Skip to content

README.md

Latest commit

 

History

History
166 lines (121 loc) · 6.07 KB

README.md

File metadata and controls

166 lines (121 loc) · 6.07 KB
Raw

Backbone Conditional Alpha Solver With Informed Bank

Purpose

This experiment is the integrated Layer 3 pass on top of the new upstream stack:

  • Layer 1: persistent-mode informed bank
  • Layer 2: specialized informed-bank gate
  • Layer 3: candidate-conditioned conditional alpha refinement

The question is whether the informed bank plus a tighter Layer 2 rule can make the Layer 3 point solver behave cleanly on the hard sparse_partial_high_noise branch.

Scope

This run stays on the hard fresh branch that the informed-bank atlas already covers:

  • condition: sparse_partial_high_noise
  • alpha strength: moderate
  • skew bins:
    • low_skew
    • mid_skew
    • high_skew
  • splits:
    • holdout
    • confirmation

It is a cached integrated pass:

  • Layer 1 candidate families come from the informed-bank atlas
  • Layer 2 metrics come from the informed-bank trial table
  • Layer 3 reruns only the local candidate-conditioned refinement

Layer 2 Rule

This version uses the specialized informed-bank Layer 2 rule from backbone observability gate informed-bank specialized ratio sweep:

  • metric: mean_candidate_count * mean_anchored_alpha_log_span / mean_anchored_effective_count
  • threshold: 0.700453
  • classifier direction: le
  • open comparator: gt

The gate opens only when that ratio is strictly above the threshold. The classifier and open rule are now serialized separately, so threshold ties stay on the classifier side instead of being reopened by a direction flip.

This run also restores the intended Layer 3 refinement weighting:

  • refined seed weights use exp(-score_offset / band)
  • band comes from the observation regime, matching the earlier conditional solver math
  • the Layer 2 rule is loaded from the serialized specialized sweep output rather than from a hardcoded override
  • the all_refine artifact is truly gate-independent, so every fresh trial now has a real refined output cached even when the current Layer 2 rule keeps it closed

Main Result

The math-corrected integrated pass is more conservative than the earlier optimistic version.

From backbone_conditional_alpha_solver_informed_bank_summary.json:

  • nominal final bank size: 300
  • mean band candidate count: 209.9259
  • point-output count: 5 / 18
  • point-output rate: 0.2778
  • gate precision: 0.8000
  • gate reject-unrecoverable rate: 0.9000
  • holdout classifier balanced accuracy: 0.4167
  • confirmation classifier balanced accuracy: 0.2000
  • best open-trial alpha error: 0.2075
  • anchored open-trial alpha error: 0.1356
  • refined open-trial alpha error: 0.1807

So the corrected stack still gives a reasonably selective gate, but the current Layer 3 refinement does not beat the anchored answer overall on the opened trials.

By Split

Holdout

  • point-output count: 2 / 9
  • point-output rate: 0.2222
  • gate balanced accuracy: 0.5833
  • gate precision: 0.5000
  • best open-trial error: 0.1484
  • anchored open-trial error: 0.0406
  • refined open-trial error: 0.1499

Confirmation

  • point-output count: 3 / 9
  • point-output rate: 0.3333
  • gate balanced accuracy: 0.8000
  • gate precision: 1.0000
  • best open-trial error: 0.2468
  • anchored open-trial error: 0.1989
  • refined open-trial error: 0.2013

The confirmation side remains close, but holdout is clearly not there yet.

Cell Read

The gate opens in five fresh cells:

  • holdout low_skew
  • holdout high_skew
  • confirmation low_skew
  • confirmation mid_skew
  • confirmation high_skew

The cell picture is mixed, not uniformly good:

  • holdout high_skew: refined beats best but not anchored
  • confirmation low_skew: refined beats both anchored and best
  • confirmation mid_skew: refined loses to both anchored and best
  • confirmation high_skew: refined beats both anchored and best

Interpretation

This is still a useful integration result, but it is a corrective one.

What it now shows is:

  • Layer 1 improves the candidate family
  • Layer 2 can be serialized cleanly and consumed without hardcoded drift
  • the restored Layer 3 math is less flattering than the earlier cached result

The current open problem is not gate serialization anymore. That part is fixed, and the cached refinement path is now complete enough for downstream Layer 2 to Layer 3 rule selection. The remaining question is whether Layer 3 can produce a net-improving refined answer under the corrected weighting math.

So the next implementation move is:

  1. keep this Layer 2 rule as the current informed-bank source of truth
  2. revisit Layer 3 correction behavior under the restored band-scaled weights
  3. only widen coverage after refinement is net-positive again

Artifacts

Data:

Figures:

Code: