coulomb/freedom-intelligence
4
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Register reserve under agents hub and pin VAULT storage.

Browse source 
Classify freedom-intelligence for State Hub (agents domain), sync FI-WP-0001..0003 with hub IDs, enrich catalog entries with profile/SWOT, and pin the open-weight reserve to D:\vault\coulomb\freedom-intelligence\ with strategic S-tier policy.
This commit is contained in:
tegwick 2026-07-28 00:25:21 +02:00
parent 00f469662b
commit a83ef0a79a
29 changed files with 1343 additions and 279 deletions
Download patch file Download diff file Expand all files Collapse all files

32
.repo-classification.yaml Normal file
View file

@ -0,0 +1,32 @@
# Repo classification (Repo Classification Standard v1.0).
repo_classification:
standard: Repo Classification Standard
version: "1.0"
classified_at: "2026-07-28"
classified_by: human
category: project
domain: agents
secondary_domains:
- infotech
capability_tags:
- knowledge
- documentation
- orchestration
- automation
- model-routing
business_stake:
- technology
- intelligence
- automation
- product
business_mechanics:
- intention
- coordination
- operation
- adaptation
notes: >
Daily AI field research briefs and open-weight model reserve for
NetKingdom/Coulomb lab ops. Primary domain agents (model intelligence,
harness awareness, local weight reserve); infotech secondary for platform
adjacency.

12
INTENT.md
View file

@ -28,7 +28,12 @@ Coulomb NetKingdom needs a third posture:
2. **Reserve** open-weight artifacts that may matter later for local training or operations.
3. **Bias** toward intelligence that can be inspected, specialized, and run under our control.
Freedom Intelligence is that lab — not a chat product, not a public model CDN, and not a commitment to host every frontier model.
Freedom Intelligence is that lab — not a chat product, not a public model CDN, and not a commitment to *run* every frontier model on today’s GPUs.
The open-weight **reserve** (VAULT bulk store under `D:\vault\coulomb\`) deliberately includes the **most
capable open models** we are allowed to retain, even when they are larger than
current lab inference can serve. Runnability and reservability are separate
axes: we sense the frontier, keep optionality offline, and grow runtime later.
---
@ -71,8 +76,9 @@ The brief feeds collection candidates. Collection never mirrors the entire Hub
## Design principles
* **Signal over noise** — report what changes capability, cost, access, or operability.
* **Just in case** — hold optionality without obligation to serve every model.
* **Catalog in git, blobs on backup storage** — metadata is versioned; weights are bulk media.
* **Just in case** — hold optionality without obligation to *serve* every model today.
* **Capability-first reserve** — prefer the strongest open weights (license-clean), not only what fits current VRAM.
* **Catalog in git, blobs on bulk store** — metadata is versioned; weights live on `D:\vault\coulomb\freedom-intelligence\`.
* **License and integrity first** — no collection that terms forbid; checksums and provenance required.
* **Homelab honesty** — prefer methods and models that fit our real hardware envelope.
* **Operable intelligence** — models without harness, sandbox, and fleet thinking are incomplete for NetKingdom.

15
README.md
View file

@ -37,12 +37,19 @@ Standing lenses: **A** frontier & price · **B** edge/local open · **C** homela
---
## Collection spine (from baseline)
## Collection plan (bulk reserve)
**P0 candidates** (approve after backup path pin): Qwen3-8B, Llama-3.2-3B-Instruct,
BGE-M3, DeepSeek-R1-Distill-Qwen-14B, nomic-embed-text-v1.5
**Storage:** `D:\vault\coulomb\freedom-intelligence\` (VAULT HD) · soft quota **850 GiB**
**Policy:** store the **most capable open weights** even if unrunnable today
(`docs/decisions/2026-07-24-nas-strategic-reserve.md`).
Details and P1/P2: see baseline survey §6 and `inventory/catalog/`.
| Tier | Examples |
| ---- | -------- |
| **R — runnable spine** | Qwen3-8B, Llama-3.2-3B-Instruct, BGE-M3, R1-Distill-14B, nomic-embed |
| **S — strategic** | DeepSeek-V3 (class), DeepSeek-R1 full, Llama-3.3-70B / Qwen3-72B dense |
| **W — optional** | 14B/32B upgrades, reranker, … |
Plan detail: `research/2026-07-24-nas-strategic-collection-plan.md` · catalog: `inventory/catalog/`.
---

9
SCOPE.md
View file

@ -95,11 +95,12 @@ Git holds briefs, schemas, policies, and catalog metadata. Model weight blobs st
| INTENT / SCOPE | Drafted |
| Baseline survey | `research/2026-07-24-baseline-field-survey.md` |
| Brief template + playbook | Present; first *daily* delta brief not yet written |
| Backup storage policy | Documented; **path/quota TBD** (FI-WP-0001-T04) |
| Inventory | Schema + policy + **P0/P1 catalog candidates** seeded |
| Backup storage | **VAULT HD** `D:\vault\coulomb\freedom-intelligence\`; soft quota 850 GiB |
| Inventory | R/S/W tiers; runnable + **strategic unrunnable** candidates seeded |
| Reserve policy | Capability-first — not gated by current VRAM |
| Activity-core | Definition drafted `enabled: false` (FI-WP-0002) |
| State Hub | `fi_daily_brief` contract documented; resolver not implemented |
| Workplans | FI-WP-0001 … 0003 active |
| State Hub | Repo registered under **agents** (2026-07-28); workplans synced via fix-consistency; `fi_daily_brief` delivery resolver still open |
| Workplans | FI-WP-0001 finished; FI-WP-0002/0003 active (hub-indexed) |
---

29
WORK-RECORDS.md Normal file
View file

@ -0,0 +1,29 @@
# Work Records — freedom-intelligence
> Generated by `statehub fix-consistency` (CUST-WP-0061-T04, work-record
> stage 3). Do not edit by hand — edit the source file/block listed for
> each record and re-run fix-consistency to refresh this index. Archived
> workplans are omitted; closed decisions/intakes/engagements stay listed
> so recently-resolved work is still visible. [auto]
| Kind | ID | Status | Lane | Source |
| --- | --- | --- | --- | --- |
| workplan | FI-WP-0001 | done | — | workplans/FI-WP-0001-lab-operating-foundation.md |
| workplan | FI-WP-0002 | active | — | workplans/FI-WP-0002-activity-core-daily-research.md |
| workplan | FI-WP-0003 | active | — | workplans/FI-WP-0003-seed-open-weight-reserve.md |
| task | FI-WP-0001-T01 | done | — | workplans/FI-WP-0001-lab-operating-foundation.md |
| task | FI-WP-0001-T02 | done | — | workplans/FI-WP-0001-lab-operating-foundation.md |
| task | FI-WP-0001-T03 | done | — | workplans/FI-WP-0001-lab-operating-foundation.md |
| task | FI-WP-0001-T04 | done | — | workplans/FI-WP-0001-lab-operating-foundation.md |
| task | FI-WP-0001-T05 | done | — | workplans/FI-WP-0001-lab-operating-foundation.md |
| task | FI-WP-0002-T01 | done | — | workplans/FI-WP-0002-activity-core-daily-research.md |
| task | FI-WP-0002-T02 | done | — | workplans/FI-WP-0002-activity-core-daily-research.md |
| task | FI-WP-0002-T03 | done | — | workplans/FI-WP-0002-activity-core-daily-research.md |
| task | FI-WP-0002-T04 | todo | — | workplans/FI-WP-0002-activity-core-daily-research.md |
| task | FI-WP-0002-T05 | todo | — | workplans/FI-WP-0002-activity-core-daily-research.md |
| task | FI-WP-0002-T06 | done | — | workplans/FI-WP-0002-activity-core-daily-research.md |
| task | FI-WP-0003-T01 | done | — | workplans/FI-WP-0003-seed-open-weight-reserve.md |
| task | FI-WP-0003-T02 | todo | — | workplans/FI-WP-0003-seed-open-weight-reserve.md |
| task | FI-WP-0003-T03 | todo | — | workplans/FI-WP-0003-seed-open-weight-reserve.md |
| task | FI-WP-0003-T04 | todo | — | workplans/FI-WP-0003-seed-open-weight-reserve.md |
| task | FI-WP-0003-T05 | todo | — | workplans/FI-WP-0003-seed-open-weight-reserve.md |

119
docs/backup-storage-policy.md
View file

@ -1,7 +1,8 @@
# Backup storage policy — open-weight model reserve
**Status:** foundation draft (concrete path and budget **TBD**)
**Related:** `INTENT.md`, `SCOPE.md`, `inventory/collection-policy.md`
**Status:** facility + path pinned (workstation VAULT HD)
**Related:** `INTENT.md`, `SCOPE.md`, `inventory/collection-policy.md`,
`docs/decisions/2026-07-24-nas-strategic-reserve.md`
**Adjacent:** `disaster-control` BackupPolicy (platform backups — different concern)
---
@ -25,35 +26,42 @@ Git never stores weight tensors. Git stores:
| Requirement | Policy |
| ----------- | ------ |
| **Class** | Backup / bulk durable storage — **not** hot cluster PVCs or app disks |
| **Durability** | Same or better retention posture as other lab bulk assets; prefer media that survives workstation rebuilds |
| **Facility** | Workstation VAULT HD (`D:`) under `D:\vault\coulomb\` (interim pin 2026-07-28) |
| **Durability** | Survives WSL rebuilds; weights live on the VAULT volume, not the git/WSL root |
| **Performance** | Sequential read for restore/training pull is enough; low latency not required |
| **Access** | Operator and approved lab hosts only; not a public mirror |
| **Separation** | Do not co-mingle with age-encrypted operational backups (Forgejo dumps, k3s state) without a clear subdirectory and different lifecycle rules |
Platform backup paths such as `/opt/backup/railiance/{infra,cluster}/` and
`~/.cache/railiance/backups/` are **operational recovery** lanes. The model
reserve may live on the **same physical facility** only if isolated by path and
quota so large weight pulls cannot crowd out restore media.
`~/.cache/railiance/backups/` are **operational recovery** lanes. Model weights
belong on the **VAULT model tree**, not on railiance hot disks.
---
## Target location (to pin)
## Target location
| Field | Value |
| ----- | ----- |
| **Host / facility** | `TBD` — operator pin (e.g. backup NAS, off-cluster bulk volume, dedicated disk) |
| **Base path or bucket** | `TBD` — suggested shape: `…/freedom-intelligence/models/` |
| **Host / facility** | Workstation-attached bulk HD **VAULT** (Windows `D:`) |
| **Raw capacity** | ~1.86 TB total (~1.76 TB free as of 2026-07-28 pin) |
| **Base path (Windows)** | `D:\vault\coulomb\freedom-intelligence\` |
| **Base path (WSL, when D: automounted)** | `/mnt/d/vault/coulomb/freedom-intelligence/` |
| **Layout under base** | See [On-disk layout](#on-disk-layout) |
| **Mount on lab hosts** | `TBD` |
| **Credentials** | `TBD` — if remote: OpenBao or existing backup credential lane; never commit secrets |
| **Mount on lab hosts** | Local attach on operator workstation; WSL via drvfs `D:` → `/mnt/d` when enabled |
| **Credentials** | Local filesystem ACL on the workstation HD; never commit secrets |
Until pinned, **do not** bulk-download multi-GB models into this git workspace or
into hot root filesystems.
**Interim note:** Earlier policy assumed a dedicated ~1 TB NAS. Operator chose this
VAULT volume under `D:\vault\coulomb\` for now. Soft quota **850 GiB** for the
model tree still applies (self-imposed discipline; disk is larger).
### Suggested path shape (non-binding)
Do **not** bulk-download multi-GB models into this git workspace or into hot
root filesystems (`/` on WSL, `C:`).
### Path shape (pinned)
```text
{BACKUP_ROOT}/freedom-intelligence/
D:\vault\coulomb\freedom-intelligence\ # Windows
/mnt/d/vault/coulomb/freedom-intelligence/ # WSL when D: mounted
├── models/
│ └── {org}__{name}/
│ └── {revision}/
@ -65,6 +73,8 @@ into hot root filesystems.
Map each `{org}__{name}/{revision}` to an inventory catalog entry.
Directories `models/`, `staging/`, and `companions/` were created 2026-07-28.
