Research
Defined hypotheses, controlled comparisons, and reproducible measurements.
In plain terms
Each line of research is a question paired with the test that could prove it wrong. Nothing here is treated as settled until that test has been run on material the method has never seen.
Direction relative to an outcome
H1: a projection-weighted signal improves outcome prediction over volume alone on a stated coordinate basis. The hypothesis is rejected if it adds no out-of-sample improvement. Outcome labels, fitting procedure, and uncertainty intervals are defined before evaluation.
Temporal change
H2: acceleration improves alert lead time over volume thresholds at matched false-alarm rates. The hypothesis is rejected if that advantage disappears on future observations. Independent measurement errors give acceleration noise √6σ/h²; correlated errors use the covariance expression in WP-01.
Subject relevance and negation
R1 compares subject guards with word overlap on held-out pairs. R2 compares explicit readings with cosine-only decisions on negation and critical-field changes. Each proposed benefit is rejected if the selected method fails to improve its prespecified baseline. Specialist contradiction methods are included where appropriate.
Reader calibration and dependence
R3 tests confidence corrections against label-only weights. R4 measures within-family and across-family correlations after accounting for item difficulty. Their proposed benefits are rejected if calibration or grouped weighting fails to improve held-out decisions. Family boundaries are grouping inputs rather than a guarantee of independence.
Basis stability and reason relevance
R5 evaluates decision changes across embedding bases, with more than five percent flips exceeding the proposed stability criterion. R6 tests off-subject reason detection and incremental decision value. Its benefit is rejected if the guard fails to distinguish the test groups or improve decisions beyond labels alone.
Origin recovery and decision thresholds
R7 tests controlled copies, paraphrases, false splits, and over-merges against known source maps. The operational benefit is rejected if origin errors defeat repetition resistance or false Allow exceeds the selected tolerance. R8 freezes thresholds before holdout comparison; its benefit is rejected if the policy fails to outperform the selected majority baseline.
Study interpretation
The reported 60-statement study and five-model embedding comparison motivate this research design. They are preliminary, reported measurements. Item-level reproduction, label provenance, and independent holdout results are needed to establish broader performance. Optional multiscale and symbolic models are evaluated separately from the core procedure.