docs: start running model & dependency size tally
Living ledger of architecture/model/dependency footprint. Headline: CORE is a 0-parameter, 0-weight architecture (substrate is algebra, not weights), enforced by the teacher versor-invariance test. Tracks: learned params (0), optional teacher lanes (4 declared, all inert, no model size chosen), 10 Python runtime deps (datasets dominates), 8 Rust crates, 0 weight files. Includes a Whisper size reference for when/if a lane is wired, and an update-on-PR + changelog discipline. Lives at docs/model_dependency_size_tally.md.
This commit is contained in:
Shay
parent
ae48a3cbf9
commit
c7290f5ab9
1 changed files with 149 additions and 0 deletions
149
docs/model_dependency_size_tally.md
Normal file
149
docs/model_dependency_size_tally.md
Normal file
|
|
@ -0,0 +1,149 @@
|
|||
# Model & Dependency Size Tally
|
||||
|
||||
A running ledger of how big CORE's model/architecture and its dependency
|
||||
footprint get. The headline metric is **learned parameters** — CORE's thesis is a
|
||||
deterministic, algebra-based engine, so this number should stay **0** for the
|
||||
substrate, forever. Everything else here is secondary footprint.
|
||||
|
||||
**Update this doc in the same PR** whenever you: add/remove a runtime dependency,
|
||||
wire a learned model (teacher/shadow lane), change a model size, or add weight
|
||||
files. Append a dated line to the [Changelog](#changelog).
|
||||
|
||||
---
|
||||
|
||||
## Tally summary (headline)
|
||||
|
||||
| Metric | Value | Trend |
|
||||
|---|---|---|
|
||||
| **CORE substrate learned parameters** | **0** | flat (by design — must stay 0) |
|
||||
| **Model weights on disk (committed or required)** | **0 bytes** | flat |
|
||||
| **Learned models wired (runtime)** | **0** | — |
|
||||
| **Learned models declared but inert (optional, gated)** | 4 (whisper, nemo, clap, encodec) | — |
|
||||
| Python runtime dependencies | 10 | see §3 |
|
||||
| Rust crate dependencies | 8 | see §4 |
|
||||
|
||||
> **The number that matters:** CORE is a **0-parameter architecture**. The
|
||||
> reasoning substrate is geometric algebra (Cl(4,1) versors + deterministic
|
||||
> operators), not weights. The versor-invariance test
|
||||
> (`tests/test_audio_teachers.py::test_teacher_hint_does_not_change_versor`)
|
||||
> structurally guarantees that even when a teacher model is wired, it **cannot
|
||||
> fold into the substrate** — so this stays 0 regardless of teachers.
|
||||
|
||||
---
|
||||
|
||||
## 1. CORE architecture — learned parameters
|
||||
|
||||
| Component | Learned params | Weights | Notes |
|
||||
|---|---|---|---|
|
||||
| Algebra / versor engine (`algebra/`, `core-rs/`) | 0 | 0 | pure Cl(4,1) math |
|
||||
| Field / propagation / vault (`field/`, `vault/`) | 0 | 0 | exact CGA recall, no ANN/embeddings |
|
||||
| Language packs (`language_packs/`, `packs/`) | 0 | 0 | curated symbolic data, not trained |
|
||||
| Audio compiler substrate (`sensorium/audio/`) | 0 | 0 | DSP + frozen operator rotor table |
|
||||
| Generation / comprehension (`generate/`) | 0 | 0 | deterministic |
|
||||
| **Total** | **0** | **0** | — |
|
||||
|
||||
**Principle:** capability comes from strengthening the deterministic path, not
|
||||
from adding parameters. If a "learned weights" line ever appears in this table,
|
||||
it is a doctrinal change and belongs in an ADR, not a quiet dependency bump.
|
||||
|
||||
---
|
||||
|
||||
## 2. Optional learned models (teacher / shadow lanes)
|
||||
|
||||
These are **bolt-on evidence lanes**, never the substrate (ADR-0181 §2, eval-plan
|
||||
§4). They are *declared and gated* behind optional extras; **none is currently
|
||||
wired** (`load_teacher(...)` raises `TeacherUnavailable`, no weights downloaded).
|
||||
Track these as a **dependency footprint**, separate from the architecture tally
|
||||
above — they are removable and never part of the engine's reasoning.
|
||||
|
||||
| Lane | Status | Extra | Model selected | Params / disk |
|
||||
|---|---|---|---|---|
|
||||
| Whisper | declared, **inert** | `audio-whisper` | **none chosen** | 0 (not installed) |
|
||||
| NeMo (Parakeet/Canary) | declared, **inert** | `audio-nemo` | none | 0 |
|
||||
| CLAP | declared, **inert** | `audio-clap` | none | 0 |
|
||||
| EnCodec | declared, **inert** | `audio-encodec` | none | 0 |
|
||||
|
||||
### Whisper size reference (for when/if it is wired)
|
||||
|
||||
Only declare a size here once an adapter is actually wired and the size is pinned
|
||||
in the audio pack manifest (with checksum). Pick the smallest size that clears
|
||||
the transcript-evidence bar — teachers produce *labels*, not comprehension.