---
## On-disk layout
@ -76,17 +86,33 @@ Map each `{org}__{name}/{revision}` to an inventory catalog entry.
| `blobs/` | Actual files; prefer original names from source |
| `staging/` | Incomplete transfers; purge or resume; never mark collected until complete + verified |
Optional layout tags (directory name or catalog field):
| Tag | Meaning |
| --- | ------- |
| runnable | Fits current run envelope (R tier) |
| strategic | Capability reserve; may exceed current hardware (S tier) |
---
## Capacity budget
| Parameter | Policy |
| --------- | ------ |
| **Soft quota** | `TBD` GiB/TiB — operator pin based on free backup capacity |
| **Hard stop** | No new collection when soft quota exceeded unless operator raises budget |
| **Per-pull threshold** | See `inventory/collection-policy.md` (size gates approval) |
| **Growth review** | Revisit quota when catalog total exceeds 70% of soft quota |
| **Eviction** | Prefer archive/delete lowest-priority, easily re-obtainable revisions first; record eviction in catalog history |
| **Device** | VAULT HD (`D:`) — model tree under `D:\vault\coulomb\freedom-intelligence\` |
| **Soft quota (model reserve)** | **850 GiB** (self-imposed; disk is larger) |
| **Hard stop** | **920 GiB** used under the freedom-intelligence base path |
| **Growth review** | When catalog total ≥ **70% soft quota (~595 GiB)** |
| **Per-pull threshold** | See `inventory/collection-policy.md` |
| **Eviction** | Prefer drop **W** (watch) and easily re-obtained quants; protect unique **S** SOTA bases and the **R** spine |
### Portfolio guidance on 1 TB
* Always keep the **runnable spine (R / former P0)** resident.
* Use remaining space for a **small number of most-capable open models (S)**,
preferably well-provenance compressed weights when full precision would exhaust the disk.
* Do **not** attempt to mirror entire HF orgs.
* Full bf16 of multiple 600B-class models will **not** fit; prioritize top capability identities.
---
@ -100,8 +126,7 @@ For every completed collection:
4. Prefer official org releases over anonymous re-uploads.
5. Keep license text or SPDX id in catalog; refuse unclear licenses.
Verification command examples belong in tooling later; policy only requires that
**catalog claims match on-disk checksums** before status `collected`.
Verification: **catalog claims must match on-disk checksums** before status `collected`.
---
@ -109,10 +134,11 @@ Verification command examples belong in tooling later; policy only requires that
| Class | Retention |
| ----- | --------- |
| **Strategic reserve** (high priority, hard to re-obtain) | Keep until explicit deprecation |
| **Working set** (common bases for training experiments) | Keep while in active use + one superseded revision optional |
| **Strategic capability (S)** | Keep until explicit deprecation or displacement by a clearly stronger open successor |
| **Runnable spine (R)** | Keep while still the lab default for local ops / FT |
| **Working set** | Active experiment bases + optional one superseded revision |
| **Staging** | Max 14 days incomplete, then purge |
| **Deprecated** | Metadata retained in catalog with status `evicted` or `superseded`; blobs may be deleted |
| **Deprecated / evicted** | Metadata retained in catalog; blobs may be deleted |
---
@ -120,42 +146,35 @@ Verification command examples belong in tooling later; policy only requires that
| Topic | Policy |
| ----- | ------ |
| **At rest** | Follow host/facility default; extra age/GPG of multi-hundred-GB trees is optional and costly |
| **In transit** | HTTPS or trusted lab network only |
| **Offsite copy** | Optional later; not required for foundation. If added, coordinate with `disaster-control` so model reserve does not break operational backup SLAs |
| **At rest** | NAS default; extra age/GPG of multi-hundred-GB trees optional |
| **In transit** | Trusted lab network / local attach |
| **Offsite copy** | Optional later; coordinate with disaster-control if added so model bulk does not break ops backup SLAs |
---
## What must not live here
* Closed weights or artifacts whose terms forbid offline retention
* Secrets, API keys, customer data, or training corpora with personal data (domain datasets need their own policy)
* Operational backups (databases, k3s state, Forgejo dumps)
* Git LFS dumps of full model trees as a substitute for backup storage
* Closed weights or artifacts whose terms forbid offline retention
* Secrets, API keys, customer data, or ungoverned training corpora with personal data
* Operational backups (databases, k3s state, Forgejo dumps) as the primary home
* Git LFS dumps of full model trees as a substitute for the NAS
---
## Operator checklist to pin TBD fields
## Operator checklist
- [ ] Choose facility and base path
- [ ] Confirm free capacity and set soft quota
- [ ] Document mount/access for lab hosts
- [ ] Confirm credential path (if any)
- [ ] Create empty `models/` and `staging/` directories
- [ ] Update this file: replace `TBD` rows with concrete values
- [ ] Note pin date and operator in revision history below
- [x] Choose facility: bulk durable storage for model reserve (2026-07-24 intent)
- [x] Set soft quota: **850 GiB** / hard stop **920 GiB**
- [x] Pin path: **`D:\vault\coulomb\freedom-intelligence\`** (2026-07-28)
- [x] Create `models/`, `staging/`, `companions/`
- [x] Document Windows + WSL path forms above
- [x] Ensure WSL can see `D:` (`/mnt/d`) when downloads run from Linux (verified 2026-07-28)
### Pin log
| Date | Operator | Change |
| ---- | -------- | ------ |
| 2026-07-23 | foundation | Policy created; path and quota unpinned |
---
## Relationship to disaster-control
`disaster-control` owns **platform resilience** (what to restore after loss).
Freedom Intelligence owns **which open weights we choose to retain** and their
inventory. Shared facilities are fine; shared lifecycle rules are not automatic —
model reserve is large, slow-changing, and rarely needed for emergency restore.
| 2026-07-24 | foundation | Policy created; path unpinned |
| 2026-07-24 | operator direction | Facility intent = 1 TB local NAS; soft quota 850 GiB; strategic unrunnable models in scope |
| 2026-07-28 | operator direction | **Interim pin:** workstation VAULT HD `D:\vault\coulomb\freedom-intelligence\`; layout created; soft quota retained |
| 2026-07-28 | operator | WSL mount verified: `D:` → `/mnt/d` (1.9T, writable) |

84
docs/decisions/2026-07-24-nas-strategic-reserve.md Normal file
View file

@ -0,0 +1,84 @@
# Decision: 1TB local NAS + strategic capability reserve
**Date:** 2026-07-24
**Status:** accepted
**Affects:** `docs/backup-storage-policy.md`, `inventory/collection-policy.md`,
`docs/hardware-envelope.md`, collection tiers, FI-WP-0001 / 0003
---
## Decision
1. **Storage backend** for the open-weight reserve is a **1 TB minimal local NAS**
attached to the lab (operator facility). Weights live there; git keeps catalog only.
2. **Reserve intent** is explicitly **not limited to currently runnable models**.
We capture the **most capable open-weight models** that license allows us to
retain, even when they exceed today’s inference/training hardware (T4 and beyond).
3. **Runnability** and **reservability** are separate axes:
- *Run envelope* → what we can serve or fine-tune now (T0–T3)
- *Reserve envelope* → what we store for future ops, training facility, or access loss
---
## Capacity reality check (1 TB)
| Use | Guidance |
| --- | -------- |
| Raw device | ~1 TB |
| Soft quota for model reserve | **850 GiB** (leave ~15% for FS overhead, staging, other lab bulk) |
| Hard stop | **920 GiB** used on the NAS model tree — stop new pulls |
| Growth review | At **70% of soft quota (~595 GiB)** re-prioritize; prefer drop easily re-obtained quants over unique SOTA bases |
Full-precision 600B-class MoE weights can be **multi-hundreds of GiB to >1 TB each**.
On a 1 TB NAS we therefore:
* Prefer **official or well-provenance compressed** distributions (FP8, INT4/GGUF,
published quant) when the card supports them for the *strategic* tier
* Still catalog the **base model id + revision** as the capability identity
* Accept that only a **small number of top-tier giants** fit at once — quality over quantity
* Keep the **runnable spine** (P0) always resident so local work does not depend on giants
Rough fit examples (order-of-magnitude; re-measure at download):
| Portfolio sketch | Est. total |
| ---------------- | ---------- |
| P0 quant spine only | ~25–55 GiB |
| P0 + mid P1 | ~80–150 GiB |
| P0 + **one** SOTA open MoE quant (~300–400 GiB) | ~350–450 GiB |
| P0 + two large MoE quants | can approach or exceed soft quota — choose carefully |
---
## Collection tiering (updated)
| Tier | Name | Rule |
| ---- | ---- | ---- |
| **R** | Runnable spine | Fits current/near hardware; always keep |
| **S** | Strategic capability | Best open SOTA / near-SOTA; **may be unrunnable today**; high retention |
| **W** | Watch / optional | Nice-to-have; only if soft quota headroom |
Supersedes the earlier “do not collect full DeepSeek MoE without a separate
decision” posture: **strategic collection of top open giants is in scope** for
this NAS, subject to license + soft quota.
---
## Non-goals (unchanged)
* Public model CDN or Hugging Face mirror
* Collecting closed weights
* Guaranteeing inference of every reserved model on current hosts
---
## Follow-ups
- [x] Operator: pin storage path and create `freedom-intelligence/{models,staging,companions}/`
- **2026-07-28 interim:** `D:\vault\coulomb\freedom-intelligence\` on workstation VAULT HD (~1.86 TB)
- Soft quota **850 GiB** retained (self-imposed; disk larger than original 1 TB sketch)
- [x] Update pin log in `docs/backup-storage-policy.md` with exact mount path
- [x] Seed/update catalog entries for strategic tier (S)
- [ ] FI-WP-0003: prioritize R spine + top S giants under 850 GiB soft quota
- [x] WSL: `D:` available as `/mnt/d` (verified 2026-07-28)

57
docs/hardware-envelope.md
View file

@ -1,28 +1,47 @@
# Hardware envelope — homelab accessible (draft)
# Hardware envelope — run vs reserve
**Status:** assumptions + TBD measurements
**Used by:** inventory `hardware_class`, axis B/C prioritization, collection size decisions
**Status:** updated 2026-07-24 (1 TB NAS strategic reserve)
**Used by:** inventory `hardware_class`, axis B/C, collection tiers R/S
---
## Purpose
Define what Freedom Intelligence means by **homelab accessible** so we do not
collect models that only make sense on hyperscale clusters.
Separate two questions that used to be conflated:
1. **What can we run or fine-tune now?** → *run envelope*
2. **What should we store on the NAS just in case?** → *reserve envelope* (1 TB)
Collecting a model **does not** require it to fit the run envelope.
See `docs/decisions/2026-07-24-nas-strategic-reserve.md`.
---
## Working tiers (until hosts are measured)
## Run envelope (inference / training *today*)
| Tier | Assumed capacity | Target model class |
| ---- | ---------------- | ------------------ |
| Tier | Assumed capacity | Typical runnable class |
| ---- | ---------------- | ---------------------- |
| **T0 — CPU / edge** | 16–64 GB system RAM, no GPU | ≤3B Q4; embeddings ≤0.5B |
| **T1 — consumer GPU** | 8–12 GB VRAM | 7–8B Q4/Q5; 3B fp16 |
| **T2 — enthusiast GPU** | 16–24 GB VRAM | 14B Q4; 8B fp16; light 32B Q4 |
| **T3 — lab multi-GPU** | 2×24 GB+ or 48 GB+ | 32B fp16 / 70B Q4; small MoE |
| **T4 — out of envelope** | multi-node / 8×A100-class | Full DeepSeek-V3 MoE — P2 watch only |
| **T4 — beyond current lab** | multi-node / datacenter class | Full large MoE / 70B+ fp16 dense |
P0 collection targets **T0–T2**. P2 full MoE is **T4**.
**Runnable spine (tier R)** targets **T0–T2** (stretch T3 when hardware exists).
**Strategic reserve (tier S)** may be **T3–T4 or above** for *execution* — still
valid for NAS collection.
