README.md

---
license: cc-by-4.0
pretty_name: cactus-axons
---

# SubstrateCommons/cactus-axons

Replay packs (`.rpk`) built with `dmipy_sim`: one converged Monte-Carlo walk each, stored so any acquisition can be replayed on it. Load one with `ReplayPack.load("hf://SubstrateCommons/cactus-axons/<path>")`; the manifest (`manifest.json`) holds the sha256 every load is checked against. This file is rendered from the manifest by `dmipy_sim.replay.publish`.

## Substrate

`cactus/cactus_bundle_00000`

* box: 30 × 30 × 30 µm
* boundary: periodic, periodic, periodic

| pool | D (m²/s) | water fraction |
|---|---|---|
| extra | 1.7e-09 | 1 |
| intra | 1.7e-09 | 1 |
| myelin | 0 | 0.4 |

## Packs

| id | duration | walkers | bands | channels | floor | size | commit | path |
|---|---|---|---|---|---|---|---|---|
| `cactus/single-bundle-1s-c3-seg125ms` | 1000 ms (8 × 125 ms, 472.0 MB/window) | 100000 | 320 | positions, boundary_local_time, compartment, susceptibility_grid, susceptibility_path | 0.0266 | 3.8 GB | e8c640e4 | `packs/single_bundle_1s_c3_seg125ms.rpk` |

## Citation

* Villarreal-Haro et al., CACTUS; dmipy-sim#285, #528

## The single-bundle 1 s walk in 125 ms windows (`cactus/single-bundle-1s-c3-seg125ms`)

`packs/single_bundle_1s_c3_seg125ms.rpk`: the CACTUS single-bundle walk of 1 s (100,000 walkers in five shards, 13,017
saves at 76.83 µs, D 1.7e-9 m²/s, C1 + C2 + C3) stored as **8 windows of 1,628 saves (125.0 ms) sharing their boundary
saves** (RPK.md §4.3), **built from the kept raw walks** (`bank._build_segmented`): every window a pack of its own from
the walk's saves of that window, positions at K 320 per window (1280 Hz), the contact channel the window's own
cumulative local time at K 8 per window, the C3 path channel at K 64 per window. A consumer reads only the windows an
acquisition reaches: `ReplayPack.load("hf://SubstrateCommons/cactus-axons/packs/single_bundle_1s_c3_seg125ms.rpk",
windows=range(k))` fetches the header and the byte ranges of windows `0..k-1` (`segment_bytes` in the manifest row says
what one window costs); every class up to 125 ms reads window 0, up to 250 ms two.

**Certificate.** Shard 1 measured every window against its raw walk (`fidelity.segments[i]`: path-tier error
5.6e-3 to 7.9e-3 against floors 1.0e-2 to 2.7e-2, surface error 1e-11, positions over the whole walk), shards 2-5
inherited window by window (dmipy-sim#535); the whole's `err_max` 5.1e-2 is the sum of the eight windows' path errors
(the additive bound of `combine_segment_fidelity`) against `floor_max` 2.7e-2. Measured equality with the one-window
parent `cactus/swoop-1s-c3` on the same waveforms (max |ΔS|, 61 measurements at b 1000 and 3000, the pack's spec at
3 T): PGSE TE 100 ms (one window) bare 3.8e-8, relaxation 9.4e-9, contact 1.0e-4, field 1.0e-4, all tiers 2.3e-5;
PGSE TE 160 ms and PGSTE TE 184 ms (two windows) bare ≤ 3.4e-7, all tiers ≤ 7.0e-5.

**Contact tier.** Each window's contact is its own, so the pack states no contact envelope: at ρ/D 5,882 and 11,765 1/m (ρ 1e-5 and 2e-5 m/s) it agrees with its one-window parent to 1.8e-4 and 3.2e-5, and its attenuation stays in (0, 1]. The spec declares the myelin pool's susceptibility, so the field tier reads the stored field channel.