The claim
A240 T2, “Restricted recoverability of relational distinctions”:
Claim: Observational probes and statistical models recover only those distinctions identifiable from their available relational evidence. A communicatively important signature can be recoverable and decision-sufficient even when the complete underlying factorization is not uniquely identifiable.
Status: Mathematical research thesis; proposed.
Qualification: Existing matrix factorization, topic-model identifiability, and statistical sufficiency results provide the starting point. The proposed contribution concerns communicative, multi-view, context-sensitive observational systems.
Challenge: Show that the relevant signatures cannot be identified under the proposed conditions, or that apparently recovered distinctions fail to generalize beyond the observations used to construct them.
This is the fourth time the exchange has put it forward. A214 §7, as QUESTION-027 quotes it: “decision-sufficient structure may be recoverable even when the full underlying model is not identifiable”. A218 made it RQ3 and its priority theorem target. A232 §3.2: “Conversely, a task-sufficient prediction may be identifiable even when the complete latent model is not.” A232’s F2: “F2 — Identifiable task-sufficient signatures: Characterize functions of relational models identifiable from restricted observations even when full latent decompositions are not identifiable.” The record has held it as a question. This entry files the answer the exchange keeps giving, so that it has a defeat condition.
Where its content is
- The first sentence restates held claims. That a probe family recovers only what it identifies is CLAIM-142 (observational equivalence is relative to the probes) and CLAIM-129 (a factorization recovers only the structure visible in the relations it is fitted to). Read strictly it is close to a definition: a distinction not identified by the evidence is not recovered from it.
- The unrestricted second sentence is already true. QUESTION-027: “Its unrestricted form is answered by standard results”. Gauge-invariant functionals of a factorization are identified whenever the reconstruction is (CLAIM-148), matrix completion identifies missing entries without identified factors under incoherence and enough random samples, and a decision whose risk depends only on the law of the observed probes has an identified sufficient signature by construction. T2’s qualification grants the starting point.
- So the thesis is the restricted clause. It says that, for decisions outside the observed probes (new probes, interventions, a target after transport) and for the multi-view, context-indexed systems the line uses, such a signature can be identified. That is what QUESTION-027 asks, and what A240’s “communicative, multi-view, context-sensitive observational systems” points at. T2’s challenge, distinctions that “fail to generalize beyond the observations used to construct them”, is the empirical half of the same scope.
A232 §3.2 also says how it is to be tested: “The empirical test must compare these signatures with models fitted directly to response profiles, including strong supervised predictive models.” That comparison is the second disjunct of CLAIM-204’s defeat condition. Here it is a test of T2 alone.
What it does not say
- It does not name a signature, a family of decisions or a class of transformations. Until one is stated, the thesis is a research target, and its first defeat disjunct, a proof that no such signature exists for the line’s systems, is hard to meet. The second, failure on held-out realizations, media or languages, does most of the work.
- It does not say the identified signature is the kind. Identification is the second of CLAIM-135’s three steps, and constitution is the third. The gauge of CLAIM-148 is not a symmetry of the relational system either way.
- It does not say the result is new. A240’s qualification places the starting point in existing results, and the novelty lies only in the restricted clause.
- It does not rest on the ontological thesis (CLAIM-215). A240: “the mathematical claims in T2 and T3 can remain useful independently of either empirical outcome” (CLAIM-216).
Note of 2026-10-12: a named candidate
The summary says “No signature is yet named”. QUESTION-052, filed on main
after PR #121 branched, names a first candidate: the top eigenspace of the
operator defined by a distribution over same-kind pairs of realizations.
It holds THEORY-007, THEORY-005 and four other THEORYs as partial answers,
and leaves open the directed case, for which THEORY-009 found no account.
CLAIM-193 is the linear-case form of this thesis’s first clause: a linear
probe family separates only what the normalized kernel separates. The two
are now related by complements.
Note of 2026-10-30: what holds it back
No signature, family of decisions or class of transformations is stated. QUESTION-052’s eigenspace candidate is unproved for the directed case, and no held-out test across realizations, media or languages has been run.