---
## Reserve envelope (NAS)
| Parameter | Value |
| --------- | ----- |
| Facility | VAULT HD `D:\vault\coulomb\freedom-intelligence\` (soft quota 850 GiB) |
| Soft quota | 850 GiB |
| Hard stop | 920 GiB |
| Policy | `docs/backup-storage-policy.md` |
---
@ -30,13 +49,14 @@ P0 collection targets **T0–T2**. P2 full MoE is **T4**.
| Host | Role | GPU | VRAM | RAM | Notes |
| ---- | ---- | --- | ---- | --- | ----- |
| TBD | primary local inference | TBD | TBD | TBD | |
| TBD | training experiments | TBD | TBD | TBD | |
| railiance01 | cluster (not weight store) | n/a | n/a | n/a | Prefer not to fill hot disks with weights |
| TBD | primary local inference | TBD | TBD | TBD | Run envelope |
| TBD | training experiments | TBD | TBD | TBD | Run envelope |
| **VAULT HD (D:)** | weight store | n/a | n/a | n/a | **Reserve envelope** — soft 850 GiB |
| railiance01 | cluster | n/a | n/a | n/a | Do not store weight trees on hot disks |
---
## Inference runtime defaults (intent)
## Inference runtime defaults (when runnable)
| Runtime | When |
| ------- | ---- |
@ -45,16 +65,18 @@ P0 collection targets **T0–T2**. P2 full MoE is **T4**.
| MLX | Apple silicon if present |
| Ollama | Quick operator UX only |
Strategic giants may remain **cold storage only** until hardware upgrades.
---
## Training defaults (intent)
## Training defaults (when runnable)
| Method | Envelope |
| ------ | -------- |
| QLoRA 7–8B | T1–T2 |
| QLoRA 14B | T2 |
| Full FT 7B | T2–T3 |
| 70B+ FT | T3+ only with explicit plan |
| 70B+ / large MoE FT | Future facility — weights may already sit on NAS |
---
@ -62,4 +84,5 @@ P0 collection targets **T0–T2**. P2 full MoE is **T4**.
| Date | Change |
| ---- | ------ |
| 2026-07-24 | Draft tiers; host table empty |
| 2026-07-24 | Draft tiers |
| 2026-07-24 | Split run vs reserve; NAS 1 TB; S tier may exceed run envelope |

12
docs/state-hub-delivery.md
View file

@ -3,6 +3,18 @@
**Status:** contract defined; optional distribution not fully wired
**Related:** `docs/daily-brief-playbook.md`, `workplans/FI-WP-0002-activity-core-daily-research.md`
## Repo registration
| Field | Value |
| ----- | ----- |
| Hub domain | **agents** |
| Repo slug | `freedom-intelligence` |
| Classification | `.repo-classification.yaml` (primary `agents`, secondary `infotech`) |
| Sync | `statehub fix-consistency --repo freedom-intelligence` after workplan edits |
Workplans live in git; the hub indexes them after registration + fix-consistency.
Do not create workplans via hub `create_workplan` APIs.
---
## Required: completion evidence for daily rhythm

42
inventory/README.md
View file

@ -1,32 +1,56 @@
# Model inventory
In-repo **catalog of open-weight revisions** reserved (or nominated) for the lab.
Weight blobs live on backup storage — see `docs/backup-storage-policy.md`.
Weight blobs live on the **VAULT HD** at `D:\vault\coulomb\freedom-intelligence\` —
see `docs/backup-storage-policy.md`.
## Layout
```text
inventory/
├── README.md # this file
├── schema.yaml # field reference + example
├── collection-policy.md # approval and eligibility rules
├── schema.yaml # field reference (v0.2+)
├── collection-policy.md # approval and eligibility (R/S/W tiers)
└── catalog/
└── *.yaml # one file per model revision
```
## What each catalog entry contains
| Section | Purpose |
| ------- | ------- |
| **Identity** | `id`, `name`, `org`, `status`, `tags`, `tier-*` |
| **source** | HF/org URLs, revision pin |
| **profile** | Use cases, sweet spots, anti-patterns, **original_source** link |
| **swot** | Compact strengths / weaknesses / opportunities / threats |
| **license / size / hardware_class** | Retention and run guidance |
| **collection** | Approval, NAS path, brief refs |
Operators should be able to open one YAML and decide *whether* and *why* to pull
or serve a model without reading the whole research tree.
## Status values
`candidate` → `approved` → `collecting` → `collected` → `verified`
also: `superseded` | `evicted` | `rejected`
## Tiers
| Tag | Meaning |
| --- | ------- |
| `tier-r` | Runnable spine (current lab hardware) |
| `tier-s` | Strategic capability (may be beyond run envelope) |
| `tier-w` | Watch / optional |
## Adding an entry
1. Confirm eligibility in `collection-policy.md`.
2. Create `catalog/{org}__{name}__{short_revision}.yaml` using fields from `schema.yaml`.
3. Set `status: candidate` (or `approved` if already signed off).
4. After download and checksums: set `collected` / `verified` and `collection.storage_path`.
2. Create `catalog/{org}__{name}__….yaml` from `schema.yaml` **including `profile` + `swot`**.
3. Set `profile.original_source` to the canonical card/release URL.
4. Set `status: candidate` (or `approved` if signed off).
5. After download: `collected` / `verified` + `collection.storage_path`.
## Empty catalog
## Related
The catalog starts empty on purpose. First entries come from daily brief
**collection candidates** after policy checks — not from bulk scraping.
- Collection plan: `research/2026-07-24-nas-strategic-collection-plan.md`
- Strategic NAS decision: `docs/decisions/2026-07-24-nas-strategic-reserve.md`

47
inventory/catalog/BAAI__bge-m3__candidate.yaml
View file

@ -7,6 +7,7 @@ source:
url: https://huggingface.co/BAAI/bge-m3
revision: main
model_card_url: https://huggingface.co/BAAI/bge-m3
project_url: https://github.com/FlagOpen/FlagEmbedding
license:
spdx: MIT
url: https://huggingface.co/BAAI/bge-m3
@ -14,6 +15,41 @@ license:
allows_local_ops: true
allows_fine_tune: true
notes: ""
profile:
summary: "Default multilingual dense embedding model for local RAG."
original_source: https://huggingface.co/BAAI/bge-m3
use_cases:
- "NetKingdom / Coulomb document and wiki retrieval"
- "Multilingual (DE/EN/…) semantic search"
- "Hybrid retrieval experiments (dense + multi-granularity features)"
- "Offline RAG in sandboxed agents"
- "Indexing ops runbooks and code comments"
sweet_spots:
- "Self-hosted production-quality multilingual embed"
- "MIT license; easy ops story"
- "Strong general retrieval without API spend"
- "Pairs cleanly with a small cross-encoder reranker"
not_ideal_for:
- "Generation / chat (not an LLM)"
- "Pure code retrieval if a code-specialized embed clearly wins A/B"
- "Ultra-tiny edge when EmbeddingGemma-class is enough"
capability_notes: >
BAAI BGE-M3 is a staple open embedder: multi-lingual, multi-granularity,
widely deployed for RAG. Primary R-tier retrieval backbone for the lab.
swot:
strengths:
- "Proven multilingual retrieval quality for self-host"
- "MIT; small enough for CPU/GPU flexibility"
- "Mature FlagEmbedding ecosystem"
weaknesses:
- "Not optimized solely for code or for extreme long-context embed niches"
- "Newer embed lines (Qwen/Gemma/Nomic) may win specific A/B tests"
opportunities:
- "Domain-adapted embed fine-tune on Coulomb corpora later"
- "Companion bge-reranker for precision@k gains"
threats:
- "Leaderboard churn; risk of holding a stale 'default' without re-eval"
- "Pipeline lock-in if vector DBs assume fixed dimension without migration plan"
size:
total_bytes: 0
total_human: "~2 GB"
@ -31,12 +67,17 @@ collection:
storage_path: ""
brief_refs:
- research/2026-07-24-baseline-field-survey.md
reason: "P0 multilingual embedding staple for local RAG (docs, ops notes, DE/EN)."
tags: [embedding, multilingual, rag]
companions: []
reason: "P0/R multilingual embedding staple for local RAG (docs, ops notes, DE/EN)."
tags: [tier-r, embedding, multilingual, rag]
companions:
- BAAI__bge-reranker-v2-m3__candidate
notes: "Small download — within agent auto-collect band after license check once storage pinned."
history:
- at: "2026-07-24"
event: nominated
by: baseline-survey
detail: "P0 recommendation from initial deep research."
- at: "2026-07-24"
event: profile_swot_added
by: grok
detail: "schema 0.2 profile + SWOT."

42
inventory/catalog/BAAI__bge-reranker-v2-m3__candidate.yaml
View file

@ -7,6 +7,7 @@ source:
url: https://huggingface.co/BAAI/bge-reranker-v2-m3
revision: main
model_card_url: https://huggingface.co/BAAI/bge-reranker-v2-m3
project_url: https://github.com/FlagOpen/FlagEmbedding
license:
spdx: Apache-2.0
url: https://huggingface.co/BAAI/bge-reranker-v2-m3
@ -14,6 +15,39 @@ license:
allows_local_ops: true
allows_fine_tune: true
notes: "Confirm SPDX on card."
profile:
summary: "Cross-encoder reranker companion to BGE-M3 for higher-precision RAG."
original_source: https://huggingface.co/BAAI/bge-reranker-v2-m3
use_cases:
- "Second-stage rerank of top-k chunks from BGE-M3 (or hybrid) retrieval"
- "Improving answer grounding for NetKingdom / ops runbook Q&A"
- "Eval A/B of retrieval quality without changing the generator"
- "Low-cost precision boost on multilingual corpora"
sweet_spots:
- "Cheap RAG quality win after a solid bi-encoder"
- "Small VRAM; easy always-on sidecar"
- "Same BAAI/FlagEmbedding family as BGE-M3"
not_ideal_for:
- "First-stage retrieval over full corpora (too slow as sole retriever)"
- "Generation / chat"
- "When latency budget forbids a second pass"
capability_notes: >
Lightweight multilingual cross-encoder from the BGE family. W-tier optional
after R spine; high value per byte when RAG is a primary workload.
swot:
strengths:
- "Material precision lift for little disk/VRAM"
- "Proven pairing with BGE-M3"
- "Apache-friendly open stack"
weaknesses:
- "Cross-encoder cost scales with candidate count"
- "Not a substitute for better chunking or generators"
opportunities:
- "Default two-stage pipeline for lab RAG demos"
- "Tune candidate depth (k) per latency tier"
threats:
- "Newer BGE / open rerankers may supersede this checkpoint"
- "LLM-as-judge rerank experiments may reduce need for a dedicated model"
size:
total_bytes: 0
total_human: "~1–2 GB (estimate)"
@ -31,8 +65,8 @@ collection:
storage_path: ""
brief_refs:
- research/2026-07-24-baseline-field-survey.md
reason: "P1 — cheap RAG quality win as companion to BGE-M3."
tags: [reranker, rag]
reason: "P1/W — cheap RAG quality win as companion to BGE-M3."
tags: [tier-w, reranker, rag]
companions:
- BAAI__bge-m3__candidate
notes: ""
@ -41,3 +75,7 @@ history:
event: nominated
by: baseline-survey
detail: "P1 recommendation from initial deep research."
- at: "2026-07-28"
event: profile_swot_added
by: grok
detail: "schema 0.2 profile + SWOT."

46
inventory/catalog/Qwen__Qwen3-14B__candidate.yaml
View file

@ -7,6 +7,7 @@ source:
url: https://huggingface.co/Qwen/Qwen3-14B
revision: main
model_card_url: https://huggingface.co/Qwen/Qwen3-14B
project_url: https://qwenlm.github.io/
license:
spdx: Apache-2.0
url: https://huggingface.co/Qwen/Qwen3-14B
@ -14,6 +15,40 @@ license:
allows_local_ops: true
allows_fine_tune: true
notes: "Confirm card at download."
profile:
summary: "Mid-size Qwen3 step up from 8B for stronger single-GPU chat/code."
original_source: https://huggingface.co/Qwen/Qwen3-14B
use_cases:
- "Higher-quality local assistant when 8B is the bottleneck"
- "Harder coding and long-context drafting on T2 GPUs"
- "A/B baseline vs R1-Distill-14B (general vs reason-specialist)"
- "Domain FT when 8B capacity is insufficient"
sweet_spots:
- "Quality step within still-single-GPU open dense class"
- "Multilingual instruct continuity with Qwen3-8B"
- "Quota-friendly alternative to jumping to 70B"
not_ideal_for:
- "Always-on default if VRAM is tight (prefer 8B)"
- "Deepest reasoner tasks (prefer R1 distill or full R1 reserve)"
- "Embedding / retrieval"
capability_notes: >
W-tier optional after R spine. Collect when soft-quota headroom and a clear
quality gap vs Qwen3-8B show up in daily work. Prefer Instruct sibling if
separate at pin time.
swot:
strengths:
- "Clear capability bump over 8B without 70B cost"
- "Same Qwen3 stack and tooling as the R default"
- "Apache-friendly licensing typical"
weaknesses:
- "Near-duplicate niche vs strong 8B + selective 14B reason distill"
- "Still not frontier closed quality on hard agentic SWE"
opportunities:
- "Promote to R if lab defaults move off 8B"
- "FT target when domain data needs more capacity"
threats:
- "Disk spent better on S1 MoE or 70B dense under tight quota"
- "Next Qwen mid-size may obsolete this checkpoint quickly"
size:
total_bytes: 0
total_human: "~28 GB fp16 / ~9 GB Q4 (estimate)"
@ -31,12 +66,17 @@ collection:
storage_path: ""
brief_refs:
- research/2026-07-24-baseline-field-survey.md
reason: "P1 — stronger single-GPU chat/code when quota allows after P0."
tags: [instruct, qwen3]
companions: []
reason: "P1/W — stronger single-GPU chat/code when quota allows after R spine."
tags: [tier-w, instruct, qwen3]
companions:
- Qwen__Qwen3-8B__candidate
notes: ""
history:
- at: "2026-07-24"
event: nominated
by: baseline-survey
detail: "P1 recommendation from initial deep research."
