A Predeclared Negative Result for Paid Calibration of Token-Sparse Planning Readouts in a Frozen JEPA World Model

Separates cost projection from sparse prediction; paid calibration and later schedules fail promotion gates, and a single-seed signal fails replication. An internal candidate claim was withdrawn, not an external paper retracted.

Historical anonymous reading copy explicitly labelled unsubmitted. Internal withdrawal of a candidate superiority claim is not an external paper retraction.

Why we tried it

Separate cost projection from sparse predictor inputs and test whether paid calibration yields task and runtime gains.

What the evidence shows

One three-seed comparison gave full/gradient/random success 33/48, 21/48 and 15/48; a separate schedule experiment gave full 34/48 versus t8 28/48 and failed expanded replication.

Where the work stands

Near-full-success reranking variants were 3%–8% slower; the second domain lacked a usable dense floor, so no sparse comparison was made. Internal claim withdrawal is not an external retraction.

Possible next steps

First establish a competent full-readout floor, then freeze candidates, randomness, budgets and endpoints and compare actual runtime and task outcomes across seeds.

Suggested next steps are untested research questions, not new experimental results. The archive records manuscript provenance, licensing and status.