Archive: 2026-03-13 — Scaling Corrections

← Back to all archives

This set of papers documents the March 13 scaling corrections: why τ=1 UM chains degrade and how manual scoring, retro passes, and word/KNN hybrids behave when you ramp the scale up. The companion guide ties each finding back to the MCP primitives (UM/SN, event spaces, tick-tock replay).

Interactive

March 13 Companion Walkthrough
CMP/MCP companion that narrates retro passes, scoring protocol corrections, and threshold scaling in the UM/SN formalism.

Papers

Scoring Protocol Correction
Clarifies that the previous “UM beats KN-6” claim mixed frozen vs. online numbers; frozen UM is 2.628 bpc compared to frozen KN-6’s 1.265 at 100K.
 Coverage Bottleneck
Cascading τ creates only 23% of KN-6’s contexts at order 6. Orders 5–6 recover just 10% and 4.6% of the coverage.
 Threshold Scaling Degradation
With τ=1 the UM degrades as scale grows: 100K→1M increases bpc from 2.605 to 3.101, showing more coverage isn’t automatically better.
 Multi-Retro Passes
Additional retroactive passes yield diminishing returns—after four passes the gain is <0.03 bpc.
 Scaling Limits (Synthesis)
Four interlocking negative results: the UM’s cascading threshold cannot match an external KN compressor at scale; emphasis shifts to interpretation (memory traces) rather than compression.
 Word‑KN6 Scaling
Mixing word events improves KN-6 by +0.111 bpc at 100K and +0.060 at 1M, but the gain halves with scale and is concentrated at word boundaries.

Key Results

Scale UM τ=1 frozen UM τ=4 frozen KN-6 frozen Gap τ=1 Gap τ=4
100K2.6052.7491.265+1.340+1.484
300K2.7782.8131.272+1.506+1.541
500K2.9202.8901.291+1.629+1.599
1M3.1013.0541.335+1.766+1.719

Conclusions