibility of the certificate, never about the mathematics. Collapsing those two is how a tool starts lying politely.

The ladder 0–1 built · 2–4 designed

level	what the tool says	status
0	silence, or a generic failure	useless to anyone. Indistinguishable from a crash.
1	named mode — REFUSED [derivative-sign]: interval straddles 0 (lo −1.0137)	BUILT. The certifier does this today, and each refusal carries a decidable named mode.
2	certified obstruction — not “I failed to find r” but “there is provably NO admissible r here”	DESIGNED. This is the increment: the refusal becomes a certificate in its own right, proved in interval arithmetic, rather than a report about a search.
3	distance to feasibility — “refuses; the smallest change that would certify is: narrow the window 4.1×, or raise N to 22”	DESIGNED. Turns a refusal into a specification. This is the level a buyer can act on.
4	the certified frontier — the locus where the method crosses from certify to refuse	PROTOTYPED in a scratchpad, never productised. A map of where a method works is an asset nobody else holds for that method.

The clever part: certify the NEGATIVE underexplored

Today the machinery certifies existence and uniqueness in a ball. The same machinery, pointed the other way, certifies statements almost nobody produces rigorously:

certified non-existence	“there is provably no solution in this region”. An exclusion certificate. For a safety envelope this is the statement that actually matters — not “we found nothing bad” but “nothing bad can be here”.
certified multiplicity	“there are provably at least two distinct solutions”. The 2D herding pitchfork in this project is exactly this phenomenon and is currently traced numerically, uncertified. Certifying a bifurcation is a real, hard, saleable statement.
certified non-uniqueness of a decision	the commercial translation: “this optimisation has multiple optima and your solver returned one of them without telling you”.

Neighbour to check before any claim of novelty here: computer-assisted proofs of saddle-node bifurcations already exist, and a fold proof is intrinsically parametric. The literature gate that found that is binding — the contribution would be the automation and the decision framing, not the mechanism.

Why a refusal is worth money the commercial case

it is the only honest safety signal	in safety-critical work the operative question is not “what is the answer” but “when should I not trust it”. Nothing on the market answers that.
it is what qualification requires	standards demand a stated operating envelope with justification. Today that is testing plus engineering judgement. A certified boundary is strictly stronger evidence, and qualification cost is where the money is.
it is un-fakeable	a competitor can claim accuracy. A competitor cannot claim principled refusal without building it, and a refusal that never fires is visibly not a refusal.
it converts narrowness into precision	a narrow certified scope reads as a limitation until it is stated as a boundary. Then it reads as the only honest offer in the room.

What must be true before any of this ships

a refusal that cannot fire is not a refusal	every refusal mode needs a red control proving it fires, and a green control proving it does not fire everywhere. Both, or the mode is decoration.
level 2 is a claim, so it needs a proof	“provably no admissible r” is a theorem about an interval, not a summary of a failed loop. It is only worth saying if it is checked.
the frontier is a measurement, not a picture	a boundary drawn by sampling is an illustration. A boundary with certified sides is an asset. The difference must be stated on the artifact, every time.