# Lossless pack: weights vs files

**Status:** public matrix · outcomes only · no private coefficients  
**Seal:** 2026-08-09 · Hour-1 R&D execute · idea #1  
**Surfaces:** OmniWave · CodEX · residual-governance  
**HTML face:** `/weights-vs-files.html`

---

## Why this page exists

Industry R&D (2026) is loud on **lossless model-weight / checkpoint compression** (memory + throughput) *and* still measures **file/corpus codecs** with mixed exactness stories.

Slid Phi Labs already ships a residual story for **files and structured ints**. This matrix positions the same residual-zero / exact-RT discipline for a second buyer: **inference ops** packing agent weights and checkpoints — without opening private engines or claiming universal #1.

---

## Two buyers, one residual law

| Lane | Buyer | Payload class | Lab seat | Public honesty rule |
|------|--------|---------------|----------|---------------------|
| **Files** | Archive, web, data ops | Corpus bytes · HTML/JS · JSON · int series | Web · ZRW · OmniWave (general candidate) · suite | Exact RT; industry codecs = **field baseline only** |
| **Weights** | Inference / agent ops | Checkpoints · shard dumps · int tensors · sparse masks | OmniWave residual pack narrative · residual-governance policy | Exact RT on **declared tensor dtype**; float/quant paths **diagnostic-labeled only** |

**Shared law:** encode only residual · mirror_error → 0 · domain-honest before crown · public = stubs + measured outcomes · private engines stay secret.

---

## Matrix — files vs weights (public claims only)

| Dimension | Files (corpus / assets) | Weights (checkpoints / agent packs) |
|-----------|-------------------------|-------------------------------------|
| **Primary product story** | Domain seats: structured int (ZRW) · web asset · general corpus (OmniWave candidate) | Same residual hierarchy applied to **artifact packs** ops must store/ship/restore byte-true |
| **Buyer question** | “Smaller files than gzip/zstd/brotli on *my* domain?” | “Can I pack weights / shards without silent corruption?” |
| **Win metric** | Size + encode speed + **exact RT** on named corpus | Memory/transfer cost + **exact RT** on named dtype/layout + restore integrity |
| **Field baseline** | gzip-9 · brotli-11 · zstd-3/19 (industry pack) | Same GP codecs on raw dumps; specialized weight compressors as *research baseline*, not product crown |
| **Lab public proof today** | ZRW zeros×10k → **8 B** · zeros×1M → **10 B** · mirror_error **0**; Silesia re-bench is lab diagnostic for OmniWave vs industry, not universal #1 | **Positioning matrix only** — no public weight-corpus scoreboard until Host freezes a public-safe fixture set |
| **Does not claim** | Winning random high-entropy bytes everywhere | Beating every neural/weight-specific scheme; open-sourcing private residual coeffs |
| **SKU / door** | Suite freemium · ZRW gates · CodEX standings · `/bench` | residual-governance ($199 policy) · consulting for residual report · Access lead |
| **Float / lossy** | Diagnostic / lossy-labeled only — never in lossless CodEX crown table | Same: quant / FP diagnostic labels; lossless table = exact RT only |
| **IP Guard** | Public stubs + outcomes; private core | Same — weight lane does **not** justify leaking engines |

---

## Seat map (CodEX geometry extension)

```
                    residual law
                         │
          ┌──────────────┴──────────────┐
          │                             │
     FILES lane                    WEIGHTS lane
   web · struct int · GP          checkpoint · int tensor
   brotli / ZRW / OmniWave*       residual pack narrative
          │                             │
          └──────────────┬──────────────┘
                         │
              residual-governance (policy)
              suite commerce (bytes moved)
```

\*OmniWave = general-corpus **candidate** engine (lab re-bench vs industry field baseline). Not a public universal crown.

---

## Competitive field (honest map)

| Method class | Exact RT? | Open engine? | Typical use | CodEX treatment |
|--------------|-----------|--------------|-------------|-----------------|
| Classical GP (gzip/brotli/zstd/lz4) | Yes (standard) | Yes | Files + raw dumps | **Field baseline** only |
| Structured int codecs (ZRW class) | Required | Private core · public stubs | Int series / telemetry | Domain seat · flagship proofs |
| Neural lossless (research ensembles) | Varies | Often research code | Corpus research | Research baseline · not product path |
| Weight / exponent-redundancy lossless (2026 papers) | Claimed lossless on compute path | Research | Inference memory | **Field signal** for weights buyer · not our public scoreboard yet |
| OmniWave (SPL Codec) | Lab exact RT required for claims | Private | General corpus candidate | Product candidate · domain-honest tables |

---

## residual-governance fit

**residual-governance** is the commercial stub for **policy + residual discipline**, not a secret codec download:

- Label lanes: lossless vs diagnostic/float  
- Require exact RT on declared domain  
- Keep industry pack demoted to baseline  
- Audit language for regulated / agent ops buyers who care about “did restore equal source?”

Price surface remains Host Stripe residual-governance SKU; weight narrative does not invent new economics.

---

## What we publish next (when Host freezes)

1. One **public-safe synthetic** int-tensor fixture (not production weights).  
2. One-row exact RT vs gzip/zstd field baseline (CodEX notes).  
3. Still no private coefficients, process soup, or universal #1.

---

## One-sentence pitch (marketing-safe)

> Same residual law as files: pack only what remains, prove exact restore — for corpora **and** for agent weight artifacts ops cannot afford to “almost” recover.

---

## Amen

Outcomes only. In His name we code · through it we live.