
A closer match is not always a better fit for the task
A candidate gains a capability the task does not need, and its score falls. The arithmetic can be correct: resemblance rewards matching a profile, while requirement satisfaction asks whether enough capability is available. A transparent two-dimensional example separates those objectives and shows why feasibility and ties belong in a selection decision.
Four formulas, two different questions
Let a be capability and b requirement, using dimensionless features. Our internal demonstration explicitly chooses four formulas; these are not reconstructed author implementations. Cosine uses angular similarity; reciprocal Euclidean is 1/(1+distance); coverage is the fraction of positive requirement dimensions met; shortfall is 1 minus total positive deficit divided by total requirement. All examples have a positive requirement denominator. Formulas, inputs and results
The first two compare whole-vector resemblance. The latter two ask about unmet requirements and do not penalize surplus in these examples. Correctness depends on the objective. Similarity is appropriate when resemblance is intended. For sufficiency, a penalty for extra capability needs justification. If that capability adds energy use or cost, it is no longer irrelevant under this example's premise.
Surplus capability lowers similarity in the constructed case
Keep the requirement at (1,0) and increase capability from (1,0) to (1,1). In the internal experiment cosine falls from 1 to approximately 0.707, reciprocal Euclidean from 1 to 0.5, while coverage and shortfall remain 1. The candidate still meets the first dimension completely, but loses resemblance to the requirement vector.
A second case keeps capability at (1,1) and relaxes requirement from (1,1) to (1,0.5). Cosine falls from approximately 1 to 0.949 and reciprocal Euclidean from 1 to approximately 0.667; both sufficiency scores stay at 1. A positive control that fills a required capability deficit improves all four scores. An unchanged-input control leaves all scores unchanged. These controls prevent treating every variation as a defect of similarity.
These deterministic mathematical examples were executed in the same-date Analysis stage, without random sampling, robots or model calls. They establish possible behavior for the listed formulas, not its frequency in real tasks.
A stable score can still change the chosen worker
Another internal construction changes two scores from (0.5001,0.5) to (0.5,0.5001). Each changes by just 0.0001, but the top choice reverses. Small score movement therefore does not establish decision stability. Selection margins, tie rules and changes of winner under perturbation need examination alongside score variation.
An all-zero row exposes another semantic issue: deterministic maximization can still return the first index. If zero explicitly means infeasible, no feasible candidate exists; returning an index is algorithm completion, not evidence that someone can do the task. This interpretation depends on the meaning assigned to zero.
Our interpretation is that declaring hard constraints and infeasibility before explaining the ranking objective makes a selection easier to audit. Coverage cannot distinguish the value of surplus competence, and hard gates can eliminate every candidate. Passing these property tests still does not establish better execution outcomes.
Evidence and sources
The external motivation is the robot-task suitability study, version 1, sections II and IV. It compares requirement-oriented and similarity-oriented methods and examines score versus choice stability; it does not measure execution-derived ground truth. Exact author normalization, code, generator seeds and complete counts are unavailable in the inspected material.
The paper's property-violation frequencies and model results remain source-reported. The values on this page are internal experimental evidence, not independent reproduction of those results. The accompanying artifact retains the four vector controls, two selection demonstrations, formulas and upstream result hash. We did not invoke a language model or reproduce the reported 98.55% frequency. Publication or citation cannot substitute for that reproduction.
Bibliographic record: Robot–Task Suitability in Heterogeneous Multi-Robot Task Allocation, Daniel Weiner and Raj Korpan, CUNY Graduate Center / Hunter College; arXiv preprint v1, 20 October 2026. It motivates the modeling comparison, rather than establishing our formula implementation or execution outcomes.
What must be tested before using the score
The two-dimensional construction includes no contention, coalitions, scheduling or coupled constraints, and compares no actual success rate, safety or allocation benefit. Transfer from robot suitability to software agents remains a hypothesis.
Before adopting a score, state units, normalization, treatment of surplus, hard constraints and abstention conditions, then test whether its ranking predicts real outcomes. For narrow margins, the need for more information or abstention should be an explicit decision. These examples help audit score meaning; they do not identify a best allocator.