To depart from a rule, first give the rule an address. To return, say what must come back. The original state? A representative? A useful answer? An acceptable shape? These requests have different mathematics. The machine does not get to exchange one for another while you are looking at the drawing.
This is a formal treatment of irreverence as a specified departure from a representation or admissibility regime. That phrase is a proposed organizing definition, not an established mathematical discipline. Its working parts are classical: quotients, inverse problems, selections, dynamical systems, probability kernels and resource costs. Their proofs are here. Their limits are here too.
0. Fill in the blanks before breaking anything
A regime specifies a source space X, a representation E:X → Z, an admissibility predicate a, and allowed operations. A departure d changes a state, its representation or a rule; we report which. A recovery contract supplies observations o, controls U, an endpoint map T, a nonnegative cost c, budget B and target relation R. A policy p may depend only on the declared observation and memory.
u = p(z) is permitted,
c(x,u) <= B,
(x,T(x,u)) belongs to R.
The relation R can mean equality to a frozen original, membership in an ideal family, agreement on a task, or a stated tolerance. Each is legitimate. They are not interchangeable. The repair chapter gives precise quantifiers and shows why separate successful repairs for hidden states do not necessarily give one implementable policy.
We call the specified departure irreverent relative to this regime when it violates the named admissibility condition or changes the rule that determines admissibility. Its recovery profile then records what can be repaired under which information, budget and regularity conditions. Unrecoverability is one possible outcome, not the definition of all irreverence. A reversible deviation can still depart from a rule. A noisy accident can be expensive without supplying intention, wit or agency.
That last distinction keeps the mathematics honest about the song. A formal interpreter can enact a specified refusal. These theorems do not establish that it experiences joy, chooses freely, or escapes computation.
follow a contradiction
- “It is bounded, therefore we can invert it.”
- Look at fibres first. Some bounded maps retain all information; some identify different inputs. Even an exact inverse can be unstable.
- “Every layer is continuous on a compact set, therefore every word behaves.”
- Try repeated LayerNorm near zero. Finite-depth compactness and equicontinuity of all words are different claims.
- “Changing the target restored the original.”
- Ask Underprint's frozen witness. New grammar, same reading and same full state are separate checks.
- “Two primitives score alike, so their compositions score alike.”
- Use the same hidden coin twice. Equal marginal scores need not give equal joint laws.
- “Every ellipse can be repaired, so a coherent repair exists.”
- Keep the orientation and go round the singularity. Individual endpoints and continuous faithful choices do not have the same existence conditions.
a small ledger of honesty
Proof means an argument under displayed assumptions. Numerical illustration means local finite computation of named formulas, not a proof of a supremum. Proposed definition fixes this work's vocabulary. Research hypothesis awaits a discriminating test. Classical results retain their names and attribution; this synthesis claims no research priority.
The computational diagrams use explicit small mathematical systems. No model runs in these pages. The evidence register states what the separate operations experiment has and has not established. All prose, equations and static diagrams remain available without JavaScript. The interactive diagrams add inspection, not missing premises.