Anchor latency analysis

Why stale anchor learning breaks the merger.

Three reports cover the same failure from different angles: validation collapse, Q-basis stability, and dense-gradient drift under stale anchor cadence.

K=5stable near dense ceiling
K=20collapses below no-merger floor
Q lowcompression basis recovers after bootstrap
20/20delay and cadence confounded
K=20 collapse mechanism chart.
Main report

Delay Failure Report

The detailed argument that K=20 turns the same signed EMA merger from useful correction into a stale, reward-flat length ratchet.

collapse cadence validation
delay_failure_report
Q-basis telemetry chart.
One-pager

K-Instability Q-Basis

A compact readout: K ages the gradient signal, while Q remains a low-error activation basis and cannot judge reward quality.

Q basis off-policy K comparison
k_instability_q_basis
Staleness error versus anchor latency chart.
Joint evidence

Dense Drift Joint

Uncompressed-run drift evidence showing the gradient-anchor route loses useful alignment as lag grows across tasks.

drift GSM8K Big-Math
dense-drift-joint