Deterministic property
A mathematical or structural property established by derivation and reproducible tests.
Status system
Every result says what kind of evidence it is.
Mathematical properties, benchmark measurements, partial causal evidence, prototypes, and retired ideas require different language. The label is part of the result.
Observed benchmark
A reported evaluation result with a known setup; not automatically a general capability.
Evidence with open controls
A promising result that still needs matched ablations, multiple seeds, or broader datasets.
Active experiment
An idea, prototype, or mechanism still being evaluated and not ready for adoption claims.
Did not survive testing
A claim or direction kept in the record because knowing what failed is useful evidence.
Baseline before story
Chance, simple linear methods, ordinary neural probes, leakage controls, and matched comparisons come first.
Measured AI
One result, its baselines, and the missing proof.
Supervised Collapse is an experimental learned-attractor classifier. The reported SNLI result is real, but the current record does not establish a unique causal advantage for collapse.
Did attractor collapse outperform an ordinary MLP?
Not in the current reported experiment. Collapse reached 68.92% on the SNLI test set; the frozen-embedding MLP probe reached 70.13%. Matched multi-seed testing remains pending.
Because the baseline is part of the result.
A result reported without its strongest ordinary baseline is marketing, not measurement. The full comparison, method, and open controls are below and in the repository.
Inspect the method →Supervised Collapse68.92% SNLI testOne reported run; experimental learned-attractor model.
Frozen-embedding linear probe64.06%Simple baseline on the same learned representation.
Frozen-embedding MLP probe70.13%Stronger than Collapse in the reported comparison.
Next required testPENDINGMatched end-to-end, multi-seed ablations across datasets.
Correct interpretation: trained geometry matters for that checkpoint. The current evidence does not show that collapse is uniquely sufficient, state of the art, or a general reasoning mechanism.
Research programs
Models, representations, combinatorics, and methodology.
Each program links back to code or written evidence in the public repository. Research language is intentionally narrower than marketing language.
Supervised Collapse NLI
Learned entailment, neutral, and contradiction attractors evaluated on SNLI, with ordinary probes reported beside the model.
Inspect NLI work →Noun-Collapse representations
Word vectors learned through collapse dynamics and evaluated on semantic-similarity tasks.
View model record →Compact premise generation
A small conditional generation program exploring whether learned-attractor representations can support label-directed text generation.
Vision Collapse
An early investigation of attractor dynamics over image data. Smoke-tested at small resolution; not a trained production vision system.
Ramsey and rotation tools
Solvers, witnesses, and group-action constructions. Standard and verified results are separated from retired novelty claims.
Prediction investigation
A preserved case study showing why a compelling geometric story is not evidence of predictive power.
Known limits
Core preserves structure. It does not learn meaning.
Static Core geometry cannot learn or abstract by design. On artifact-free ANLI it performs around chance. Learned models built above Core are separate systems and must earn their own evidence.
General reasoning from geometry
Core is not an AI model. It does not understand language, discover semantics, separate signal from noise, or replace ordinary learning.
Exact reversible structure
Core supports deterministic encoding, addressing, permutation, conserved weights, and inspectable group actions.
Collaborate with evidence
Have a hypothesis worth testing carefully?
Bring the claim, the strongest ordinary baseline, the data, and the result that would change your mind.