|
||||
|
||||
| Model | Params | ~Disk | Notes |
|
||||
|---|---|---|---|
|
||||
| `tiny(.en)` | 39M | ~75 MB | fastest, CPU-friendly, weakest |
|
||||
| `base(.en)` | 74M | ~145 MB | likely sweet spot for label-only use |
|
||||
| `small(.en)` | 244M | ~0.5 GB | |
|
||||
| `medium(.en)` | 769M | ~1.5 GB | |
|
||||
| `large-v3` | 1550M | ~3 GB | best, heaviest |
|
||||
|
||||
> When a lane goes live: change its **Status** to "wired", fill **Model
|
||||
> selected** + **Params / disk**, bump the §0 summary counts, and log it in the
|
||||
> Changelog with the pinned size + manifest checksum.
|
||||
|
||||
---
|
||||
|
||||
## 3. Python runtime dependencies
|
||||
|
||||
From `pyproject.toml` `[project].dependencies`. Disk sizes are **approximate
|
||||
installed footprints** (order-of-magnitude); the exact, auditable number is the
|
||||
learned-params tally above. Measure precisely with
|
||||
`du -sh .venv/lib/python*/site-packages/<pkg>` if needed.
|
||||
|
||||
| Package | Pin | Purpose | ~Installed |
|
||||
|---|---|---|---|
|
||||
| `numpy` | >=1.26 | core array math | ~30–40 MB |
|
||||
| `datasets` | >=4.8.5 | GSM8K eval corpus loading (pulls pyarrow/pandas) | **~300–500 MB** (heaviest) |
|
||||
| `ruff` | >=0.15.12 | lint/format (Rust binary) | ~25–40 MB |
|
||||
| `pytest` | >=9.0.3 | test runner | ~5 MB |
|
||||
| `pytest-asyncio` | >=1.3.0 | async tests | <1 MB |
|
||||
| `pytest-xdist` | >=3.6 | parallel tests | <1 MB |
|
||||
| `hypothesis` | >=6.152.7 | property tests | ~5 MB |
|
||||
| `psutil` | >=7.0 | process/resource probing | ~2 MB |
|
||||
| `pyrage` | ==1.2.3 | age encryption (Rust binding) | ~3 MB |
|
||||
| `pyyaml` | >=6.0 | config/manifest parsing | ~1 MB |
|
||||
|
||||
> **Footprint note:** `datasets` dominates the Python install. If the GSM8K
|
||||
> loading path is ever isolated behind an extra, the base runtime footprint drops
|
||||
> sharply. None of these carry learned weights.
|
||||
|
||||
---
|
||||
|
||||
## 4. Rust crate dependencies
|
||||
|
||||
From `core-rs/Cargo.toml`. Compiled into the `core_rs` extension; contribute to
|
||||
binary size, not learned params.
|
||||
|
||||
| Crate | Version | Purpose |
|
||||
|---|---|---|
|
||||
| `pyo3` | 0.21 | Python↔Rust bindings |
|
||||
| `numpy` | 0.21 | zero-copy numpy views |
|
||||
| `rayon` | 1.10 | data-parallel scoring |
|
||||
| `nalgebra` | 0.33 | linear algebra |
|
||||
| `ndarray` | 0.16 | n-d arrays (diffusion) |
|
||||
| `ndarray-rand` | 0.15 | rng for arrays |
|
||||
| `bytemuck` | 1.16 | zero-copy slice casts |
|
||||
| `thiserror` | 1.0 | error types |
|
||||
|
||||
---
|
||||
|
||||
## 5. Model weights on disk
|
||||
|
||||
| Path | Type | Size |
|
||||
|---|---|---|
|
||||
| — | — | **none** |
|
||||
|
||||
No `.pt` / `.safetensors` / `.onnx` / `.gguf` / `.ckpt` / `.bin` weight files are
|
||||
committed or required. (Verify: `git ls-tree -r --name-only HEAD | grep -iE
|
||||
'\.(pt|onnx|bin|safetensors|ckpt|gguf|pth)$'` → empty.)
|
||||
|
||||
---
|
||||
|
||||
## Changelog
|
||||
|
||||
Append one dated line per change. Newest first.
|
||||
|
||||
- **2026-05-29** — Doc created. Tally baseline: **0 learned params, 0 model
|
||||
weights**. 4 teacher lanes (whisper/nemo/clap/encodec) declared but inert
|
||||
(ADR-0181 PR-6, merged #479); no model size chosen for any. 10 Python runtime
|
||||
deps, 8 Rust crates.
|
||||
Loading…
Reference in a new issue