freedom-intelligence/inventory/collection-policy.md

# Collection policy — open-weight reserve

**Status:** updated for Scaleway object reserve (2026-09-13)  
**Related:** `schema.yaml`, `docs/backup-storage-policy.md`,
`docs/decisions/2026-09-13-scaleway-object-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
Scaleway (streamed through a remote worker with bounded RAM; no local staging).

---

## Goals

* 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
* Stay within the **€15 / month** soft budget — 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 euro budget allow.

---

## Eligibility (must pass all)

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 monthly storage cost fits under remaining
   €15 soft budget (Glacier for S, One Zone IA for R)

Fail any gate → `rejected` or never enter catalog.

---

## 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** | 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 |

---

## Approval rule of thumb

| Estimated total size | Approval |
| -------------------- | -------- |
| **< 5 GiB** | Operator or lab agent after license check |
| **5–40 GiB** | Explicit operator approval |
| **> 40 GiB** | Operator approval + remaining **euro-budget** check |
| **> 200 GiB (typical S giants)** | Operator approval + written note that the object goes to **Glacier** |
| **Any size if month ≥ 70% of €15** | Operator approval required |
| **Unclear license or ToS risk** | Do not collect; `rejected` |

Agents may **nominate** freely; they **collect** only in the < 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 blow the euro budget; Glacier makes full official trees
  of the *top* S identities affordable (V4-Flash ~€0.42/mo, K3 ~€3.69/mo)  
* 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  

---

## Lifecycle

```text
brief nominates
    → candidate
    → approved
    → collecting     (remote HTTP stream → S3 multipart)
    → collected
    → verified
    → superseded|evicted
```

---

## Brief integration

1. Brief **Collection candidates** nominates R/S/W.  
2. Catalog YAML under `inventory/catalog/`.  
3. Download only after the Scaleway bucket is live (FI-WP-0004-T08) and
   approval rules pass. Use [diskless streaming](../docs/streaming-reserve.md); SoT is `s3://`.
4. `collection.brief_refs` / research refs for provenance of the nomination.

---

## Eviction rule of thumb

When over the euro soft budget:

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`