ULR Current Research Status

The verdict in one sentence#

ULR (Unified Latent Representation) is the current Main research programme. That does not mean that it has discovered a separate neural-specific entity within the present candidate registry and the declared baseline scope. The canonical verdict sealed on August 3, 2026 is:

This NO is not a metaphysical denial that "neural networks have no organisation" or that "no ULR candidate can ever exist." It is a finite research verdict: after reducing the registered candidates against the strongest typed baselines, no meaningful incremental explanatory value survived the combined B baseline, and no candidate passed the preregistered nonabsorption gate*. The programme itself remains active. Its role is to test symmetry, representation comparison, predictive state, causal testing, dynamics, external relation, and family transport together without collapsing them into a single ontology.

Current canonical coordinates#

When authorities conflict, the order is Research Charter → claim ledger → individual proof and experimental evidence → Canon → hypothesis index → progress and execution records. The Canon does not create new claims; it is a readable mirror that reconstructs this evidence.

The four questions posed by the programme#

1. Sharing#

Why do different models, objectives, and modalities arrange some relations among the same data items similarly? The programme distinguishes among explanations based on data structure, supervision, architecture priors, untrained baselines, and metric choice.

2. Identity#

What should count as "the same representation"? Even the same function can have different raw geometry and coordinate interventions, while different functions can exhibit high relational similarity. No single level—weight, function, geometry, or relational statistic—automatically owns identity.

3. Role#

Learning and inference can use the same local computation. The distinction therefore cannot rest on the presence of a weight update or the form of an equation alone. The analysis must instead consider, jointly, the information source, write target, persistence, reset/clone response, and future causal role.

4. Formation#

An increasing scalar, or several curves moving together, is not the same as the emergence of new organisation. Formation must be an event in which the structural type of an object or family changes under a chosen identity and cross-time transport. The current definition is conditionally closed, but there is no universal neural-specific target to which that definition can be applied.

Results that remain live#

1. Six empirical and gauge-robustness atoms. Relational geometry above a shuffled null was observed in independent vision-only and language-only pairs, together with a typed trend that strengthens during training. This does not imply a shared observer-independent latent space or an exact intrinsic dimension.
2. 44 gauge, function-fibre, and assembly results. These results establish exact and conditional boundaries for architecture-realisable symmetry, local-to-global lifts, function fibres and quotient rigidity, and generic assembly. The same function need not force the same raw geometry; a non-gauge residual does not automatically imply meaning, a circuit, or a ULR.
3. Three observer-relative role and port boundaries. The reset response $HRb$, partial-port response $HRB$, and measurability–total variation–adaptive policy theorem turn the learning–inference distinction into an identifiability problem relative to a declared observer and experiment.
4. One small-neural application. In an engineered matched pair, a neural bridge in which passive-equivalent systems separate under reset and clone responses was established. However, standard controlled testing of the same probe and architecture routing absorbed its predictive gain, so Z1 is WEAKENED.

For the detailed mathematical connections, read the mathematical flow. For individual atoms, see the claim ledger; for experimental verdicts, see the M1–M8 programme.

The explanation adopted now—and the overstatements rejected#

The present winner is Z0 typed pluralism/deflation. Rather than explaining all four questions as one hidden substance, it treats them separately through representation similarity, function-fibre identifiability, testing, realisation, causal abstraction, and system-family theory.

The current Canon therefore does not claim:

• that ULR is CKA, RDM, or a shared representation space itself;
• that the same function implies the same raw geometry;
• that the unexplained 75% is therefore the size of a ULR;
• that co-onset across three metrics alone proves formation;
• that the concepts of a response quotient or predictive state are themselves the novelty of ULR;
• or that using the language of bundles, quotients, or categories creates a neural ontology.

What changed from SCC#

On July 24, 2026, the programme was re-centred on ULR alone. SCC's cohesion, formation predicate, intervention certification, causal unity, and null-model stack were not deleted; they are preserved as a historical evidence layer. That layer no longer determines the current research verdict or its execution priorities. The reasons for the transition and the scope of what was preserved are documented in From SCC to ULR.

What is complete and what remains open#

There is no mandatory ontology workstream. M1–M8 are closed, and research after Canon 24 is an optional frontier kept separate from the canonical state. Of those materials, 119–142 are tracked in Git, 143 is tracked but under modification, and 144–159 are local, uncommitted, and untracked. Internal PASS and sealed labels in the latter group indicate verification stages only; they do not mean canonical, committed, or public promotion.

Any new candidate must preregister object identity, gauge and physical port, B0–B4 and combined B*, nonreduction, held-out behaviour, external boundaries, rivals, falsifiers, and termination before the result is known. Any claim of formation additionally requires cross-time transport and causal manipulation.

The Post-Canon frontier preserves follow-up work 119–159 in a separate layer without changing Canon 24. Open problems and restart conditions explains what the current Main research programme must prove or refute next.