Causal rather than decorative

AI experiments

A graph, a fluent answer, or a benchmark gain is not enough. The retained formation must mediate a predicted difference.

Status: DESIGNS + NARROW FROZEN RESULTS
Evidence class: Per experiment
Current question: Can one formed susceptibility cause a later reasoning difference?
Current attack: Host policy, transcript leakage, retrieval, or unequal budget
Next test: Primary MFM reversal/recovery/transplant/ablation sequence
Code / data: Linked where public
Baseline: RAG, RETE, production, procedural, temporal KG, pub-sub, bare host
Null policy: Preserve it

1. Bare-host profile

Measure capability and host policy before formation. If the host already supplies the effect, acquired formation is not identified.

2. Reversal / negative transfer

Train one history, reverse the contingency, and predict where prior formation should initially hurt.

3. Recovery formation

Test whether subsequent formation produces faster or structurally different recovery.

4. Narrow transplant

Move one validated susceptibility or bounded formed object and predict the transferred difference.

5. Ablation

Remove the same object and require the predicted effect to disappear.

6. Matched baselines

Compare quality, latency, compute, storage, update, correction, and audit costs across native alternatives.

Current experiment record

Narrow frozen causal-state and cold-restart continuity fixtures. These establish only their declared behavior.

FROZEN / TESTED

A deterministic convergence/dissolution instrument is ready. It did not test the confirmatory history effect.

INSTRUMENT_READY

Cross-host design separates formed history from policy and includes strong controls. Execution remains blocked.

SPECIFICATION / NOT RUN

Sparse recruitment, hard admissibility, wake cost, and ablatable composition require implementation and matched tests.

BUILD-TEST TARGET

First candidate test uses two embodiments, partition/rejoin, bounded authority, and formation-locality predictions.

BUILD-TEST PENDING

Comparator discipline tests what a representation preserves and what becomes residue under a declared target.

BOUNDED ILLUSTRATION