The Full Mathematical Flow of ULR

The complete flow#

This page covers only the canonical mathematical chain of Canon 24. The post-Canon response span, minimum-effect definition, and absorption/no-go sequence are separated into the non-canonical layer on the Post-Canon frontier.

ULR mathematics does not begin by asserting that a common latent object exists. It begins with a limited alignment phenomenon and, before interpreting that phenomenon, specifies separate equivalence relations for coordinates, functions, relations, experiments, and time.

Each arrow carries a separate proof obligation. A positive result at an earlier stage does not automatically establish existence at a later stage.

1. Phenomenon: different representations share some relations#

Write the item representation of model $m$ as a matrix $H_m$. What ULR first observed was not agreement between the raw coordinates of $H_m$, but a limited alignment of inter-item relation geometry. Define the centred Gram matrix by

Provided that $\lVert K_a\rVert_F>0$ and $\lVert K_b\rVert_F>0$, linear CKA can be read as the cosine between two centred Gram matrices.

Experiments found values above a shuffled null for some model–dataset pairs. From the outset, however, this expression is an observation protocol. High CKA does not imply an observer-independent common space, the same mechanism, the same meaning, or a single ULR object.

The first conclusion licensed by the phenomenon#

Anything stronger requires an identity contract for the observable.

2. Observable typing: what was measured, and what was quotiented out?#

Every representation observable declares at least the following triple.

• $O$: the raw observable, such as a Gram matrix, relation cosine, or activation response.
• $\mathcal G$: the architecture-realisable gauge action under which representations are identified.
• $N$: the normalisation or quotient map applied to the observable.

Comparisons across models or times may require extending this contract to the following quadruple.

Here the resulting quotient representation is $Q=N(O)$, and $\mathcal S$ is a comparison map such as Procrustes. $\mathcal S$ is neither the same as a function-preserving gauge nor the same as $O$ itself.

Every observable has its own gauge#

CKA, Procrustes residual, and relation cosine do not automatically share the same invariant contract. In the actual typed audit, relation cosine collapsed from 0.4712 to −0.0091 under the independent product gauge, while remaining at 0.4910 under the shared diagonal subgroup. A passing result for one observable therefore cannot be transferred to another observable.

3. Architecture gauge: separating function from coordinates#

Let $\theta$ denote parameters and $f_\theta$ the function implemented by the architecture. An admissible gauge $g\in\mathcal G$ may satisfy

while changing hidden coordinates or the raw Gram matrix. Head permutation, positive scaling, QK and VO co-transformations, expert permutation, normalisation, and residual coupling differ by architecture.

An identity candidate must therefore be relative at least to the orbit space, not to the raw parameter space $\Theta$.

Even here, the two directions must be distinguished.

whereas

The converse does not follow without additional function-fibre analysis.

4. From local gauge to global network symmetry#

A local change permitted by an equation at a node becomes a global gauge only if it lifts to an actual parameter reparameterisation, is compatible across branches and residual paths, and preserves the entire network function. This layer is organised by the filtration

The G1–G6 mathematics of ULR classified these distinctions through local characterisations, lift obstruction, shared lifts, occurrence gluing, attention gluing, block sectors, and jet specifications.

Scope boundary#

Completing the work plan for a minimal bias-free Transformer block does not classify the maximal gauge of every Transformer. Global exact completeness holds under declared assumptions. The only unconditional result at that scope is generic infinitesimal completeness.

5. Function fibre: every parameterisation that implements the same function#

For a function $f$, define its fibre by

The central question is whether $\mathcal F_f/\mathcal G$ contains a residual outside the gauge. Results on the ATT quotient, relative MLP fibre, channel rigidity, atlas admissibility, effective residual, and joint witnesses address the finiteness, rigidity, and generic structure of this fibre for particular models and open sets.

The following inference is prohibited, however.

A non-gauge residual means only that function equivalence did not remove it. Semantic, causal, or behavioural significance requires a separate bridge.

6. Lift, gluing, quotient, and assembly#

Even when the same local description is available, a global lift into parameter space may not exist. A local certificate may hold in several charts while the transitions fail to agree on their overlaps. Assembly must therefore separate all of the following.

1. Local admissibility;
2. existence of a parameter lift;
3. overlap compatibility;
4. definition of the quotient source;
5. injectivity of the quotient map;
6. unramifiedness of its differential.

This separation is the core of the UAR correction.

Two different maps in UAR#

Channel incidence asks whether off-diagonal pairs can be excluded in

The assembly side instead asks about the tangent and ramification of a quotient map on $Q=\Theta^{\mathrm{adm}}/G_\Sigma$. Consequently,

cannot be combined into a single monomorphism theorem. The current Canon preserves typed channel exclusion and assembly unramifiedness separately, and retracts the composite monomorphism.

