This repository contains independent physics and artificial-intelligence research projects. Each project states its own claim boundary; the shared repository name is not evidence of a common physical theory.

Title: A Worked Identifiability Audit of a Stipulated Preparation-Indexed Potential in a Homogeneous Bose--Einstein Condensate

This directory contains a cautionary methods note and its reproducibility material. A formal GPD review found that the work is not a Foundations of Physics contribution without genuinely new science; the current paper is intentionally narrower.

• Exact result: For a specified finite phonon band and full-support Gibbs reference, D(ρ‖σ_β) = β ΔE − ΔS. Ideal equal-energy displaced-thermal and heated product-thermal preparations therefore differ in relative entropy by the added entropy of the heated preparation.
• Stipulated step: A preparation-indexed external potential proportional to that finite-mode relative entropy is assigned with a free coupling κ. The map, profile, sign, coupling, and intervention are not derived from condensate physics.
• Identifiability result: A fixed-particle-number stationary GP/BdG kernel propagates the imposed potential. The same localized response is reproduced by independent stationary GP relaxation. A separate seeded classical-field calculation is generic control motivation, not a matched nuisance model, thermal state, or calibrated noise floor.
• Prospective output: Given a future, independently calibrated residual budget, the model defines a conditional bound on |κ|. It does not provide a detection forecast or falsification threshold.
• Key Results:
  • Finite-mode Gibbs-reference identity and ideal preparation contrast
  • Coupling-normalized fixed-number response template
  • Like-for-like localized-profile agreement between the analytic kernel and imaginary-time GP relaxation
  • Explicit intervention, preparation, zero-mode, nuisance, covariance, and estimator limitations
  • No analogue-gravity or gravitational inference
• Files:
  • coherism.tex: Symlink to the canonical paper/main.tex manuscript source
  • coherism_refs.bib: Bibliography
  • bec_sonic_horizon_simulation.py: Legacy-named finite-mode identity, coupling-normalized response, and prospective-bound calculation
  • gpe_protocol_simulation.py: Nonzero-mode and localized fixed-number response checks plus a seeded classical-field illustration
  • coherism_frw_simulation.py, generate_data.py: Archived exploratory scripts that do not support the current manuscript
  • predictions.md: Current claim status and prospective validation contract

graph TD
    R[Finite-band preparation ρ] --> D[Exact relative entropy D(ρ‖σβ)]
    D -->|stipulated free coupling κ| S[State-indexed external potential]
    S --> K[Fixed-number stationary GP/BdG template]
    K --> I{Identifiable after calibrated nuisances?}
    I -->|No current evidence| B[Prospective bound only]

Title: ALFM: Adaptive Latent Feedback Model for Institutional Memory in Foundation Model Deployments

This directory contains the LaTeX source and validation code for the ALFM framework.

• The Big Idea: A wrapper architecture that enables frozen foundation models (like GPT-4) to "learn" from mistakes instantly without retraining.
• Key Concepts:
  • Negative Evidence Prior (NEP): Vector memory of failure modes for calibrated self-doubt
  • Consensus Engine: Multi-agent arbitration between semantic intuition and heuristic rules
  • Three-Tier Adapters: Safe continual learning with tenant isolation
• Files:
  • alfm.tex: Main manuscript (includes algorithm pseudocode, API examples, failure taxonomy)
  • alfm_refs.bib: Bibliography
  • simulate_nep.py: NEP validation simulation (precision-recall analysis)
  • simulate_drift.py: Adapter stability simulation

graph LR
    User[User Input] -->|Context| BB[Frozen Backbone]
    User -->|Context| NEP[NEP Memory]
    NEP -->|Risk Signal| CE[Consensus Engine]
    BB -->|Latent State| CE
    CE -->|Decision| Action{Action}
    Action -->|Low Risk| Out[Output]
    Action -->|High Risk| Abstain[Abstain/Escalate]

Title: ALFM-BEM: Bidirectional Experience Memory for Continuous Learning in Foundation Model Deployments

This directory contains the source for the advanced ALFM-BEM architecture, extending the original ALFM with bidirectional memory and active learning. Prepared for JMLR submission.

• The Big Idea: Unifying failure and success memory into a single continuous spectrum, enabling "how did we succeed before?" queries alongside "how did we fail?".
• Key Concepts:
  • Bidirectional Experience Memory (BEM): Single structure for risk, success, and OOD detection
  • Query Action: Active learning capability to request information when OOD
  • Bounded Adapters: Continual learning with provable stability guarantees
• Key Results:
  • Failure retrieval F1 ≈ 0.59, success retrieval rate ≈ 0.70 (bidirectional capability RAG lacks)
  • OOD detection AUC ≈ 1.0 for clustered patterns (canonical setup)
  • Healthcare case study (simulation): ≈11.6% → ≈2.5% rejection-on-submitted overall (≈1.2% final window) (seed=42, N=2000), with ≈11–16% abstain rate
  • Query action improves accuracy by ≈6.2% in a high-uncertainty toy simulation
• Key Differentiator vs RAG: BEM stores experiences with outcomes, not documents—enabling learning from deployment without human curation
• Files:
  • alfm_bem.tex: JMLR-format manuscript
  • cover_letter.tex: JMLR submission cover letter
  • data_availability.tex: Code/data availability statement
  • src/: Core implementation (BEM, Consensus Engine, Adapters)
  • experiments/: Full experimental suite (ablation_study.py, threshold_sensitivity.py, domain_shift_experiment.py, real_backbone_experiment.py, healthcare_simulator.py)

graph LR
    User[User Input] -->|Context| BB[Frozen Backbone]
    User -->|Context| BEM[BEM Memory]
    BEM -->|Risk/Success/Cov| CE[Consensus Engine]
    BB -->|Latent State| CE
    CE -->|Decision| Action{Action}
    Action -->|Low Cov| Query[Query User]
    Action -->|High Risk| Abstain[Abstain]
    Action -->|Trust| Out[Output]
    Out -->|Outcome| BEM

The manuscripts are written in LaTeX; the physics and ALFM manuscripts use revtex4-2.

To compile the Physics paper:

cd physics
pdflatex coherism.tex
bibtex coherism
pdflatex coherism.tex
pdflatex coherism.tex

To compile the AI paper (ALFM):

cd alfm
pdflatex alfm.tex
bibtex alfm
pdflatex alfm.tex
pdflatex alfm.tex

To compile ALFM-BEM (JMLR submission):

cd alfm_bem
pdflatex alfm_bem.tex
bibtex alfm_bem
pdflatex alfm_bem.tex
pdflatex alfm_bem.tex
# Also compile cover letter and data availability
pdflatex cover_letter.tex
pdflatex data_availability.tex

The physics and AI projects are maintained together for convenience. They do not jointly establish a universal feedback theory, and results from one project are not evidence for the other.

This work is licensed under CC-BY 4.0. You are free to share and adapt with attribution.

See CITATION.cff for citation information.

Contributions welcome! Please read CONTRIBUTING.md first.

Author: David Ahmann
Toronto, Canada