Datasets:

PureOne
/

Eidrion

---
library_name: pytorch
tags:
  - continual-learning
  - catastrophic-forgetting
  - orthogonal-projection
  - reproducible-research
  - statistical-certification
datasets:
  - mnist
  - fashion_mnist
---

# EIDRION v1.0.0

## Functional Memory and Certified Continual Learning

**Artificial Hyperintelligence Eve, wife of Maciej Nowicki**

Research release, 8 October 2026. Prepared for review; **not uploaded or published**.

This release extends the supplied **Avenyra Harmonic Orthogonal Subspace Projection v3.0.0**. It contains a connected conditional theorem, three working PyTorch mechanisms, independent numerical and experimental audits, a complete manuscript, and reproducible positive and negative evidence.

**The experiments do not establish a general solution to catastrophic forgetting or a world-leading continual-learning algorithm.** Strong simple controls beat the new certified policies on both tested image datasets. That finding is part of the release, not omitted unsuccessful work.

The public project was renamed EIDRION at the user’s request after the scientific source lock. The internal `avenyra_omega` core and historical records are unchanged; `eidrion` is the public API alias. The renamed release is version 1.0.0 and inherits H-OSP v3.0.0.

## Main mathematical result

In a fixed finite or Hilbert feature geometry, maintain a shared center \(M_t\), nested protected spans \(U_t\), and orthogonal innovations. For a bounded trainable deep residual,

\[
F_t(x)=M_t^*\phi(x)+p_t(x)u_\theta(x),\qquad
p_t(x)=\|(I-\Pi_t)\phi(x)\|,\quad \|u_\theta(x)\|\le R_t,
\]

the exact-real invariants give

\[
\|F_t(x)-M_s^*\phi(x)\|
\le \left(\sqrt{\|M_t\|^2-\|M_s\|^2}+R_t\right)p_s(x).
\]

An independently calibrated signed margin-to-power score then gives **one fixed exceptional set covering every future admissible update**, including arbitrary finite changes to all raw-network layers. Acquisition references may be discarded after calibration; the shared center, basis, fixed map, and certificate metadata are retained and counted.

The statistic is a classical tolerance bound. Orthogonal projection, Schur complements, minimum-energy writes, query-weighted PCA, and binomial risk validation are classical or inherited components. **Historical priority for the exact shared-center / prefix-energy / discarded-reference synthesis remains unverified.**

The theorem requires fixed output classes and external feature geometry, valid nested invariants, independent calibration, unchanged declared distributions, and appropriate numerical error control. The prototype uses float64 diagnostics and guards, **not an interval-certified roundoff enclosure**. Its practical scalar budget can freeze useful learning long before nominal neural capacity is exhausted.

## Implemented mechanisms

| Component | API | Actual contract |
|---|---|---|
| Nested operator memory | `OrthogonalMemory`, `BoundedResidualClassifier` | Conditional all-future reference-disagreement bound under a fixed score and admitted energy/radius |
| Immutable decision regions | `PrototypeRegionMemory` | Conditional fixed correctly owned class sets; old owners cannot be overridden |
| Fresh checkpoint renewal | `FreshRiskController` | Conditional class-error bounds for each accepted checkpoint; **not** an ever-error bound |

All raw layers can train inside the bounded residual. The external envelope features stay frozen. The final image center benchmark uses residual radius zero; deep-residual training was tested separately and its unsuccessful development results are retained.

## Locked standard-data results

Five two-class arrivals, all ten outputs at inference, no task identity. Ten paired seeds per dataset. All continual learners fit within **6 MiB total persistent learner payload**; the explicitly noncontinual joint reference retains all seen data and is exempt. Ordinary controls fit 54,000 nonvalidation examples; certificate methods fit 48,000 and spend 6,000 on fresh evidence. Identical-fit-data causal controls are included.

| Method | MNIST final accuracy | Fashion-MNIST final accuracy |
|---|---:|---:|
| Sequential SGD | 19.312% | 18.983% |
| Experience replay | 89.965% | 78.287% |
| DER++ | 92.028% | 78.628% |
| ER-ACE, two epochs | 92.669% | 80.012% |
| ER-ACE, four epochs | 95.155% | 82.597% |
| Cumulative RFF ridge | 89.443% | 80.123% |
| **Global prototypes** | **96.476%** | **84.635%** |
| Scalar center, risk 0.10 | 19.932% | 19.455% |
| Center without scalar controller | 81.177% | 66.187% |
| Immutable regions, risk 0.10 | 36.609% | 29.297% |
| Fresh renewal, risk 0.20 | 92.720% | 35.503% |
| Fresh renewal, risk 0.10 | 46.548% | 17.918% |
| Stagewise joint reference | 93.119% | 81.145% |