Separating existence from universal quantification#

Suppose the local certificate is sound, so that $\mathrm{Good}(\theta,x)\Rightarrow\mathrm{NoCollision}(\theta,x)$ at each audited point. Even then, the existence of one good point in a fibre does not mean that the entire fibre is collision-free.

This quantifier error was the central audit finding that overturned the impression of completion in Canon 19.

7. Role: distinguished by response, not internal location#

Because learning and inference can use the same local operator, the presence or absence of a weight update is not a sufficient definition. The current candidate typing is

The minimality of this tuple has not yet been proved. What can be identified at a declared port can, however, be stated exactly.

Reset response quotient#

For a linear system, let $R$ be the persistence law, $H$ the read port, and $B$ the preparation basis. The reset response is

If $H$ has full column rank and $B$ has full row rank—equivalently, $H$ has a left inverse and $B$ has a right inverse—then $R$ can be recovered from $HRB$. At a partial port, however, even matrices $R$ of different ranks may produce the same $HRB$. The identifiable object is not the internal $R$ itself but an interface-relative equivalence class.

Passive equivalence and a richer port#

Two systems may agree on passive trajectories and outputs while differing under write; reset; read or a clone policy. This shows that a richer interface can strictly refine a passive quotient; it does not show that the resulting quotient is a ULR object.

8. Observer-relative discrimination#

Let $C\in\{L,I\}$ be the role and $\mathcal F^O$ the observer filtration. A deterministic exact classifier exists if and only if

For probabilistic transcript laws $P_L,P_I$, the equal-prior Bayes risk is

Thus, if $P_L=P_I$, no classifier can beat chance; if the laws are mutually singular, zero error is possible.

For a set $\Pi$ of admissible policies, the optimal discrimination power of an adaptive observer is

A universal classifier that works for every event exists only when the role label factors through the full response quotient.

Internal and external information over time#

If the internal transcript is a Markov garbling of the external transcript, then

Arbitrarily complex computation after observation cannot recover discarded provenance. Under a fixed infinite policy, finite-prefix error converges to 0 if and only if the full path laws are singular.

9. Comparing the phenomenon and role with the strongest baseline#

ULR does not recognise a new object merely because a candidate quantity $Q$ outperforms a passive baseline. The baseline hierarchy is

An incremental residual must at least satisfy

on a preregistered held-out target. In the engineered small-neural experiment, $Q$ substantially outperformed B2 but did not cross the non-absorption threshold relative to B3/B4.

10. Ordering hypotheses by elimination, not existence#

Z0–Z7 form a registry of competing explanations. The present decision places Z0, which requires the least ontology, first; another hypothesis defeats Z0 only by making an independent prediction.

At present Z1 and Z5 are weakened, Z2 and Z6 are baselines, Z3 is exploratory, Z4 is a scaffold, and Z7 is conditional language.

11. Carrier reduction#

Any new ontology requires a carrier capable of bearing it. The same four-gate reduction rule was applied to the seven candidates in the current registry.

1. Are the object and identity typed?
2. Do they descend through the gauge, observer, and port?
3. Are they irreducible to the strongest existing theory?
4. Do they make an independent prediction of held-out behaviour?

The result is 0/7.

12. Formation: from curve onset to structural transition#

If formation is defined only as the time $t_0$ at which a scalar $s_t$ becomes large, then choices of coordinates, metric, and threshold manufacture the event. Let $T$ be the declared time index. The conditional ULR definition first requires a time-indexed object package.

where $\sim_t$ is identity, $\{\tau_{t\to s}\}_{s\in T}$ is the declared family of cross-time transport maps, and $\operatorname{Type}_t$ is structural type. A formation event must be a change of structural stratum under transport, not merely a change in the quotient value.

Because there is currently no carrier survivor, the universal neural instance is NO_ADMISSIBLE_TARGET. Co-onset measurement remains only as an auxiliary probe after a candidate object has been established.

13. How the final conclusion is reached#

The end of the mathematical flow does not deny the phenomena.

1. Alignment phenomena exist.
2. They cannot be explained away entirely as raw-coordinate artefacts.
3. There is exact mathematics for gauge and function fibre.
4. There is also observer-relative response separation.
5. Each result is nevertheless absorbed by an existing typed theory.
6. No carrier survives, and no universal formation target exists.
7. The UAR composite claim was retracted because it was not well typed.
8. There is no independent neural-specific prediction beyond the strongest baseline.

Therefore,

is the most compressed statement of the Canon 24 conclusion.

Logical guardrails#

Individual atoms and their statuses appear in the claim ledger; experimental evidence for each milestone continues in M1–M8.