An empirical study of delayed generalization, built around fit competition across candidate scaling laws rather than the curve that flatters the story.

• Scaling sweeps, bootstrap intervals, a LaTeX paper.
• Negative and boundary results treated as part of the output.

A semantic-dynamics engine that inserts a testable intermediate representation between a prompt and generated language, with the decoder kept as an explicit boundary.

• Pure engine path runs without API keys; optional prompted and activation-steering decoders.
• Steering diagnostics found shared-direction collapse, then added decorrelation and query-dependent seeding.

A controlled generation study testing whether chaotic logit modulation could make LLM prose feel more alive. The result sharpened the hypothesis: coherence scaffolding mattered more than token turbulence.

• Matched chaos against OU-noise, white-noise, sampling, and prompt-agency controls.
• Negative result preserved: the wrong lever was part of the finding.

Strange attractors, Bessel-mode resonances ("frozen music"), and tensegrity normal modes rendered as light: every visual derived from a verified invariant.

• Browser-native works built from real mathematical systems.
• Math verified before it is drawn.

A coverage-aware computational ethnobotany benchmark for testing structured search over plant data without mistaking database attention for biological signal.

• Decoupled reward provenance, negative controls, and matched-density sweeps.
• Findings separate raw cue-reward coupling from study-effort artifacts.

A research engine that maps ambiguous problems into causal structures, tracks provenance, and keeps hypotheses, simulations, and evidence behind explicit review gates.

• Tree of Knowledge graph for structural signatures, provenance, and causal-coordinate retrieval.
• Research workflows keep candidates, simulations, and evidence transitions deliberately separated.