Skip to content

Latest commit

 

History

History
184 lines (120 loc) · 6.73 KB

File metadata and controls

184 lines (120 loc) · 6.73 KB

Backbone Correction-Pressure-Triggered Alpha Solver

Purpose

This experiment is a bounded Layer 3 follow-up in the backbone-first solver program.

The earlier Backbone Correction-Flux-Triggered Alpha Solver showed that raw flux is a real signal, but also suggested that flux alone is too primitive.

This variant tests one simpler claim:

  • the higher Layer 3 activation is governed by a hidden ratio, not by raw flux alone

Research Question

If the existing Layer 3 anchored and refined candidates are held fixed, can a simple correction-pressure ratio beat the raw-flux trigger on both fresh blocks?

Layer Position

This is still Layer 3 of the staged solver plan in docs/SOLVER_CHALLENGES.md.

Layer target:

  1. stable backbone recovery
  2. extension-coordinate observability gate
  3. conditional alpha recovery inside the gate-open region

Not attempted here:

  1. full confirmation-stable solver policy across the whole focused slice

Method

This experiment reuses the cached output bundle from Backbone Correction-Flux-Triggered Alpha Solver. That keeps the candidate generator fixed and changes only the final Layer 3 control quantity.

The pressure metric is:

  • pressure = (correction_flux * correction_sign_majority) / mean_anchored_alpha_log_std

where:

  • correction_flux is the mean absolute anchored-to-refined bank move
  • correction_sign_majority is the fraction of banks agreeing on the dominant correction direction
  • mean_anchored_alpha_log_std is the Layer 2 gate uncertainty already measured by the anchored posterior

The trigger rule is:

  • refine if pressure >= threshold
  • otherwise keep the anchored answer

The threshold is selected on calibration gate-open trials only, with the same objective as the raw-flux variant:

  • minimize mean open-trial alpha output error
  • tie-break toward sparser switching

The executable artifact is run.py.

Main Result

The pressure ratio is materially stronger than raw flux.

The summary file is backbone_correction_pressure_triggered_alpha_solver_summary.json.

Global result on the gate-open region:

  • trial count: 72
  • point-output count: 53
  • point-output rate: 0.7361
  • calibration-frozen pressure threshold: 0.2427
  • trigger fire rate on gate-open trials: 0.5849
  • anchored ensemble alpha error: 0.1432
  • always-refine ensemble alpha error: 0.1364
  • raw-flux-triggered ensemble alpha error: 0.1358
  • pressure-triggered ensemble alpha error: 0.1331
  • anchored ensemble bank log-span: 0.0684
  • raw-flux-triggered bank log-span: 0.1031
  • pressure-triggered bank log-span: 0.1408
  • always-refine bank log-span: 0.1738

That is the core result.

The pressure ratio is now the strongest Layer 3 trigger tested so far. It improves on the raw-flux trigger on both fresh blocks and overall.

By Split

  • calibration:

    • trigger fire rate: 0.6000
    • raw-flux-triggered output error: 0.1281
    • pressure-triggered output error: 0.1279
  • holdout:

    • trigger fire rate: 0.6429
    • anchored output error: 0.1598
    • always-refine output error: 0.1650
    • raw-flux-triggered output error: 0.1592
    • pressure-triggered output error: 0.1584
  • confirmation:

    • trigger fire rate: 0.5000
    • anchored output error: 0.1353
    • always-refine output error: 0.1163
    • raw-flux-triggered output error: 0.1262
    • pressure-triggered output error: 0.1172

The split read is precise:

  • on holdout, the pressure ratio is the best of all tested Layer 3 policies
  • on confirmation, it nearly matches always-refine and is much stronger than the raw-flux trigger
  • on both fresh blocks, it beats the raw-flux trigger

Stability Tradeoff

The pressure ratio spends more correction capacity than raw flux, but still stays meaningfully below always-refine.

Open-trial bank log-span:

  • anchored output: 0.0684
  • raw-flux-triggered output: 0.1031
  • pressure-triggered output: 0.1408
  • always-refine output: 0.1738

So the new metric does pay some stability to recover the missed confirmation gains, but not all the way back to always-refine.

Interpretation

This experiment strengthens the hidden-ratio diagnosis.

Raw flux was real, but incomplete. The more natural control quantity is closer to:

  • correction size
  • times directional agreement
  • scaled by the anchored gate uncertainty

That simple ratio changes the Layer 3 behavior in exactly the direction we wanted:

  • more active than the raw-flux trigger
  • less reckless than always-refine
  • better on holdout
  • much better on confirmation

What This Establishes

This experiment does show:

  • a simple ratio-governed trigger is stronger than raw flux
  • directional coherence matters
  • the Layer 3 control quantity is not just move size
  • the hidden-ratio framing is consistent with the current data

This experiment does not show:

  • that Layer 3 is fully solved
  • that the branch has already cleared the strict fresh-block stop rule
  • that the solver is ready to advance to Layer 4

Figures

The clearest figure is backbone_correction_pressure_triggered_alpha_solver_alpha_error.png, because it shows the ratio-based trigger sitting between raw flux and always-refine in exactly the way the hidden-ratio idea predicts.

Artifacts

Data:

Code:

Outcome

This result is significant.

The correction-event branch did not collapse into ad hoc patching. It simplified into a cleaner ratio law that beats the raw-flux trigger on both fresh blocks.