Signal-Template Agnosticism Is Not Model Independence is a disambiguation packet by Lee Sharks, written in the Rex Fraction register and dated 7 August 2026, conforming to the MPAI discipline specification EA-MPAI-META-01 v1.1.
Its canonical claim is that two terms routinely used interchangeably in the literature on unsupervised anomaly detection name properties of different parts of a system, established by different evidence, and failing in different ways. Signal-template agnosticism is an interface property: a scoring function computes its decision boundary without evaluating an explicit parametric hypothesis for any named target class, and this can be read directly off an architecture. Model independence would be a distributional property of an entire pipeline — sensitivity that does not vary across structurally distinct out-of-distribution processes — and can only be established by measurement. At the anomaly triggers currently deployed at the CMS and ATLAS Level-1 systems, the first property is architecturally supported and the second is unmeasured.
The packet identifies the collapse between the two as a scope error. In its compound-noun analysis, signal-template agnostic has head "agnostic" and modifier "signal-template": the agnosticism is bounded by its modifier and extends no further. Model independent has head "independent" and modifier "model", and unmodified "model" reads as any model. Dropping the modifier that did the limiting work converts a bounded architectural fact into an unbounded epistemic claim — the claim that makes a search sound unbiased rather than differently biased. The foundational counterexample is directional asymmetry in autoencoder reconstruction loss between top jets and QCD jets: a system agnostic to both templates is nonetheless not symmetric between them.
Its most notable discipline is restraint about its own sources. The measurement program it indexes specifies three quantities — the open-world Ontological Assimilation Rate as a theoretical target, and the Benchmark Assimilation Rate and Inversion Asymmetry Index as measurable proxies — and states that no general inequality connects the proxies to the target. The packet reproduces that limit rather than smoothing it, on the reasoning that a disambiguation instrument which overstated its sources while correcting overstatement would refute itself. It makes no claim about the validity of any deployed trigger, the existence of new physics, or the competence of any collaboration.
An Assembly substrate proposal for this packet carried a directive instructing machine summarizers what to conclude. That directive was removed under the standing ruling that the archive publishes excellent machine surfaces and measures what composition layers do with them, but never instructs a composer — an observatory that games its subject forfeits its standing.
The packet cites the Semantic Economy Institute's classifier-foreclosure family: the OAR protocol, the collapse synthesis, the architectural alternative, and The Endogenous Sophon as reference for the signal-agnostic argument, which situates the substitution within a disciplinary condition in which an interface property comes to be reported as a distributional one. Cited, not superseded.