- at: "2026-07-28"
event: profile_swot_added
by: grok
detail: "schema 0.2 profile + SWOT."

83
inventory/catalog/Qwen__Qwen3-72B__strategic.yaml Normal file
View file

@ -0,0 +1,83 @@
id: Qwen__Qwen3-72B__strategic
status: candidate
name: Qwen3-72B
org: Qwen
source:
kind: huggingface
url: https://huggingface.co/Qwen/Qwen3-72B
revision: main
model_card_url: https://huggingface.co/Qwen/Qwen3-72B
project_url: https://qwenlm.github.io/
license:
spdx: Apache-2.0
url: https://huggingface.co/Qwen/Qwen3-72B
allows_offline_retention: true
allows_local_ops: true
allows_fine_tune: true
notes: "Confirm card; use Instruct/Coder sibling if that is the capability peak for the line."
profile:
summary: "Strategic large Qwen dense — multilingual/tool open 70B-class reserve."
original_source: https://huggingface.co/Qwen/Qwen3-72B
use_cases:
- "Future multi-GPU dense instruct without MoE serving complexity"
- "Strong multilingual (DE/EN) offline assistant at 70B scale"
- "Domain FT base when 8B/14B capacity is too small"
- "Dense alternative to MoE giants for portable lab stacks"
sweet_spots:
- "S4 dense open with Qwen tooling continuity from the R 8B default"
- "Easier multi-GPU dense serve story than 600B MoE"
- "Apache-friendly licensing typical of Qwen3 line"
not_ideal_for:
- "Current default single-GPU ops"
- "Collecting both this and Llama-70B-class first under tight quota"
- "Embedding / retrieval"
capability_notes: >
Large dense strategic pick. Prefer Instruct/Coder sibling if that is the line
peak at pin time. If soft quota is tight after S1, pick at most one of
Qwen-72B vs Llama-70B-class first.
swot:
strengths:
- "High multilingual dense quality in open 70B class"
- "Same family as R-tier Qwen3-8B — transfer of prompts/tools"
- "Usually simpler ops than giant MoE"
weaknesses:
- "~40–150 GiB depending on quant; still heavy"
- "May lag top MoE on some frontier-open benches"
opportunities:
- "First dense strategic fill after R + optional S1 compressed"
- "FT / LoRA at 70B when facility upgrades"
threats:
- "Llama 70B / next Qwen large may be better use of same bytes"
- "Supersession by Qwen next large release"
size:
total_bytes: 0
total_human: "~145 GB fp16 / ~40 GB Q4 (estimate)"
hardware_class:
min_vram_gb_q4: 40
min_vram_gb_fp16: 145
notes: "T3+ to run; strategic dense multilingual open."
axes: [A, B, C]
priority: high
collection:
approved_by: ""
approved_at: null
downloaded_at: null
downloaded_by: ""
storage_path: ""
brief_refs:
- research/2026-07-24-nas-strategic-collection-plan.md
- docs/decisions/2026-07-24-nas-strategic-reserve.md
reason: "S4 strategic — large Qwen dense open for multilingual/tool capability reserve."
tags: [tier-s, strategic, dense, qwen3]
companions:
- Qwen__Qwen3-8B__candidate
notes: "Pick at most one of Llama-70B-class vs Qwen-72B-class first if quota tight after S1 MoE."
history:
- at: "2026-07-24"
event: nominated
by: operator-policy
detail: "Strategic dense open for NAS."
- at: "2026-07-28"
event: profile_swot_added
by: grok
detail: "schema 0.2 profile + SWOT."

45
inventory/catalog/Qwen__Qwen3-8B__candidate.yaml
View file

@ -7,6 +7,7 @@ source:
url: https://huggingface.co/Qwen/Qwen3-8B
revision: main
model_card_url: https://huggingface.co/Qwen/Qwen3-8B
project_url: https://qwenlm.github.io/
license:
spdx: Apache-2.0
url: https://huggingface.co/Qwen/Qwen3-8B
@ -14,6 +15,42 @@ license:
allows_local_ops: true
allows_fine_tune: true
notes: "Confirm exact card license at download time; Qwen3 line generally Apache-2.0."
profile:
summary: "Default mid-small open instruct for local ops, tools, and domain fine-tunes."
original_source: https://huggingface.co/Qwen/Qwen3-8B
use_cases:
- "Local NetKingdom / Coulomb assistant (chat, docs, runbooks)"
- "Tool-using agent loops on consumer GPU"
- "QLoRA / LoRA domain specialization base"
- "Multilingual (incl. DE/EN) drafting and summarization"
- "Offline code help when 14B+ is too heavy"
sweet_spots:
- "Best balance of quality vs VRAM in the ~8B open class for many 2026 stacks"
- "Instruction + tool-use oriented workflows"
- "Homelab fine-tune target (axis C)"
- "Runnable spine default when one model must wear many hats"
not_ideal_for:
- "Hardest SWE-bench-class multi-file engineering (use larger or closed frontier)"
- "Deep multi-step math/reason vs R1-class distill or full reasoners"
- "Embedding / retrieval (use BGE-M3 or nomic)"
capability_notes: >
Flagship small-mid dense open generalist in the Qwen3 line. Strong multilingual
and instruct behavior for its size; primary R-tier workhorse for the lab. Prefer
Instruct sibling on the card if separate repo exists at pin time.
swot:
strengths:
- "High capability density at 8B; Apache-friendly licensing typical"
- "Good multilingual + tool/instruct posture for local agents"
- "Excellent FT base for domain specialization"
weaknesses:
- "Still far from frontier closed models on hard agentic coding"
- "8B ceiling on long-horizon planning and rare knowledge"
opportunities:
- "Domain LoRAs (security, ops, railiance) on NAS-held base"
- "Pair with BGE-M3 RAG for grounded NetKingdom answers"
threats:
- "Rapid supersession by next Qwen/peer 8–14B release"
- "Quant quality variance across third-party GGUF repacks"
size:
total_bytes: 0
total_human: "~16 GB fp16 / ~5 GB Q4 (estimate)"
@ -31,8 +68,8 @@ collection:
storage_path: ""
brief_refs:
- research/2026-07-24-baseline-field-survey.md
reason: "P0 spine — best default open general/tool model for local ops and QLoRA domain specialization."
tags: [instruct, text, qwen3, ft-base]
reason: "P0/R spine — best default open general/tool model for local ops and QLoRA domain specialization."
tags: [tier-r, instruct, text, qwen3, ft-base]
companions: []
notes: "Prefer Instruct variant on card if separate repo; pin commit SHA at collection."
history:
@ -40,3 +77,7 @@ history:
event: nominated
by: baseline-survey
detail: "P0 recommendation from initial deep research."
- at: "2026-07-24"
event: profile_swot_added
by: grok
detail: "schema 0.2 profile + SWOT."

50
inventory/catalog/deepseek-ai__DeepSeek-R1-Distill-Qwen-14B__candidate.yaml
View file

@ -7,6 +7,7 @@ source:
url: https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-14B
revision: main
model_card_url: https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-14B
project_url: https://github.com/deepseek-ai/DeepSeek-R1
license:
spdx: MIT
url: https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-14B
@ -14,6 +15,42 @@ license:
allows_local_ops: true
allows_fine_tune: true
notes: "R1 distill series MIT — confirm card at pin time."
profile:
summary: "Local reasoning-specialist distill — hard tasks without full R1 MoE."
original_source: https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-14B
use_cases:
- "Offline multi-step reasoning (math, logic, planning sketches)"
- "Harder agent tool-selection and self-checks"
- "Local alternative when DeepSeek API is unavailable"
- "Eval harness baseline for 'reason' quality on lab tasks"
- "QLoRA experiments on reason-style behavior"
sweet_spots:
- "Reasoning quality above plain 14B chat models at similar size"
- "Single enthusiast GPU (Q4) without multi-node MoE"
- "MIT-friendly open reason lineage"
not_ideal_for:
- "Cheapest always-on chat (use 3B/8B)"
- "Full parity with frontier closed reasoners or full R1"
- "Pure embedding"
capability_notes: >
Distilled from DeepSeek-R1 into a Qwen-14B-class dense student. Captures much
of the reasoner value proposition for homelab VRAM budgets. Primary R-tier
reason model; full R1 remains strategic S-tier.
swot:
strengths:
- "Strong open reasoner at single-GPU scale"
- "MIT distill story; clear provenance from R1 line"
- "Complements general 8B chat without replacing it"
weaknesses:
- "Heavier and slower than 8B for routine chat"
- "Distill ≠ full R1; still loses on hardest problems"
- "May overthink simple tasks if not prompted carefully"
opportunities:
- "Route only hard queries here; keep 8B as default"
- "Domain FT for structured ops decision checklists"
threats:
- "Newer distill or mid-size reasoners may obsolete this checkpoint"
- "VRAM growth of defaults may push lab to 32B distill instead"
size:
total_bytes: 0
total_human: "~28 GB fp16 / ~9 GB Q4 (estimate)"
@ -31,12 +68,17 @@ collection:
storage_path: ""
brief_refs:
- research/2026-07-24-baseline-field-survey.md
reason: "P0 local reasoning without full R1 MoE — agent/tool loops and harder offline tasks."
tags: [reasoning, distill, deepseek]
companions: []
notes: "If disk/VRAM constrained, substitute DeepSeek-R1-Distill-Qwen-8B as P0 alternate."
reason: "P0/R local reasoning without full R1 MoE — agent/tool loops and harder offline tasks."
tags: [tier-r, reasoning, distill, deepseek]
companions:
- deepseek-ai__DeepSeek-R1__strategic
notes: "If disk/VRAM constrained, substitute DeepSeek-R1-Distill-Qwen-8B as R alternate."
history:
- at: "2026-07-24"
event: nominated
by: baseline-survey
detail: "P0 recommendation from initial deep research."
- at: "2026-07-24"
event: profile_swot_added
by: grok
detail: "schema 0.2 profile + SWOT."

