Non-record: Cross-Base Regularizer Transferability — methodological study (20+ cells, 10 figures)

[BharathSShankar]

Summary

Methodological (non-record) contribution. Studies how seven novel
regularizers — QAHSP, ES, AOS, HSU, WBC, WOP, PCS — transfer across two
distinct training bases for the 10-min / 16 MB track:

• Base A: SP10240 + SimCTG (PR Record: SP8192 + LQER + Sparse Attn Gate + BOS-Fixed SmearGate + 9-Hparam Greedy Stack — val_bpb 1.06108 (3-seed mean) #1855 lineage).
• Base B: PR Record candidate: long-context no-QV rank56/prefix3000 TTT — val_bpb 1.05875 #1965 phased-TTT lineage (caseops vocab).

The headline finding is that regularizer benefit does not transfer
across bases: a regularizer like QAHSP that helps Base A by ≈0.1 mBPB
moves Base B by ≈ +3.7 mBPB (i.e. it hurts). ES shows a similar but
smaller swing. We also measure a real-data reg × quant interaction
matrix across 7 quantization schemes (int4/6/8 sym/asym per-tensor/per-row,
AWQ-lite, GPTQ-lite) and decompose the eval pipeline into pre-quant
grade → post-quant → sliding-window → TTT stages.

This is a research/analysis contribution intended to inform reviewers
and future submitters about regularizer-base interaction. No model
artifact is shipped.

Contents

• README.md — 17-section writeup with hypotheses, methodology,
  cross-base swings, pipeline-stage attribution, and statistical caveats.
• figures/ — 10 figures (cross-base signs, pipeline waterfall, real
  pre-vs-post quant, PQT × reg compounding, lambda-budget pair-vs-single,
  3D PCA per-reg, canonical metrics, coord and L2-norm distributions,
  reg × quant matrix heatmap).
• *.json — analysis tables (real_canonical_metrics, real_reg_quant_matrix,
  pipeline_attribution, eval_pipeline_breakdown).
• run_reg_quant_matrix.py, build_real_data_figures.py,
  build_synergy_figures.py — reproduction scripts.

What we DO claim

1. The sign and magnitude of regularizer benefit depends on the training
   base; "transferability" is not a free assumption.
2. On Base A, post-quant val_bpb is largely insensitive to which of the
   7 regs is trained with — quant cost is reg-independent within ~14-15 mBPB.
3. PQT × ES compounds modestly beyond PQT alone; PQT × QAHSP does not —
   suggesting direction-shaping vs codebook-shaping regs interact
   differently with TTT.

What we do NOT claim

• Statistical significance from a single seed per cell (most cells are
  seed=42 only). The headline cross-base sign change is robust enough to
  call out, but mBPB-magnitude claims are point estimates.
• A predictive theory of regularizer-base interaction. The synthetic
  embedding geometry analysis is suggestive, not predictive of post-quant
  val_bpb at this scale.
• Recommendations beyond Base A's stack. Numbers are bound to the exact
  cells listed in §6.

Test plan

• Reviewer reads §1 + §6 + §13 (statistical caveats) + §14 (what we
  do NOT claim).
• Reviewer confirms cross-base sign change (Table in §6.5) is the
  load-bearing finding, not the per-cell mBPB numbers.
• Reviewer confirms this is filed under track_non_record_16mb/ as
  a methodological contribution, not as a record entry.

🤖 Generated with Claude Code