# Phase 181 remaining-capability audit

Audit date: 2026-07-15 (UTC)

This audit separates an unavailable production capability from a locally
actionable implementation gap. It is evidence for why the six unchecked items in
`phase-181.md` remain unchecked; it is not a waiver or a substitute for those
items.

## Host and repository evidence

- Host: 15 GiB RAM, 31 GiB swap, and 23 GiB free workspace disk at audit time.
- No `nvidia-smi`, `nvcc`, `/dev/nvidia*`, CUDA/NVIDIA scheduler variables, or
  SLURM/PBS/LSF/Ray/Kubernetes execution endpoint is present.
- No POYO, NDT2/NEDS, LaBraM, MindEye2, BrainLM, MICrONS-FNN, FFN, ONNX,
  safetensors, PyTorch-checkpoint, or model-zoo artifact is present in the
  repository. The only Phoebe implementations for these six items are
  deterministic tokenization/preprocessing and explicit fail-loud deployment
  seams.
- The locally actionable gaps found during this audit were completed first: real
  GPFA/LFADS/AutoLFADS/pi-VAE/CEBRA fits (`181.10.2.1`), real
  SpikeInterface/MountainSort5, drift correction, QC, and edge sorting
  (`181.10.1.1`), and executable/durable tiled FFN backend orchestration.

## Requirement-by-requirement disposition

| Task                                             | Evidence inspected                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                 | Why the remaining requirement is not local                                                                                                                                                                                                                                               |
| ------------------------------------------------ | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| `181.5.2.1` POYO/POYO+ training and serving      | Phoebe has spike tokenization but no weights/corpus; the official [POYO repository](https://github.com/nerdslab/poyo) does not place a trusted checkpoint in this workspace.                                                                                                                                                                                                                                                                                                                                       | Training and cross-session stitching require a multi-session neural corpus and accelerator execution. Neither a corpus, an official checkpoint, nor a GPU/HPC endpoint is available locally. A synthetic miniature would not prove foundation-model training or cross-session stitching. |
| `181.5.2.2` NDT2/NEDS and LaBraM                 | The [NDT2 codebase](https://github.com/joel99/context_general_bci) trains against multi-context BCI data. The official [LaBraM implementation](https://github.com/935963004/LaBraM) specifies roughly 2,500 hours of EEG and recommends eight RTX 3090-class GPUs for tokenizer/pretraining.                                                                                                                                                                                                                       | No required neural corpora or accelerator exist locally. Tokenization alone already exists; a small synthetic MAE would be a narrower substitute and would not satisfy this task.                                                                                                        |
| `181.5.2.3` MindEye2-style fMRI/EEG-to-image     | The official [MindEyeV2 repository](https://github.com/MedARC-AI/MindEyeV2) documents about one GPU-day each for full pretraining and fine-tuning over 40 NSD sessions. Phoebe has ROI preprocessing but no NSD data, CLIP/diffusion weights, or trained subject mapper.                                                                                                                                                                                                                                           | The data and GPU execution needed to train and validate the shared-subject latent → CLIP → diffusion path are absent. Emitting pixels from an untrained CPU toy would violate the research-capability requirement.                                                                       |
| `181.7.1.1` learned C++/CUDA FFN                 | There is no CUDA compiler/device or learned FFN checkpoint. Phoebe now executes injected backends tile-by-tile, validates input/output voxel counts, persists before checkpointing, resumes, and retains a safe replay point on failure.                                                                                                                                                                                                                                                                           | Only the learned C++/CUDA segmentation kernel and petavoxel validation remain. They cannot be compiled or exercised without the CUDA toolchain, GPU, weights, and EM corpus.                                                                                                             |
| `181.9.1.1` BrainLM-style training and model zoo | The official [BrainLM repository](https://github.com/vandijklab/BrainLM) states that UK Biobank pretraining data cannot be publicly shared and exposes 111M/650M pretrained models for inference. Its model card describes pretraining on 6,700 hours of UK Biobank/HCP fMRI. No approved data, weights, or hosted artifact store is configured here.                                                                                                                                                              | Downloading one inference checkpoint would not implement or validate the required training plus multimodal hosted model zoo. The governed training corpus and accelerator/model-hosting endpoints are missing.                                                                           |
| `181.9.2.1` function→structure digital twin      | The official [MICrONS FNN release](https://github.com/cajal/fnn) supports CPU inference, but it predicts neural responses, not by itself the required cell-type/connectivity outputs. The public example training prefix contains 26,294 objects totaling 89.36 GiB. The published computed-response/stimulus arrays total about 31.68 GiB; the small anatomy table contains position/area, not cell type or connectivity. MICrONS itself spans about 75,000 functionally imaged neurons and 0.5 billion synapses. | The exact function→cell-type/connectivity fit cannot be trained or validated with the 23 GiB available disk and absent aligned structural labels/model assets. Integrating response-only inference would not satisfy the stated function→structure requirement.                          |

## Re-entry conditions

Resume these items when the corresponding real assets are supplied: a CUDA GPU
or approved remote scheduler, governed training-data access, trusted checkpoints
with their licenses/checksums, and enough storage for the exact evaluation
corpora. Re-run the host/resource checks before broad training or evaluation.
Keep each task unchecked until its named model, dataset-scale evaluation, and
serving/training acceptance evidence all exist.