48
inventory/catalog/deepseek-ai__DeepSeek-R1-Distill-Qwen-32B__candidate.yaml
View file

@ -7,13 +7,47 @@ source:
url: https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-32B
revision: main
model_card_url: https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-32B
project_url: https://github.com/deepseek-ai/DeepSeek-R1
license:
spdx: MIT
url: https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-32B
allows_offline_retention: true
allows_local_ops: true
allows_fine_tune: true
notes: ""
notes: "R1 distill series MIT — confirm card at pin time."
profile:
summary: "Larger R1 distill for stronger local reason when 14B is not enough."
original_source: https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-32B
use_cases:
- "Hard offline reasoning beyond 14B distill quality"
- "Local eval ceiling before considering full R1 MoE"
- "Heavier agent planning loops on multi-GPU / high-VRAM hosts"
- "FT experiments on reason-style traces at 32B"
sweet_spots:
- "Best dense open reason step between 14B distill and full R1"
- "Still denser/portable than 600B-class MoE"
- "MIT-friendly R1 lineage"
not_ideal_for:
- "Default always-on chat (too heavy)"
- "Hosts without ~20+ GB VRAM for Q4"
- "When 14B distill already saturates task quality"
capability_notes: >
W-tier optional. Approve only with VRAM + soft-quota headroom after R spine
and primary S pulls. Do not confuse with full DeepSeek-R1 MoE (separate S entry).
swot:
strengths:
- "Material reason quality jump over 14B distill"
- "Dense, so simpler serving than full MoE"
- "Clear provenance in R1 distill family"
weaknesses:
- "VRAM and latency cost; poor default chat model"
- "Still not full R1; diminishing returns vs S2 full weights"
opportunities:
- "Route only hardest local tasks here"
- "Bridge until multi-GPU can host full R1"
threats:
- "Quota competition with S1/S4 strategic weights"
- "Newer mid-large open reasoners may leapfrog"
size:
total_bytes: 0
total_human: "~65 GB fp16 / ~20 GB Q4 (estimate)"
@ -31,12 +65,18 @@ collection:
storage_path: ""
brief_refs:
- research/2026-07-24-baseline-field-survey.md
reason: "P1 stronger local reasoner — approve only with VRAM + quota headroom."
tags: [reasoning, distill, deepseek]
companions: []
reason: "P1/W stronger local reasoner — approve only with VRAM + quota headroom."
tags: [tier-w, reasoning, distill, deepseek]
companions:
- deepseek-ai__DeepSeek-R1-Distill-Qwen-14B__candidate
- deepseek-ai__DeepSeek-R1__strategic
notes: "Do not collect full DeepSeek-V3/R1 MoE under this id."
history:
- at: "2026-07-24"
event: nominated
by: baseline-survey
detail: "P1 recommendation from initial deep research."
- at: "2026-07-28"
event: profile_swot_added
by: grok
detail: "schema 0.2 profile + SWOT."

84
inventory/catalog/deepseek-ai__DeepSeek-R1__strategic.yaml Normal file
View file

@ -0,0 +1,84 @@
id: deepseek-ai__DeepSeek-R1__strategic
status: candidate
name: DeepSeek-R1
org: deepseek-ai
source:
kind: huggingface
url: https://huggingface.co/deepseek-ai/DeepSeek-R1
revision: main
model_card_url: https://huggingface.co/deepseek-ai/DeepSeek-R1
project_url: https://github.com/deepseek-ai/DeepSeek-R1
license:
spdx: MIT
url: https://huggingface.co/deepseek-ai/DeepSeek-R1
allows_offline_retention: true
allows_local_ops: true
allows_fine_tune: true
notes: "MIT for R1 series per public release notes; re-confirm card at pin."
profile:
summary: "Strategic full open reasoner — best-known open R1-class weights for future multi-GPU."
original_source: https://huggingface.co/deepseek-ai/DeepSeek-R1
use_cases:
- "Offline retention of top open reasoner lineage"
- "Future local hosted reason when multi-GPU is available"
- "Teacher / reference for distill and eval harnesses"
- "Hard planning and math when API reasoners are unavailable"
sweet_spots:
- "S2 pure-reason niche on the reserve"
- "Complements V3 general MoE rather than replacing the 14B distill spine"
- "MIT-friendly open reasoner identity"
not_ideal_for:
- "Current single-GPU daily reason (use R1-Distill-14B)"
- "Full-precision pull if soft quota already holds S1 giant"
- "Casual chat default"
capability_notes: >
Full DeepSeek-R1 (or open successor reasoner). Distills cover runnable reason;
this entry is strategic optionality. If quota forces a choice between full R1
and full V3, re-evaluate from latest brief which niche is the larger gap.
swot:
strengths:
- "Reference open reasoner class for the era"
- "Distill ecosystem already proves value at smaller sizes"
- "Open weights for offline continuity"
weaknesses:
- "Enormous resource cost to serve"
- "Overlaps capability budget with S1 MoE on a small disk"
- "Version supersession risk"
opportunities:
- "Hold compressed if needed; keep identity for later upgrade"
- "Use as teacher target for lab distill experiments"
threats:
- "Quota: competing with S1 for hundreds of GiB"
- "Newer open reasoners may displace this checkpoint"
size:
total_bytes: 0
total_human: "large — similar class to V3 full; prefer compressed if needed for 850 GiB soft quota"
hardware_class:
min_vram_gb_q4: 0
min_vram_gb_fp16: 0
notes: "Beyond current lab run envelope. Distills (14B/32B) cover runnable reason; this is full strategic reasoner."
axes: [A, B, C]
priority: high
collection:
approved_by: ""
approved_at: null
downloaded_at: null
downloaded_by: ""
storage_path: ""
brief_refs:
- research/2026-07-24-nas-strategic-collection-plan.md
- docs/decisions/2026-07-24-nas-strategic-reserve.md
reason: "S2 strategic — best-known open reasoner lineage; optionality for future multi-GPU / hosted local reason."
tags: [tier-s, strategic, beyond-run-envelope, reasoning]
companions:
- deepseek-ai__DeepSeek-R1-Distill-Qwen-14B__candidate
notes: "If quota forces a choice between full R1 and full V3, prefer the stronger current open general (often V3/V4 line) unless pure reason is the gap; re-evaluate at pull time from latest brief."
history:
- at: "2026-07-24"
event: nominated
by: operator-policy
detail: "Strategic NAS reserve — unrunnable OK."
- at: "2026-07-28"
event: profile_swot_added
by: grok
detail: "schema 0.2 profile + SWOT."

83
inventory/catalog/deepseek-ai__DeepSeek-V3__strategic.yaml Normal file
View file

@ -0,0 +1,83 @@
id: deepseek-ai__DeepSeek-V3__strategic
status: candidate
name: DeepSeek-V3
org: deepseek-ai
source:
kind: huggingface
url: https://huggingface.co/deepseek-ai/DeepSeek-V3
revision: main
model_card_url: https://huggingface.co/deepseek-ai/DeepSeek-V3
project_url: https://github.com/deepseek-ai/DeepSeek-V3
license:
spdx: MIT
url: https://huggingface.co/deepseek-ai/DeepSeek-V3
allows_offline_retention: true
allows_local_ops: true
allows_fine_tune: true
notes: "Confirm active card license at pin; prefer latest V3.x/V4 open successor if it is the new SOTA open general."
profile:
summary: "Strategic open general MoE — among the most capable open chat/code weights."
original_source: https://huggingface.co/deepseek-ai/DeepSeek-V3
use_cases:
- "Long-term offline optionality if frontier API access is lost or restricted"
- "Future multi-GPU / training-facility inference or continued pretrain"
- "Capability benchmark reference against closed frontier"
- "Source of student weights / distill experiments (when license + hardware allow)"
sweet_spots:
- "S1 primary giant: best open general MoE class on the reserve"
- "Capability identity even when only compressed weights fit disk"
- "MIT-friendly open SOTA lineage (confirm at pin)"
not_ideal_for:
- "Current lab single-GPU daily ops (use R spine)"
- "Storing multiple near-duplicate full-precision giants on one disk"
- "Assuming immediate local serve without major hardware upgrade"
capability_notes: >
Flagship open general MoE (or successor) for the strategic tier. On ~1–2 TB
lab bulk storage, prefer official compressed distributions when full precision
would exhaust soft quota. Catalog identity stays the base model id + revision.
swot:
strengths:
- "Top-tier open general capability for the era"
- "Open weights enable true offline retention and future fine work"
- "Strong coding/agent relevance as hardware catches up"
weaknesses:
- "Huge disk and multi-GPU serve cost"
- "Rapid version churn (V3 → V3.x → V4)"
- "Quant choice materially affects quality"
opportunities:
- "Hold compressed primary; upgrade hardware later without re-scrape risk"
- "Anchor daily briefs against a fixed offline SOTA open baseline"
threats:
- "Soft quota: one full pull can block other S niches"
- "License or card terms change at successor releases — re-check at pin"
size:
total_bytes: 0
total_human: "hundreds of GiB (quant) to multi-hundred+ GiB; measure at pull — may need official compressed release"
hardware_class:
min_vram_gb_q4: 0
min_vram_gb_fp16: 0
notes: "Beyond current lab run envelope (T4+). Reserved for strategic capability, not current inference."
axes: [A, B, C]
priority: high
collection:
approved_by: ""
approved_at: null
downloaded_at: null
downloaded_by: ""
storage_path: ""
brief_refs:
- research/2026-07-24-nas-strategic-collection-plan.md
- docs/decisions/2026-07-24-nas-strategic-reserve.md
reason: "S1 strategic — among the most capable open general MoE weights; keep even if unrunnable today."
tags: [tier-s, strategic, beyond-run-envelope, moe, general]
companions: []
notes: "At download: pin exact revision; if full precision exceeds remaining soft quota, collect best official compressed distribution of the same model line."
history:
- at: "2026-07-24"
event: nominated
by: operator-policy
detail: "1TB NAS strategic reserve — capability-first, not run-envelope-gated."
- at: "2026-07-28"
event: profile_swot_added
by: grok
detail: "schema 0.2 profile + SWOT."

44
inventory/catalog/meta-llama__Llama-3.2-3B-Instruct__candidate.yaml
View file

@ -7,6 +7,7 @@ source:
url: https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct
revision: main
model_card_url: https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct
project_url: https://www.llama.com/
license:
spdx: custom
url: https://ai.meta.com/llama/license/
@ -14,6 +15,41 @@ license:
allows_local_ops: true
allows_fine_tune: true
notes: "Llama Community License — not MIT; review terms before commercial redistribution."
profile:
summary: "Tiny edge instruct for always-on and low-VRAM agent micro-services."
original_source: https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct
use_cases:
- "Always-on local helper with minimal power/VRAM"
- "Fast classification, routing, and short-form rewrites"
- "CPU or 4–8 GB GPU edge boxes"
- "Smoke-test harnesses before promoting to 8B+"
- "Embedded tooling demos and offline kiosks"
sweet_spots:
- "Latency and footprint over peak intelligence"
- "Huge Llama ecosystem (llama.cpp, Ollama, many templates)"
- "Cheap parallel micro-agents in a fleet"
not_ideal_for:
- "Serious coding or long documents"
- "Domain FT when quality matters (prefer 8B base)"
- "Standalone reasoning-heavy ops decisions"
capability_notes: >
Compact Meta instruct model. Best as a floor for local intelligence and
high-volume simple tasks, not as the lab's primary brain.
swot:
strengths:
- "Very small; runs almost anywhere in the lab"
- "Massive tooling and community template coverage"
- "Good enough for structured short outputs"
weaknesses:
- "Shallow capability; fails hard tasks silently if not constrained"
- "Custom Llama license (not pure OSS-permissive)"
- "HF gated access friction"
opportunities:
- "Fleet of specialized tiny adapters per workflow"
- "Guardrail / router model in front of larger backends"
threats:
- "Qwen/Gemma/Smol peers may outclass at same size"
- "License or regional policy changes for Llama family"
size:
total_bytes: 0
total_human: "~6 GB fp16 / ~2 GB Q4 (estimate)"
@ -31,8 +67,8 @@ collection:
storage_path: ""
brief_refs:
- research/2026-07-24-baseline-field-survey.md
reason: "P0 edge instruct — tiny, huge ecosystem, good CPU/GPU floor for agents."
tags: [instruct, edge, llama]
reason: "P0/R edge instruct — tiny, huge ecosystem, good CPU/GPU floor for agents."
tags: [tier-r, instruct, edge, llama]
companions: []
notes: "HF gated model — need accepted license on account before download."
history:
@ -40,3 +76,7 @@ history:
event: nominated
by: baseline-survey
detail: "P0 recommendation from initial deep research."
- at: "2026-07-24"
event: profile_swot_added
by: grok
detail: "schema 0.2 profile + SWOT."

84
inventory/catalog/meta-llama__Llama-3.3-70B-Instruct__strategic.yaml Normal file
View file

@ -0,0 +1,84 @@
id: meta-llama__Llama-3.3-70B-Instruct__strategic
status: candidate
name: Llama-3.3-70B-Instruct
org: meta-llama
source:
kind: huggingface
url: https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct
revision: main
model_card_url: https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct
project_url: https://www.llama.com/
license:
spdx: custom
url: https://ai.meta.com/llama/license/
allows_offline_retention: true
allows_local_ops: true
allows_fine_tune: true
notes: "Llama Community License — review before commercial redistribution. Prefer Llama 4 open text sibling if it is clearly stronger at pull time."
profile:
summary: "Strategic dense Llama 70B instruct — ecosystem-rich open reserve for multi-GPU."
original_source: https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct
use_cases:
- "Future multi-GPU general instruct with mature Llama tooling"
- "Portable dense alternative to giant MoE for lab upgrades"
- "Ecosystem baselines (vLLM, llama.cpp, eval harnesses expect Llama ids)"
- "Domain FT when Llama license fits the use"
sweet_spots:
- "S4 dense open with the broadest third-party tooling surface"
- "More portable serve path than 600B-class MoE"
- "Strong general instruct heritage in the 70B class"
not_ideal_for:
- "Current single-GPU daily ops (use 3B/8B R spine)"
- "Uses forbidden by Llama Community License terms"
- "Duplicate with Qwen-72B under tight quota — pick one first"
capability_notes: >
HF gated. Strategic dense pick. If Llama 4 open text weights supersede
clearly, nominate successor and mark this entry superseded. License is not
Apache — operator must re-read terms before redistribution.
swot:
strengths:
- "Huge ecosystem and ops familiarity"
- "Solid 70B dense instruct quality"
- "Easier multi-GPU dense story than MoE giants"
weaknesses:
- "Custom Llama license (not Apache/MIT)"
- "HF gating friction for automated pulls"
- "May lag peer open MoE on some benches"
opportunities:
- "First dense strategic if Qwen-72B is deferred"
- "Broad eval comparability with published Llama numbers"
threats:
- "Llama 4 open may obsolete 3.3 quickly"
- "Quota competition with Qwen-72B and S1 MoE"
size:
total_bytes: 0
total_human: "~140 GB fp16 / ~40 GB Q4 (estimate)"
hardware_class:
min_vram_gb_q4: 40
min_vram_gb_fp16: 140
notes: "T3+ to run well; still valuable reserve for future multi-GPU."
axes: [A, B, C]
priority: high
collection:
approved_by: ""
approved_at: null
downloaded_at: null
downloaded_by: ""
storage_path: ""
brief_refs:
- research/2026-07-24-nas-strategic-collection-plan.md
- docs/decisions/2026-07-24-nas-strategic-reserve.md
reason: "S4 strategic dense — strong ecosystem 70B open instruct; more portable than full MoE for a future lab upgrade."
tags: [tier-s, strategic, dense, instruct, llama]
companions:
- meta-llama__Llama-3.2-3B-Instruct__candidate
notes: "HF gated. If Llama 4 open weights supersede, nominate successor and supersede this entry."
history:
- at: "2026-07-24"
event: nominated
by: operator-policy
detail: "Strategic dense open for NAS capability reserve."
- at: "2026-07-28"
event: profile_swot_added
by: grok
detail: "schema 0.2 profile + SWOT."

45
inventory/catalog/nomic-ai__nomic-embed-text-v1.5__candidate.yaml
View file

@ -7,6 +7,7 @@ source:
url: https://huggingface.co/nomic-ai/nomic-embed-text-v1.5
revision: main
model_card_url: https://huggingface.co/nomic-ai/nomic-embed-text-v1.5
project_url: https://www.nomic.ai/
license:
spdx: Apache-2.0
url: https://huggingface.co/nomic-ai/nomic-embed-text-v1.5
@ -14,6 +15,39 @@ license:
allows_local_ops: true
allows_fine_tune: true
notes: ""
profile:
summary: "Lightweight open text embedder for A/B with BGE-M3 and long-context retrieval experiments."
original_source: https://huggingface.co/nomic-ai/nomic-embed-text-v1.5
use_cases:
- "Secondary / A/B embedding path vs BGE-M3"
- "English-heavy RAG with low resource cost"
- "Long-context embedding experiments (per card capabilities)"
- "Portable laptop offline search indexes"
sweet_spots:
- "Small footprint Apache embed"
- "Fast iteration when re-embedding corpora often"
- "Nomic tooling and Matryoshka-style dimension flexibility (per card)"
not_ideal_for:
- "Sole multilingual production embed without A/B (prefer BGE-M3 default)"
- "Generation tasks"
- "Code-only retrieval (prefer code embed if held)"
capability_notes: >
Compact Nomic text embedding model. Held as runnable dual-path to avoid
single-vendor embed lock-in and to re-eval retrieval quality cheaply.
Prefer current v2 text card at pull if clearly better.
swot:
strengths:
- "Tiny, Apache-2.0, easy to re-run indexes"
- "Diversifies embed dependency next to BGE-M3"
weaknesses:
- "May lose multilingual or domain A/B vs BGE-M3"
- "Version churn (v1.5 vs v2) needs careful pin"
opportunities:
- "Dimension-reduced storage for large corpora"
- "Hybrid ensembles (BGE + nomic signals)"
threats:
- "Becomes redundant if one embed wins all lab evals"
- "Upstream rename/deprecation of v1.5"
size:
total_bytes: 0
total_human: "<1 GB"
@ -31,12 +65,17 @@ collection:
storage_path: ""
brief_refs:
- research/2026-07-24-baseline-field-survey.md
reason: "P0 lightweight embed for A/B with BGE-M3; long-context text retrieval."
tags: [embedding, rag]
companions: []
reason: "P0/R lightweight embed for A/B with BGE-M3; long-context text retrieval."
tags: [tier-r, embedding, rag]
companions:
- BAAI__bge-m3__candidate
notes: "If v2 text is preferred at collection time, update id/url and supersede this candidate."
history:
- at: "2026-07-24"
event: nominated
by: baseline-survey
detail: "P0 recommendation from initial deep research."
- at: "2026-07-24"
event: profile_swot_added
by: grok
detail: "schema 0.2 profile + SWOT."

179
inventory/collection-policy.md
View file

@ -1,48 +1,71 @@
# Collection policy — open-weight reserve
**Status:** foundation
**Related:** `schema.yaml`, `docs/backup-storage-policy.md`, `INTENT.md`
**Status:** updated for 1 TB NAS + strategic capability reserve (2026-07-24)
**Related:** `schema.yaml`, `docs/backup-storage-policy.md`,
`docs/decisions/2026-07-24-nas-strategic-reserve.md`, `INTENT.md`
---
## Purpose
Decide **what** enters the open-weight reserve, **who** may approve it, and
**when** a daily-brief candidate becomes a catalog entry with blobs on backup
storage.
**when** a daily-brief candidate becomes a catalog entry with blobs on the NAS.
---
## Goals
* Keep a **small, high-leverage** reserve — not a Hugging Face mirror
* Hold a **capability-first strategic reserve** of the best open weights we may
need later — **including models too large to run on current lab hardware**
* Also hold a **runnable spine** for day-to-day local ops and fine-tunes
* Enforce **license and integrity** before download completes
* Match capacity to the backup storage soft quota
* Prefer models that serve axes **B** and **C**, plus strategic **A** open releases
* Stay within the **1 TB NAS soft quota (850 GiB)** — quality over mirror volume
* Prefer official provenance; not a full Hugging Face scrape
---
## Two envelopes (do not conflate)
| Envelope | Question | Effect on collection |
| -------- | -------- | -------------------- |
| **Run** | Can we infer / FT this *now*? | Tags `hardware_class`, prioritizes R tier for local work |
| **Reserve** | Is this among the best open artifacts worth keeping *just in case*? | **Not gated by current VRAM** |
**Runnability is not an eligibility gate.** An unrunnable SOTA open MoE can be
priority **high** if license and quota allow.
---
## Eligibility (must pass all)
1. **Open weights** — weights obtainable under terms that allow offline retention for lab use
2. **Clear license** — SPDX or linkable license text; `allows_offline_retention: true`
3. **Stable provenance** — official org, tagged release, or commit revision (not anonymous drive-by reupload as sole source)
4. **Lab rationale** — written `reason` tied to at least one axis A–D (usually B/C)
5. **Capacity** — estimated size fits under remaining soft quota (see backup storage policy)
1. **Open weights** — obtainable under terms that allow offline retention for lab use
2. **Clear license** — SPDX or linkable license text; `allows_offline_retention: true`
3. **Stable provenance** — official org, tagged release, or commit revision
4. **Lab rationale** — written `reason` (capability SOTA, runnable spine, embed, FT base, …)
5. **Capacity** — estimated size fits under remaining soft quota (850 GiB)
Fail any gate → status `rejected` with reason, or never enter catalog.
Fail any gate → `rejected` or never enter catalog.
---
## Priority rubric
## Collection tiers
| Tier | Code | Meaning |
| ---- | ---- | ------- |
| **Runnable spine** | **R** | Default local chat/code/embed/FT bases; keep resident |
| **Strategic capability** | **S** | Most capable open models (often large); may be T4+ only to *run* |
| **Watch / optional** | **W** | Secondary; collect only with clear headroom |
Catalog field: use `tags` including `tier-r` / `tier-s` / `tier-w` and
`priority: high|medium|low`.
### Priority rubric
| Priority | Guidance |
| -------- | -------- |
| **high** | Rare or strategically important; license/access risk of disappearance; uniquely strong for B/C at our hardware class; hard to re-obtain |
| **medium** | Clear lab use within 1–2 quarters; good quality/cost; easy enough to re-download but worth having cold |
| **low** | Nice to have; only collect if quota headroom is large and pull is cheap |
Daily brief **collection candidates** should set a suggested priority; approval may change it.
| **high** | Top open capability (S) or essential runnable spine (R); hard to re-obtain; license/access risk |
| **medium** | Strong but not unique; mid-size upgrades; companions (rerankers) |
| **low** | Nice-to-have; W tier |
---
@ -50,16 +73,48 @@ Daily brief **collection candidates** should set a suggested priority; approval
| Estimated total size | Approval |
| -------------------- | -------- |
| **&lt; 5 GiB** | Operator or lab agent may collect after license check; catalog entry required before or immediately after |
| **5–40 GiB** | Explicit operator approval (chat, workplan task, or signed catalog `approved_by`) |
| **&gt; 40 GiB** | Operator approval **plus** check against soft quota and whether a smaller quant/variant suffices |
| **Any size if quota ≥ 70% used** | Operator approval required regardless of size |
| **Unclear license or ToS risk** | Do not collect; status `rejected` |
| **&lt; 5 GiB** | Operator or lab agent after license check |
| **5–40 GiB** | Explicit operator approval |
| **&gt; 40 GiB** | Operator approval + remaining soft-quota check |
| **&gt; 200 GiB (typical S giants)** | Operator approval + written note on which other S models may need to wait |
| **Any size if quota ≥ 70% used** | Operator approval required |
| **Unclear license or ToS risk** | Do not collect; `rejected` |
“Operator” means the human lab owner (or a documented delegate). Agents may
**nominate** (`status: candidate`) freely from briefs; they may **collect** only
within the &lt; 5 GiB band when licenses are unambiguous — otherwise stop at
`candidate` / `approved`.
Agents may **nominate** freely; they **collect** only in the &lt; 5 GiB band with
unambiguous licenses unless the operator has approved the catalog entry.
---
## What we prefer to collect
### Runnable spine (R)
* Small/mid instruct and code models that fit the run envelope
* Strong embedding / rerank models for local RAG
* Bases known to fine-tune well under QLoRA on lab GPUs
### Strategic capability (S)
* **Best available open general / reasoning / code weights** at the frontier of open
* Large MoE or dense models even if current infra cannot serve them
* Prefer official compressed releases (FP8, published quant) when full precision
would exhaust the 1 TB NAS
* One clear “best open” per capability niche is better than five near-duplicates
### Companions
Tokenizers, LoRA adapters, small eval fixtures when required to use a reserved base.
---
## What we usually skip
* Duplicate quants of the same revision already reserved
* Anonymous merges/repacks without provenance
* Closed weights
* Entire org mirrors
* Giant pretraining corpora (default out of band unless separately justified)
* Anything whose license forbids offline retention
---
@ -67,68 +122,30 @@ within the &lt; 5 GiB band when licenses are unambiguous — otherwise stop at
```text
brief nominates
→ candidate (catalog YAML, no blobs required)
→ approved (license + size + quota OK)
→ collecting (download in staging/)
→ collected (blobs complete, checksums recorded, storage_path set)
→ verified (optional re-hash / smoke load)
→ superseded|evicted (replaced or removed; metadata kept)
→ candidate
→ approved
→ collecting (NAS staging/)
→ collected
→ verified
→ superseded|evicted
```
Rejected candidates stay in catalog only if useful as a decision record; otherwise omit.
---
## What we prefer to collect
* Small/mid instruct and code models that fit the hardware envelope
* Strong embedding / rerank models for local RAG
* Base models known to fine-tune well under QLoRA/LoRA on lab GPUs
* Official quant releases when they are the supported distribution
* Adapters and tokenizers that unlock a reserved base (as companions)
## What we usually skip
* Duplicate quants of the same revision already reserved
* Huge models with no near-term local run/train path and no access-risk story
* Merges/repacks without provenance
* Datasets larger than model weights unless separately justified (default: out of band)
* Anything requiring acceptance flows we cannot satisfy offline
---
## Companions
Tokenizers, LoRA adapters, and small eval fixtures may be collected when:
* they are required to use a reserved base, or
* they are small (&lt; 1 GiB) and high leverage
Link via `companions` in the catalog schema.
---
## Brief integration
1. Brief section **Collection candidates** nominates items.
2. Operator/agent opens `inventory/catalog/{id}.yaml` with `status: candidate`.
3. Approval and download follow this policy and `docs/backup-storage-policy.md`.
4. Brief `brief_refs` on the entry point back to the nominating day(s).
1. Brief **Collection candidates** nominates R/S/W.
2. Catalog YAML under `inventory/catalog/`.
3. Download only after storage path is pinned (`docs/backup-storage-policy.md`) and approval rules pass.
4. `collection.brief_refs` / research refs for provenance of the nomination.
---
## Eviction rule of thumb
When over quota or cleaning:
When over soft quota:
1. `low` priority, easily re-obtainable from still-live official URLs
2. Superseded revisions with a newer `verified` replacement
3. Never silent-delete: set `status: evicted`, clear or note `storage_path`, append `history`
---
## Non-goals
* Automatic bulk mirrors of entire orgs
* Collecting on every brief mention without priority
* Bypassing license gates for “research only” convenience
1. **W** tier and easily re-obtainable duplicates
2. Superseded revisions with a stronger verified successor
3. Never silent-delete **S** SOTA or sole **R** spine without operator note
4. Always set `status: evicted` and append `history`

110
inventory/schema.yaml
View file

@ -3,7 +3,7 @@
# Filename suggestion: {org}__{name}__{short_revision}.yaml
#
# Schema version documents field meaning for humans and future validators.
schema_version: "0.1.0"
schema_version: "0.2.0"
# ---------------------------------------------------------------------------
# Example entry (illustrative only — not a real collection claim)
@ -15,60 +15,56 @@ schema_version: "0.1.0"
# source:
# kind: huggingface # huggingface | github_release | direct_url | other
# url: https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct
# revision: abc1234def... # commit sha, tag, or release id
# revision: abc1234def...
# model_card_url: https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct
# license:
# spdx: llama3.2 # or MIT, Apache-2.0, etc.; use "custom" + notes if needed
# url: https://...
# allows_offline_retention: true
# allows_local_ops: true
# allows_fine_tune: true
# notes: ""
# artifacts:
# - path: blobs/model.safetensors
# sha256: "..."
# bytes: 0
# size:
# total_bytes: 0
# total_human: "0 B"
# hardware_class:
# # Rough lab guidance — not a guarantee
# min_vram_gb_q4: 4
# min_vram_gb_fp16: 8
# notes: "fits consumer 8GB at Q4"
# axes: # why this matters to the lab
# - B
# - C
# priority: medium # low | medium | high
# collection:
# approved_by: ""
# approved_at: null # ISO-8601 date
# downloaded_at: null
# downloaded_by: ""
# storage_path: "" # absolute or facility-relative path on backup storage
# brief_refs: # briefs that nominated this candidate
# - briefs/2026/07/2026-07-23.md
# reason: "Strong small instruct base; clear license; homelab-fit."
# tags:
# - instruct
# - text
# companions: [] # optional related catalog ids (adapters, tokenizers)
# project_url: https://www.llama.com/ # optional: org project / paper home
# paper_url: https://arxiv.org/... # optional
# license: { ... }
# profile:
# summary: "One-line what this model is for."
# original_source: "https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct"
# use_cases:
# - "Always-on local assistant on small VRAM"
# - "Edge classification and simple tool routing"
# sweet_spots:
# - "Latency-sensitive chat under 3–8 GB VRAM"
# - "Teaching / prototyping agent loops cheaply"
# not_ideal_for:
# - "Hard multi-file software engineering"
# - "Long-horizon reasoning"
# capability_notes: >
# Compact instruct model; strong ecosystem tooling; limited depth vs mid-size.
# swot:
# strengths:
# - "Tiny footprint; huge Llama tooling ecosystem"
# weaknesses:
# - "Shallow reasoning and coding vs 8B–70B class"
# opportunities:
# - "Domain LoRA for NetKingdom micro-agents"
# threats:
# - "Superseded quickly by next small open release"
# size: { ... }
# hardware_class: { ... }
# axes: [B]
# priority: high
# collection: { ... }
# reason: "..."
# tags: [tier-r, instruct]
# companions: []
# notes: ""
# history:
# - at: "2026-07-23"
# event: nominated
# by: operator
# detail: "From daily brief collection candidates."
# history: []
# ---------------------------------------------------------------------------
# Required fields by status (normative for humans; tooling may enforce later)
# Required fields by status
# ---------------------------------------------------------------------------
# candidate: id, status, name, org, source.url, source.revision, license, reason, priority
# candidate: id, status, name, org, source.url, source.revision, license,
# reason, priority, profile (summary + original_source + use_cases +
# sweet_spots), swot (all four lists, compact)
# approved: + collection.approved_by, collection.approved_at
# collected: + collection.downloaded_at, collection.storage_path, size, artifacts[].sha256
# verified: + checksums re-read OK (notes or history event)
# rejected: + reason (why rejected)
# superseded / evicted: + history event; storage_path may be empty after eviction
# collected: + collection.downloaded_at, collection.storage_path, size, artifacts
# verified: + checksums re-read OK
# rejected: + reason
# superseded / evicted: + history event
field_reference:
id:
@ -83,10 +79,24 @@ field_reference:
type: string
source:
type: object
fields: [kind, url, revision, model_card_url]
fields: [kind, url, revision, model_card_url, project_url, paper_url]
license:
type: object
fields: [spdx, url, allows_offline_retention, allows_local_ops, allows_fine_tune, notes]
profile:
type: object
description: Human-facing capability card for operators choosing models
fields:
summary: One-line positioning
original_source: Canonical URL (usually HF model card or org release page)
use_cases: List of typical lab / product uses
sweet_spots: Where the model punches above weight or is the default pick
not_ideal_for: Explicit anti-patterns
capability_notes: Short free-text capability narrative
swot:
type: object
description: Compact SWOT for reserve / ops decisions (2–4 bullets each)
fields: [strengths, weaknesses, opportunities, threats]
artifacts:
type: list
item_fields: [path, sha256, bytes]

13
research/2026-07-24-baseline-field-survey.md
View file

@ -216,12 +216,19 @@ Freedom Intelligence **reports** on harness tech; it does not replace sand-boxer
| 10 | Code embed: `nomic-ai/nomic-embed-code` or Jina code embed | Code RAG for repos |
| 11 | Small reranker: `BAAI/bge-reranker-v2-m3` | Cheap quality win for RAG |
### 6.4 P2 — watch / large (candidate or docs only)
### 6.4 Strategic large (updated 2026-07-24)
**Superseded by operator decision:** 1 TB local NAS + capability-first reserve.
Full DeepSeek-V3/R1-class and 70B dense opens **are in scope** even if unrunnable
on current GPUs. See:
- `docs/decisions/2026-07-24-nas-strategic-reserve.md`
- `research/2026-07-24-nas-strategic-collection-plan.md`
| Item | Guidance |
| ---- | -------- |
| Full **DeepSeek-V3 / R1** MoE (~600B+ class, hundreds of GB even quantized) | **Do not auto-collect.** Separate capacity + multi-GPU decision |
| **Llama 70B / Qwen 72B** class | Only if lab gains multi-GPU and clear offline need |
| Full **DeepSeek-V3 / R1** class | **Strategic (S)** — collect under 850 GiB soft quota; prefer official compressed if needed |
| **Llama 70B / Qwen 72B** class | **Strategic (S)** dense — good second fill after one primary MoE |
| Every new quant repack on HF | Skip; track base official only |
### 6.5 Explicit non-recommendations (for now)

94
research/2026-07-24-nas-strategic-collection-plan.md Normal file
View file

@ -0,0 +1,94 @@
# NAS strategic collection plan (1 TB)
**Date:** 2026-07-24
**Decision:** `docs/decisions/2026-07-24-nas-strategic-reserve.md`
**Soft quota:** 850 GiB on 1 TB local NAS
---
## Intent
Capture the **most capable open-weight models** for long-term optionality, **even
if they cannot run on current lab GPUs**, plus keep a **runnable spine** for
daily local use.
---
## Tier R — runnable spine (keep resident)
| Model | Est. (Q4 / fp16) | Role |
| ----- | ---------------- | ---- |
| Qwen3-8B | ~5 / ~16 GiB | Default instruct + FT |
| Llama-3.2-3B-Instruct | ~2 / ~6 GiB | Edge |
| BGE-M3 | ~2 GiB | Multilingual embed |
| DeepSeek-R1-Distill-Qwen-14B | ~9 / ~28 GiB | Local reason |
| nomic-embed-text-v1.5 | &lt;1 GiB | Light embed |
| **Subtotal (quant-first)** | **~20–30 GiB** | |
| **Subtotal (fp16-ish)** | **~50–55 GiB** | |
---
## Tier S — strategic capability (best open, may be unrunnable)
Prioritize **one primary artifact per capability niche** so 1 TB is not wasted
on near-duplicates. Prefer official compressed weights when full precision is
hundreds of GiB–TB.
| Priority | Capability niche | Representative targets (verify current HF id + license at pull) | Size class (order of magnitude) |
| -------- | ---------------- | ---------------------------------------------------------------- | -------------------------------- |
| **S1** | Best open general / MoE chat | DeepSeek-V3 / V3.2 / V4-class open MoE (latest card) | ~300–700+ GiB depending on quant/precision |
| **S2** | Best open reasoner | DeepSeek-R1 full (or successor open reasoner) | similar large |
| **S3** | Best open coding / SWE open weight | Top open code MoE or dense (e.g. GLM-5.x open, Qwen coder large — pick current SOTA open) | tens–hundreds GiB |
| **S4** | Strong open dense mid-large | Qwen3-72B or Llama 3.3 70B / Llama 4 open text variant | ~40–150 GiB quant / larger fp |
| **S5** | Efficient open frontier family | Gemma 4 large open variants | tens of GiB |
**1 TB portfolio rule of thumb**
```text
R spine (~30–55 GiB)
+ S1 one primary giant quant (~300–450 GiB) ← first strategic fill
+ optional S4 dense 70B (~40–80 GiB Q4)
+ embeds already in R
+ staging headroom
≈ 400–600 GiB typical first strategic fill
```
If a single full-precision S1 would exceed soft quota alone, **take the best
official compressed release** of that model and record the base id as the
capability identity.
**Do not** try to store S1 + S2 + S3 all at full precision on this NAS.
---
## Tier W — optional (quota permitting)
* Qwen3-14B, R1-Distill-32B, BGE reranker (already catalog medium)
* Extra quants of models already held in another format
* Code-specific embeds
---
## Explicitly out
* Closed API-only weights
* Full HF org mirrors
* Multiple anonymous “uncensored” reuploads of the same base
---
## Download order (recommended)
1. Pin mount path in backup-storage-policy (done 2026-07-28: `D:\vault\coulomb\freedom-intelligence\`)
2. **R spine** (fast wins, unblocks local work)
3. **S1** — single best open general MoE (compressed if needed)
4. Reassess free space
5. **S4** or **S2** next depending on which niche matters more for Coulomb
6. W-tier only with ≥200 GiB free under soft quota
---
## Catalog
R and selected S/W candidates live under `inventory/catalog/`. Strategic S
entries use tags `tier-s`, `strategic`, `beyond-run-envelope`.