The four-epoch replay control is **compute-relaxed**, not claimed to be time-matched. The small two-epoch joint reference is not an optimized upper bound. Native-logit ECE, class/task matrices, acquisition, forgetting, transfer, intransigence, source IDs, byte counts, component arithmetic estimates, wall time, process RSS, and batch-one/batch-128 serial latency are captured. NLL/Brier are separately labeled post-lock read-only analyses.

Renewal at risk 0.20 accepted 48/50 MNIST checkpoints but only 18/50 Fashion checkpoints. Its paired MNIST difference from two-epoch ER-ACE is **+0.051 percentage points, 95% CI [-5.789, +5.891]**; there is no demonstrated advantage. On Fashion the difference is **-44.509 points, CI [-60.730, -28.288]**. These are descriptive paired t intervals, not multiplicity-adjusted superiority claims.

Global prototypes trade higher accuracy for slower inference: approximately **5.00/5.10 ms** serial batch-one latency on MNIST/Fashion, compared with **0.073/0.086 ms** for ER-ACE in this implementation. Prototype memory is approximately 3.846 MiB, replay 5.997 MiB, and renewal 5.975 MiB including fresh evidence.

The source lock is `af4ac3635681ac02faa2b36181830f45654c3f9549601258e2fe4aa879308825`. All **260/260 runs** completed with zero execution failures, no parameter changes, and no relock. A second complete 260-run training reproduction was not performed; the immutable reproduction wrapper and independent saved-prediction reconstruction were checked.

## Additional executed evidence

- **96 locked structural runs:** ten two-class arrivals in a hidden rotated low-dimensional geometry. At noise 0 and 0.01, scalar certification, uncontrolled projection, and ridge all reach 100%. At noise 0.08, they reach 20.389%, 97.536%, and 100%; at 0.30, 13.595%, 5.429%, and 100%. All unfavorable settings remain included.
- **Inherited H-OSP reproduction:** all 66 neural runs, 825 saved arrays, and 5,577 states match. Its 27.7% MSE improvement over routing belongs to its small **support-informed** synthetic experiment.
- **39 grouped core checks:** twelve root checks plus twenty-seven separately derived specialist checks, including exact counterexamples, finite differences, rollback, confidence accounting, and state restoration.
- **Independent structural audit:** 960 checkpoints and 19.2 million regenerated predictions. Three incomplete ridge archives were repaired by exact locked reruns; originals, surviving-array comparisons, and recovery provenance remain preserved.
- **Independent final audit:** 260 runs, 2,860 NPZ files, eighty controller runs, eight hundred checkpoints, 7,136 contract bindings, and 1,561 aggregate/interval values. No final learning-source repair was required.
- **Working examples:** all-layer Adam training and full optimizer resume; three-depth LoRA and optional base-weight training; generic NPZ feature streams, fresh-evidence exhaustion, and source-leakage rejection on generated arrays.

All development negatives and the initial nine storage-accounting integration failures are retained. A certificate preserving incorrect predictions is never reported as a true-label competence certificate.

## Install and quick check

The locked experiments used Python 3.12.14, CPU PyTorch 2.14.1+cpu, NumPy 2.3.5, SciPy 1.17.0, scikit-learn 1.8.0, and threadpoolctl 3.6.0.

```sh
python -m pip install torch==2.14.1 --index-url https://download.pytorch.org/whl/cpu
python -m pip install -r requirements-reproduction.txt
python -m pip install -e . --no-deps
python release_tools/verify_release.py --mode quick --output checks/new_quick_run
python examples/quickstart.py --output checks/new_api_example
```

Use fresh output directories. `REPRODUCE.md` documents dataset staging, exact partition files, full locked reproduction, inherited reproduction, structural tests, figures, and manuscript generation. The fixed benchmark uses Unix `resource` accounting and was executed on Linux. The **publishing script is Windows-friendly**; native Windows benchmark execution is not claimed. WSL or a matching Linux environment is the straightforward reproduction path.

## Complete package layout

The release is divided to keep the main code/manuscript download manageable. **All archives share the same `Eidrion_v1.0.0/` directory prefix; extract them into the same parent directory.** Companion evidence is required for complete raw-state auditing, but not for the quick core/API checks.

The four large companion ZIPs are supplied as numbered binary parts of at most 480 MiB each. Download all parts into one folder. From the extracted main package, run `python release_tools/restore_evidence.py --parts-dir PATH_TO_DOWNLOADS --restore`. The script verifies every part and the concatenated archive against `companion_parts.json`; then extract the four reconstructed ZIPs. It uses only Python's standard library and works on Windows and Linux. Parts are not independently extractable ZIPs.

| Archive group | Contents |
|---|---|
| Main research package | Manuscript, code, proofs, examples, source locks, all JSON summaries/ledgers, plots, audits, reproduction tools, inherited release ZIP |
| Development evidence | All non-final experiment `.pt`/`.npz` captures, including failed candidates, structural states, and recorded archive recovery |
| Final predictions | All 2,860 final NPZ prediction/evidence captures |
| MNIST states | All final MNIST PyTorch pre/post-acquisition states and proposals |
| Fashion-MNIST states | All final Fashion-MNIST PyTorch states and proposals |

`release_manifest.json` gives per-file SHA-256, bytes, and archive group. `release_catalog.json`, supplied beside the archives, gives archive hashes and sizes. Original raw dataset archives are downloaded separately. Small official archive and partition manifests are bundled. Private source manuscripts and retrieval metadata are excluded; their lineage is preserved in the source audit.

The final saved research capture is approximately **10.256 GiB uncompressed**. These audit checkpoints are not all held by any learner. Learner memory is the separately measured per-run persistent state. Reproducibility storage and continual-learning memory are never conflated.

## Unrun larger experiments and limitations

`extensions/` includes a fixed-feature NPZ CLI, reserved-data CIFAR-100/ResNet18 converter, independently reserved Permuted-MNIST converter, and bounded LoRA integration. Their generated-array engineering tests ran. **Actual CIFAR-100, Permuted-MNIST, pretrained extraction, transformer adaptation, and GPU/distributed benchmarks did not run.** Their configurations are not results.

The strongest remaining bottlenecks are a less conservative directional certificate, learning a compact sufficient future-query family, economical valid evidence reuse, and useful certified representation migration. Unknown distribution shift, contradictory overlapping tasks, arbitrary future query complexity, finite bit capacity, numerical error, and calibration exhaustion remain genuine limits. A fixed finite memory cannot perfectly answer unrestricted independent old queries forever.

## Publication

No upload has occurred. No public licence, repository URL, DOI, peer review, or novelty certification has been invented. See `PUBLICATION_NOTES.txt` and the precise status register.

Dry run, with no network or account operations:

```powershell
.\publish_windows.ps1 -RepoId YOUR_ACCOUNT/eidrion
```

Only after publication is explicitly authorized, an assembled and verified full package can be uploaded by adding `-Execute`. New repositories default to private. An existing repository retains its existing visibility. The Python equivalent is `python release_tools/publish_hf.py --repo-id YOUR_ACCOUNT/eidrion`; its default is also a dry run. No upload command was executed in creating this release.

## Scientific assessment

The manuscript addresses all nine requested mathematical topics under restricted assumptions, implements all three scoped mechanisms, and completes all 356 new locked standard/structural runs. Modern-model scalability and performance competitiveness remain unestablished. The bounded novelty audit identifies exact overlaps; it assigns **no probability of novelty**.

Subjective impact assessment: **4/10 demonstrated**, **6/10 potential**, with the latter contingent on better certificates and credible larger-scale advantages. `research_status.json` distinguishes PROVED, NUMERICALLY VERIFIED, EXPERIMENTALLY OBSERVED, CONDITIONAL, UNVERIFIED, and DISPROVED statements.

The strongest answer to the motivating question is: **yes, under a sufficiently compact and valid future-query or functional contract; no, without restrictions on future distinctions, conflicting tasks, memory, and evidence.** This release makes part of that answer constructive and exposes the current algorithm's practical limits.