Skip to content

README.md

Latest commit

 

History

History
123 lines (81 loc) · 3.63 KB

README.md

File metadata and controls

123 lines (81 loc) · 3.63 KB
Raw

Persistent-Mode Informed Bank

Purpose

This experiment is the first direct informed-bank test for the backbone-first solver program.

The current atlas work established two facts:

  • the hard branch carries persistent candidate modes rather than random clutter
  • compression load is materially higher there

The question here is narrower:

  • can a scout-conditioned, mode-aware bank reduce Layer 1 compression load
  • without breaking backbone recovery

Design

The experiment compares two bank constructions on the hard branch of the current focused moderate-anisotropy slice:

  • sparse_partial_high_noise

  1. one_shot_random

    • the current baseline
    • one random bank of size 300

  2. persistent_mode_informed

    • the informed-bank candidate
    • a scout bank of size 120
    • preserve 72 carryover anchors chosen across scout clusters
    • add 192 local proposals around scout modes
    • keep 36 random exploration entries

This first pass uses 3 independent bank seeds so it can run as a focused pilot rather than a full production benchmark.

The informed stage uses only the scout-family structure:

  • geometry clusters
  • cluster mass
  • cluster size
  • cluster geometry span
  • cluster alpha span

The cluster archetypes are the same first-pass vocabulary from the atlas mining work:

  • dominant_core
  • broad_fan
  • alpha_fan
  • compact_minor
  • fringe_singleton

Acceptance Rule

The first informed bank counts as successful only if it does both:

  • cuts hard-branch compression load by at least 20% relative to one_shot_random
  • preserves hard-branch backbone geometry MAE within 5% of one_shot_random

The hard branch here is:

  • sparse_partial_high_noise

The experiment uses only the fresh evaluation blocks because this first bank design does not tune any calibration-side thresholds.

The primary checks are:

  • holdout
  • confirmation

Result

The first pilot did not clear the precommitted 20% rule on both fresh blocks, but it did produce a strong confirmation-block win while preserving backbone quality.

Fresh-block compression load:

  • holdout

    • one_shot_random: 23.3228
    • persistent_mode_informed: 21.4700
    • reduction: 7.94%

  • confirmation

    • one_shot_random: 31.7663
    • persistent_mode_informed: 16.1310
    • reduction: 49.22%

Fresh-block backbone geometry MAE:

  • holdout

    • one_shot_random: 0.07409
    • persistent_mode_informed: 0.07442
    • ratio: 1.0043

  • confirmation

    • one_shot_random: 0.07573
    • persistent_mode_informed: 0.07361
    • ratio: 0.9721

So the pilot result is:

  • 20% target: failed because holdout stayed below threshold
  • geometry-preservation rule: passed
  • confirmation behavior: very strong

This means the informed-bank idea is real, but the first allocation rule is not yet stable enough across fresh blocks to count as solved.

Artifacts

Code:

Outputs:

Figures: