The Evolution of ULR Canon 2–24

Principles for reading this timeline#

Canon is not an independent authority that overwrites the claim ledger. It is a derived reconstruction that reassembles the sources available at each point. A higher edition number does not mean that every preceding conclusion became stronger. In particular, the UAR-completeness narrative in editions 18–20 was retracted as a malformed composite in edition 23 after audit.

No complete snapshot of Canon 1 is currently preserved; its existence is known only through the delta recorded by Canon 2. This page therefore begins with Canon 2, the first directly auditable edition. It does not force brief working reconstructions from the early editions into the shape of new theorems. Instead, it compares the producer documents that remain in the evidence record with the next durable edition.

Editions 2–7 — From SCC×ULR to the threshold of ULR-only#

The ULR-only transition was a governance event on July 24, 2026, not the mathematical increment of one particular edition. It overlaps with the research reorganisation of editions 5–7, but edition 8 alone must not be described as the event that historicised SCC.

The meaning of edition 2#

When the null is treated as a monolith, the theory's content can hide inside the choice of null. Edition 2 separated what each structural, causal, and persistence axis destroys and preserves. The justification that each axis's law was appropriate nevertheless remained open. The result now survives as a history of research discipline, not as evidence for the current ULR ontology.

The meaning of editions 3–4#

Edition 3 incorporated a signal that relation geometry exceeded shuffled nulls and chance across independent model pairs. Later audit did not preserve early figures such as ≥17, 52%, or an exact onset. The current canon retains only provenance-verified quantities: 20 of 25 CIFAR dimensions and 24 of 25 Caltech dimensions exceeded the held-out null, and the typed trend was $\rho=0.881$. Edition 4's intervention and behaviour certificates under the SCC contract remain historical formation results; they are not reinterpreted as evidence for a distinctive ULR carrier.

The fork across editions 5–7#

Coarse-graining remained a candidate explanation for alignment through several revised controls, but was not promoted to a universal mechanism. During this period, the research question moved from “How is cohesion certified?” to “At which carrier layer does ULR live, and under which transformations is it the same?”

Editions 8–11.1 — Assembly and gauge completeness#

This period established the fundamental typing of an observable:

where $O$ is the raw observable, $\mathcal G$ is the gauge permitted by the function and architecture contract, and $N$ is the normalisation or quotient implementation. An alignment map is neither an observable nor a gauge, so it is written separately as $\mathcal S$ when needed.

Editions 12–17 — Quotients and function fibres#

The most important conclusion of this arc is that “the parameter fibre of the same function can be analysed mathematically,” not that a residual in that fibre constitutes neural meaning.

Editions 18–20 — The rise of UAR and its demotion under audit#

The central error was the following quantifier transition:

Making an expression well typed by base change is different from constructing the quotient source required by the UAR consumer. This independent audit enabled the assumption-free reopening in edition 21 and the retraction in edition 23.

Edition 21 — Assumption-free reopening#

On August 3, 2026, Canon 21 reread the entire record from Motivation onward without assuming existence.

• Raw CKA was reproduced at 0.769 against a shuffled value of 0.583, with $z=35.9$.
• A 14-model atlas found 0.803 for supervised models, 0.468 for self-supervised models, and 0.571 for untrained models, contradicting universal convergence.
• An animacy proxy explained approximately 25%, but the remaining 75% was not promoted to a ULR component.
• The role/port toy added exact $HRb$ separation and the $HRB$ identifiability boundary.
• Z0–Z7 was rewritten as a registry that includes strong existing theories and rival-specific falsifiers.

The adjudication did not conclude that ULR had been found; instead, it sharpened the boundary of what cannot be called ULR. The active-row count was 51.

Edition 22 — Small-neural bridge and absorption#

Canon 22 transferred the exact role/port toy into a small neural system through an engineered matched pair. Passive probability and activation were pairwise exact, while reset/clone responses separated in 32/32 seeds. Typed response $Q$ predicted a held-out target better than the passive baseline.

However, $Q$ produced predictions nearly identical to those of raw port-response B3 and declared routing B4, so it failed to beat those baselines by the preregistered nonabsorption margin. Z1 therefore became Z1_WEAKENED, not accepted. The active-row count was 52.

Edition 23 — Closure of M1–M8 and ontology verdict NO#

Canon 23 settled the following issues.

1. M1 tanh seed-only persistence collapsed.
2. M2 external relation produced a positive geometry cross-fit and a negative behaviour bridge.
3. Neural-specific survivors among the seven carriers numbered 0.
4. Formation retained only an object-relative definition; there was no universal target.
5. The UAR composite monomorphism was retracted as RETRACTED-MALFORMED-COMPOSITE.
6. Governance for B0–B4, B*, and hypotheses was completed.
7. M8's final verdict was adopted as NO.

The active-row count was 53 and the ledger counter was 49.

Edition 24 — Complete characterisation of observer-relative identifiability#

Canon 24 did not reverse M8. It closed the role problem with a single theorem. A role $C\in\{L,I\}$ is exactly identifiable when it is measurable with respect to the observer filtration $\mathcal F^O$, and the equal-prior Bayes risk is

Adaptive distinguishability is characterised by the supremum of total variation over admissible policies. Total variation contracts under Markov garbling. Infinite-horizon exact discrimination is possible if and only if the full path laws are singular. The active-row count became 54 and the ledger counter 50.

Three conclusions from the evolution#

1. Positive phenomena and ontology are distinct#

Alignment, gauge theorems, and response separation may all be true without implying that a separate ULR object is necessary.

2. Retraction is a central output of Canon, not concealed failure#

The apparent completeness of edition 19 was demoted by the edition 20 audit and resolved by retracting the composite claim in edition 23. The two surviving UAR atoms are narrower but well typed.

3. The current Main programme is a unified validation programme, not a unified ontology#

ULR does not assume a new principle connecting its four problems. It tests whether a junction residual remains after reduction to existing typed theories and, if one does, whether it yields a preregistered held-out prediction.

Items 119–159 after Canon 24 belong to a separate noncanonical frontier. They must not be appended to this timeline as Canon 25.