Skip to content

README.md

Latest commit

 

History

History
122 lines (79 loc) · 4.27 KB

README.md

File metadata and controls

122 lines (79 loc) · 4.27 KB
Raw

Backbone Bounded Correction Alpha Solver With Informed Bank

Purpose

This experiment revisits Layer 3 after the corrected informed-bank integration.

The corrected stack fixed the math and the Layer 2 handoff, but it also showed that full Layer 3 refinement is still too aggressive on the current open set.

This run tests a simpler rule:

  • keep the informed-bank Layer 2 open set unchanged
  • keep anchored and refined Layer 3 outputs unchanged
  • scale the refined move down when it tries to move too far from anchored
  • freeze the correction budget tau on a separate informed-bank calibration block before scoring fresh holdout and confirmation rows

Method

For each gate-open trial:

  1. measure correction excursion in log-alpha space
    |log(refined) - log(anchored)|

  2. normalize it by anchored alpha span
    excursion_ratio = excursion / mean_anchored_alpha_log_span

  3. apply a bounded correction weight
    w = max(0, 1 - excursion_ratio / tau)

  4. output the bounded log-alpha blend
    log(bounded) = log(anchored) + w * (log(refined) - log(anchored))

So large excursions collapse back toward anchored, while small excursions keep most of the refined move.

Calibration-Frozen Tau

tau is now chosen only on the informed-bank calibration block:

  • calibration rows: 18
  • calibration open rows: 7
  • frozen tau = 0.054314
  • calibration mean bounded open-trial error: 0.1223

So the current bounded result is no longer tuned on fresh evaluation rows.

Main Result

From backbone_bounded_correction_alpha_solver_informed_bank_summary.json:

  • mean open-trial anchored alpha error: 0.1356
  • mean open-trial full refined alpha error: 0.1807
  • mean open-trial bounded alpha error: 0.1333

So bounded correction beats both:

  • anchored by 0.0023
  • full refinement by 0.0474

That keeps bounded correction as the strongest Layer 3 behavior we have on top of the corrected informed-bank stack, now under a calibration-frozen control parameter.

Split Read

Holdout

  • anchored open-trial error: 0.0406
  • full refined open-trial error: 0.1499
  • bounded open-trial error: 0.0406

On holdout, bounded correction effectively collapses back to anchored.

Confirmation

  • anchored open-trial error: 0.1989
  • full refined open-trial error: 0.2013
  • bounded open-trial error: 0.1951

On confirmation, bounded correction keeps the useful small moves while suppressing the overreach.

Interpretation

This is the cleanest Layer 3 result after the math fix.

The key point is:

  • the problem was not that refinement existed
  • the problem was that full refinement moved too far when correction excursion got large relative to anchored span
  • once tau is frozen on calibration, the gain shrinks slightly but remains real on the same fresh open set

So the useful Layer 3 action is not all-or-nothing refinement. It is bounded correction.

That fits the current stack very naturally:

  • Layer 1 gives the better family
  • Layer 2 decides whether point output is allowed
  • Layer 3 applies only the amount of correction the anchored state can absorb

Artifacts

Data:

Figures:

Code: