README.md

Candidate Atlas Pattern Mining

Purpose

This experiment mines the instrumented candidate atlas for recurring cluster types and trial-level burden ratios.

The atlas instrumentation already showed that the current near-best family is large, structured, and branch-dependent. This follow-up asks the next question:

• what recurring candidate modes are actually present
• which trial-level ratios extracted from those modes carry solver-relevant signal
• whether the first poison-candidate idea should move from single rows to cluster-level burden

Research Question

If the candidate atlas is reduced to recurring cluster archetypes, does that reveal a simpler trial-level structure that can guide an informed bank and improve downstream solver decisions?

Method

This experiment reads the existing atlas outputs from Candidate Atlas Instrumentation and joins them to the current Layer 3 pressure-trigger outputs.

The first-pass archetype rules are empirical and explicit:

• dominant_core
  • cluster_mass_layer1 >= 0.18
  • cluster_size >= 8
• broad_fan
  • cluster_alpha_span >= 0.35
  • cluster_geometry_span >= 0.12
• alpha_fan
  • cluster_alpha_span >= 0.25
• fringe_singleton
  • cluster_size <= 2
  • cluster_mass_layer1 <= 0.02
• otherwise compact_minor

From those archetypes, the experiment builds trial-level burden metrics such as:

• fan_vs_core
• useful_structure_ratio
• residual_shell_alpha_mass

The executable artifact is run.py.

Main Result

The atlas is not just a large family. It contains recurring archetypes, and the strongest new trial-level ratio is a competition between broad_fan mass and dominant_core mass.

The summary file is candidate_atlas_pattern_summary.json.

Global result:

• trial count: 72
• gate-open trial count: 53
• archetype counts:
  • fringe_singleton: 1551
  • compact_minor: 1288
  • alpha_fan: 1182
  • broad_fan: 1036
  • dominant_core: 383

On gate-open trials:

• fan_vs_core vs improvement correlation: 0.2828
• fan_vs_core vs correction-pressure correlation: 0.1803
• useful_structure_ratio vs improvement correlation: 0.2797
• useful_structure_ratio vs refined-bank-span correlation: 0.3657

That is the core result.

The atlas is expressing a real structural competition:

• a stable core mass
• versus a broad fan of still-active alternative structure

Archetype Read

Mean archetype shape:

Archetype	Size	Mass	Alpha Span	Geometry Span
dominant_core	17.44	0.2476	0.4701	0.1959
broad_fan	9.94	0.0977	0.5357	0.1743
alpha_fan	4.14	0.0648	0.3938	0.0939
compact_minor	2.56	0.0595	0.0849	0.0574
fringe_singleton	1.19	0.0069	0.0184	0.0096

The most important split is:

• broad_fan is heavily concentrated in sparse_partial_high_noise
• compact_minor is much more common in sparse_full_noisy

That means the hard branch is not just “larger.” It is organized around a stronger broad-fan structure.

By Condition

The branch split is sharp.

• sparse_full_noisy

  • holdout mean fan_vs_core: 0.3824
  • confirmation mean fan_vs_core: 0.4371
  • holdout mean useful_structure_ratio: 0.8417
  • confirmation mean useful_structure_ratio: 0.9825

• sparse_partial_high_noise

  • holdout mean fan_vs_core: 0.4829
  • confirmation mean fan_vs_core: 0.6612
  • holdout mean useful_structure_ratio: 1.3192
  • confirmation mean useful_structure_ratio: 1.6637

The confirmation block matters most.

The hardest branch is carrying much more broad-fan mass relative to the dominant core on confirmation.

Atlas-Only Trigger Probe

The strongest simple atlas-only trigger in this pass is:

• refine if fan_vs_core >= threshold

Calibration-frozen threshold:

• 0.4107

Gate-open errors:

• calibration: 0.1278
• holdout: 0.1587
• confirmation: 0.1182

This is not better than the current pressure trigger. But it is close enough to matter.

The atlas alone is carrying a real part of the Layer 3 activation signal.

Poison Heuristic Shift

The first candidate-level poison score did not tell us much:

• old candidate poison score vs improvement: 0.0245
• old candidate poison score vs correction pressure: -0.0009

The better replacement in this pass is not a single bad candidate score. It is a cluster-level residual-shell burden:

• residual_shell_alpha_mass

This measures the mass sitting in the higher-score shell inside alpha-wide clusters.

It performs materially better:

• residual_shell_alpha_mass vs improvement: 0.1394
• residual_shell_alpha_mass vs correction pressure: 0.2664

That is the important shift.

The atlas is saying:

• the harmful or unresolved structure is not well described as one poison candidate
• it is better described as residual shell mass in alpha-wide clusters

Interpretation

This is the strongest evidence yet that the bank should be informed by cluster structure rather than by raw sample density.

The current random bank is surfacing:

• dominant cores
• broad fans
• alpha fans
• compact minors
• fringe singletons

The next informed bank should use that structure directly.

The most actionable design signal is:

• high fan_vs_core means the observation is carrying more broad active structure than stable core structure

That is exactly the kind of state variable a backbone-conditioned informed bank should react to.

What This Establishes

This experiment does show:

• the candidate atlas contains recurring cluster archetypes
• broad_fan mass is a defining structural feature of the hard branch
• fan_vs_core is a real trial-level ratio with downstream solver relevance
• the first poison-candidate framing should move from single rows to cluster-level residual-shell burden

This experiment does not show:

• that the archetype thresholds are final
• that fan_vs_core should replace the current pressure trigger outright
• that the informed bank is already solved

Figures

• candidate_atlas_archetype_mass_by_condition.png
• candidate_atlas_fan_vs_core_scatter.png

The clearest figure is candidate_atlas_archetype_mass_by_condition.png, because it shows the broad-fan expansion in the hard branch directly.

Artifacts

Data:

• candidate_atlas_archetype_rows.csv
• candidate_atlas_trial_patterns.csv
• candidate_atlas_pattern_split_summary.csv
• candidate_atlas_pattern_condition_summary.csv
• candidate_atlas_fan_rule_rows.csv
• candidate_atlas_pattern_summary.json

Code:

• run.py

Next Step

The next informed-bank pass should not sample globally.

It should use the atlas pattern state directly:

1. preserve dominant_core coverage
2. allocate extra density along broad_fan structure
3. treat residual shell alpha mass as unresolved burden, not as a one-row poison event