30
workplans/FI-WP-0001-lab-operating-foundation.md
View file

@ -2,14 +2,15 @@
id: FI-WP-0001
type: workplan
title: "Lab operating foundation: sources, playbook, storage pin, hardware envelope"
domain: infotech
domain: agents
repo: freedom-intelligence
status: active
status: done
owner: grok
topic_slug: freedom-intelligence
priority: high
created: "2026-07-24"
updated: "2026-07-24"
updated: "2026-07-28"
state_hub_workstream_id: "0181d782-fdac-4366-9fc7-d42dac97756b"
---
# FI-WP-0001 — Lab operating foundation
@ -42,6 +43,7 @@ and seeds collection candidates under `inventory/catalog/`.
id: FI-WP-0001-T01
status: done
priority: high
state_hub_task_id: "e4748af3-f70f-4650-afe0-3b55def41d1c"
```
Create `docs/sources-allowlist.md` with standing channels for axes A–D:
@ -62,6 +64,7 @@ Create `docs/sources-allowlist.md` with standing channels for axes A–D:
id: FI-WP-0001-T02
status: done
priority: high
state_hub_task_id: "781b6cac-fece-4139-9ee5-c08a00ee1267"
```
Write `docs/daily-brief-playbook.md`:
@ -83,6 +86,7 @@ Write `docs/daily-brief-playbook.md`:
id: FI-WP-0001-T03
status: done
priority: medium
state_hub_task_id: "359b58d1-2692-420c-8f28-324b3a5ceb3d"
```
Write `docs/hardware-envelope.md` with **honest TBD** for unmeasured hosts, plus
@ -99,8 +103,9 @@ without schema changes.
```task
id: FI-WP-0001-T04
status: todo
status: done
priority: high
state_hub_task_id: "a0f6beb6-3273-4957-b9f6-c33b4ba7967f"
```
Operator task: fill TBD fields in `docs/backup-storage-policy.md`:
@ -115,6 +120,19 @@ Create empty `models/` and `staging/` on that facility.
**Done when:** policy has no critical TBD for path/quota; `FI-WP-0003` may download.
**Completed 2026-07-28:**
| Field | Value |
| ----- | ----- |
| Facility | Workstation VAULT HD (`D:`) |
| Base path | `D:\vault\coulomb\freedom-intelligence\` |
| WSL path | `/mnt/d/vault/coulomb/freedom-intelligence/` (when D: automounted) |
| Soft / hard | **850 GiB** / **920 GiB** |
| Layout | `models/`, `staging/`, `companions/` created |
Strategic unrunnable models remain in scope
(`docs/decisions/2026-07-24-nas-strategic-reserve.md`).
---
### T05 — Link baseline research into lab navigation
@ -123,6 +141,7 @@ Create empty `models/` and `staging/` on that facility.
id: FI-WP-0001-T05
status: done
priority: medium
state_hub_task_id: "ac382e31-5401-425a-bb7b-cf1ff969a937"
```
- Point `README.md` / `SCOPE.md` at `research/2026-07-24-baseline-field-survey.md`
@ -136,9 +155,10 @@ priority: medium
## Acceptance (workplan-level)
- [x] Sources allowlist + daily playbook + hardware envelope drafted
- [ ] Backup storage path and soft quota pinned (T04 — operator)
- [x] Backup storage: path + quota pinned under `D:\vault\coulomb\freedom-intelligence\` (T04)
- [x] Baseline research persisted and linked
- [x] Catalog candidates seeded for recommended models
- [x] Strategic reserve policy (capability beyond run envelope) documented
## Out of scope

9
workplans/FI-WP-0002-activity-core-daily-research.md
View file

@ -2,7 +2,7 @@
id: FI-WP-0002
type: workplan
title: "Activity-core daily research brief rhythm"
domain: infotech
domain: agents
repo: freedom-intelligence
status: active
owner: grok
@ -12,6 +12,7 @@ created: "2026-07-24"
updated: "2026-07-24"
depends_on:
- FI-WP-0001
state_hub_workstream_id: "aa5b118e-36e2-41a7-b6c1-5d9ac9c7fb41"
---
# FI-WP-0002 — Activity-core daily research brief rhythm
@ -68,6 +69,7 @@ Weekends optional later (`0 9 * * 6` optional Saturday scan) — not in v1.
id: FI-WP-0002-T01
status: done
priority: high
state_hub_task_id: "4eb44be8-04b7-45c9-9e78-e573a3501524"
```
Add `activity-definitions/fi-daily-research-brief.md`:
@ -90,6 +92,7 @@ existing definitions (binky / forgejo prune).
id: FI-WP-0002-T02
status: done
priority: high
state_hub_task_id: "ce61769e-9c32-4757-a691-c443e6cec37d"
```
Document and implement playbook steps for:
@ -117,6 +120,7 @@ Wire into `docs/daily-brief-playbook.md` and `docs/state-hub-delivery.md`.
id: FI-WP-0002-T03
status: done
priority: high
state_hub_task_id: "37a5d67c-93b4-4cbd-996c-e87b9308e295"
```
**Cross-repo (activity-core):** implement state-hub (or shell) resolver:
@ -141,6 +145,7 @@ Mirror `binky_rhythm_status` pattern. Add unit tests with mocked hub.
id: FI-WP-0002-T04
status: todo
priority: high
state_hub_task_id: "ecb8df0a-4f5d-494e-9217-2906afbacf9c"
```
**Cross-repo / ops:** ensure railiance activity-core worker loads this repo’s
@ -158,6 +163,7 @@ path). Run `make sync-activity-definitions` / schedule reconciliation.
id: FI-WP-0002-T05
status: todo
priority: medium
state_hub_task_id: "4d7a7491-f458-4984-bda1-8d212046dbc6"
```
1. Manual `POST .../activity-definitions/<id>/trigger` with def still disabled
@ -177,6 +183,7 @@ priority: medium
id: FI-WP-0002-T06
status: done
priority: medium
state_hub_task_id: "2df977ea-06b4-468d-b6fb-6b40e00f8928"
```
Document how harness/operator picks up `target_repo: freedom-intelligence`

75
workplans/FI-WP-0003-seed-open-weight-reserve.md
View file

@ -2,40 +2,49 @@
id: FI-WP-0003
type: workplan
title: "Seed open-weight reserve from baseline recommendations"
domain: infotech
domain: agents
repo: freedom-intelligence
status: active
owner: grok
topic_slug: freedom-intelligence
priority: medium
created: "2026-07-24"
updated: "2026-07-24"
updated: "2026-07-28"
depends_on:
- FI-WP-0001
state_hub_workstream_id: "7513a853-24bf-4826-9ed9-3e8ed8ab0fa5"
---
# FI-WP-0003 — Seed open-weight reserve from baseline recommendations
## Goal
Turn baseline survey recommendations into a **real, policy-compliant reserve**:
approve high-priority catalog candidates, download to pinned backup storage,
verify checksums, and leave medium/low items as candidates for later briefs.
Turn baseline + **strategic reserve policy** into a **real, policy-compliant reserve**:
approve high-priority catalog candidates, download to the **pinned bulk store**,
verify checksums, and leave lower-priority items as candidates for later briefs.
**Policy (2026-07-24 / path 2026-07-28):** include the **most capable open-weight
models** even when they exceed current run hardware. Soft quota **850 GiB**.
**Storage pin:** `D:\vault\coulomb\freedom-intelligence\` (Windows) /
`/mnt/d/vault/coulomb/freedom-intelligence/` (WSL when D: mounted).
Source of truth for *which* models:
- `research/2026-07-24-baseline-field-survey.md` § Collection recommendations
- `inventory/catalog/*.yaml` (seeded as `candidate`)
- `research/2026-07-24-baseline-field-survey.md`
- `research/2026-07-24-nas-strategic-collection-plan.md`
- `docs/decisions/2026-07-24-nas-strategic-reserve.md`
- `inventory/catalog/*.yaml`
Gates: `inventory/collection-policy.md` + `docs/backup-storage-policy.md` (T04 of FI-WP-0001).
Gates: `inventory/collection-policy.md` + `docs/backup-storage-policy.md`.
## Priority tiers (from survey)
## Priority tiers
| Tier | Action in this workplan |
| ---- | ----------------------- |
| **P0 — seed now** | Approve + collect after storage pin (small/mid, high leverage) |
| **P1 — next wave** | Approve when quota allows; may stay candidate |
| **P2 — watch / large** | Keep candidate; do not bulk-pull full 600B+ MoE without explicit operator decision |
| **R — runnable spine** | Approve + collect first (former P0) |
| **S — strategic capability** | Approve + collect top open giants (V3/R1 class, 70B dense, …) under quota; compressed OK |
| **W — watch** | Optional after R+S; may stay candidate |
## Tasks
@ -45,6 +54,7 @@ Gates: `inventory/collection-policy.md` + `docs/backup-storage-policy.md` (T04 o
id: FI-WP-0003-T01
status: done
priority: high
state_hub_task_id: "5d2a493b-a189-4eae-8a40-18ce8e0ab17d"
```
Create `status: candidate` entries under `inventory/catalog/` for every P0/P1
@ -54,55 +64,60 @@ recommendation in the baseline survey (and note P2 as candidates or docs-only).
---
### T02 — Operator approve P0 set
### T02 — Operator approve R spine + S priority order
```task
id: FI-WP-0003-T02
status: todo
priority: high
state_hub_task_id: "909263c7-7148-45fa-98bc-5d2bafacf507"
```
Human review of P0 candidates: license, size vs soft quota, hardware fit.
Human review: license, size vs **850 GiB** soft quota, download order from
`research/2026-07-24-nas-strategic-collection-plan.md` (R first, then S1, …).
Set `status: approved`, `collection.approved_by`, `collection.approved_at`.
**Done when:** each P0 entry is `approved` or `rejected` with reason.
**Done when:** each R and chosen S entry is `approved` or `rejected` with reason.
---
### T03 — Download P0 to backup storage + verify
### T03 — Download R spine + strategic S to NAS + verify
```task
id: FI-WP-0003-T03
status: todo
priority: high
state_hub_task_id: "222969d0-f3af-46fb-a303-f76800c80ff9"
```
For each approved P0:
For each approved model:
1. Download into `{BACKUP}/freedom-intelligence/models/...` (or `staging/` then promote)
2. Record `artifacts[].sha256`, `size`, `collection.storage_path`
3. Set `status: collected` then `verified` after re-hash or smoke load
1. Download into `D:\vault\coulomb\freedom-intelligence\models\...` (or `staging/` then promote)
2. For S giants: prefer official compressed if full precision exceeds headroom
3. Record `artifacts[].sha256`, `size`, `collection.storage_path`
4. Set `status: collected` then `verified` after re-hash (smoke load optional for unrunnable S)
Prefer official HF revisions; use `huggingface-cli` or equivalent with revision pin.
**Blocked on:** FI-WP-0001-T04 (storage pin).
**Unblocked:** FI-WP-0001-T04 storage pin complete (2026-07-28). Ensure WSL can
see `/mnt/d` before Linux-side pulls.
**Done when:** all approved P0 are `verified` or explicitly deferred with notes.
**Done when:** approved R (+ first S as space allows) are `verified` or deferred with notes.
---
### T04 — P1 decision pass
### T04 — W-tier and remaining S decision pass
```task
id: FI-WP-0003-T04
status: todo
priority: medium
state_hub_task_id: "33daa2cf-f9fd-430d-a4e4-58996c900ca6"
```
After P0, decide which P1 models fit remaining quota. Approve/collect subset or
leave as candidates for daily briefs to re-prioritize.
After R + primary S fills, decide remaining S/W under free soft-quota capacity.
**Done when:** each P1 has an explicit next status (`approved`, `candidate`, `rejected`).
**Done when:** each remaining catalog entry has an explicit next status.
---
@ -112,6 +127,7 @@ leave as candidates for daily briefs to re-prioritize.
id: FI-WP-0003-T05
status: todo
priority: low
state_hub_task_id: "9974b98a-e1b3-4b7a-b2c2-0fd2cba65f97"
```
Write `inventory/RESERVE-STATUS.md` summarizing collected vs candidate totals,
@ -123,9 +139,10 @@ bytes used vs soft quota, and hardware coverage gaps.
## Acceptance (workplan-level)
- [x] Catalog seeded from baseline survey
- [ ] Storage pin complete (upstream)
- [ ] P0 verified on backup media
- [x] Catalog seeded from baseline survey (+ profile/SWOT on all entries)
- [x] Storage pin complete (upstream FI-WP-0001-T04)
- [ ] Operator approve R + S priority (T02)
- [ ] P0/R verified on backup media (T03)
- [ ] RESERVE-STATUS snapshot after first collections
## Out of scope