# The digests — which builds this battery actually measured

Extracted from the bench's own night record. Only the two digest tables are
reproduced here; the rest of that record is an internal working document and is
not published with this kit.

**Why this file exists.** A tag is not a model. Every verdict in this kit names
the build it measured, and the manifest digest — not the tag, and not the
reported size — is what identifies that build.

## 1. Staging: every tag re-read and checked before any scored call

Every tag was re-read from the bench store's own `/api/tags` and checked against
the digest this lane had recorded from that same store earlier. A battery that
scores a model it has not identified is measuring an unknown.

| tag | manifest digest | size as the runtime reported it | verdict |
|---|---|---|---|
| `gemma4:31b` | `6316f0629137b426c9d9…` | 19.87 GB | MATCH |
| `qwen3.6:27b` | `9d5803d493a991af27b9…` | 17.77 GB | MATCH |
| `nemotron-3.5-lightning:30b` | `e7a64ff15fb174c42b4f…` | 25.43 GB | MATCH |
| `laguna-xs-2.1:latest` | `0175be1e57f44bb4a967…` | 20.27 GB | MATCH |
| `muse-glimmer:30b` | `de878ce33ad81d060001…` | 18.16 GB | MATCH (fresh pull) |

**5 of 5 verified — safe to score.** The same five digests were re-verified after
the disk stop and before scoring resumed; `battery.log` carries both checks
(`DIGEST VERIFY 5/5 MATCH`).

## 2. The two repro rows against the 2026-08-12 published run

The earlier trials ran on a 96 GB workstation card. The prior question — whether
the repro rows are even the same models — is settled by digest, and it changes
what a repro row means.

| tag, this battery | digest now | tag, 2026-08-12 | digest then | verdict |
|---|---|---|---|---|
| `nemotron-3.5-lightning:30b` | `e7a64ff15fb174c42b4f463e5c888c4f2c7b9cabf9e8d65a1c0874405426c1b2` | `nemotron-3.5-lightning:30b-a3b` | `e7a64ff15fb174c42b4f463e5c888c4f2c7b9cabf9e8d65a1c0874405426c1b2` | **IDENTICAL — same build** |
| `qwen3.6:27b` | `9d5803d493a991af27b9441c098aa56f2ed7bbd260877f075ec09b575c049bc3` | `qwen3.6:27b` | `a50eda8ed977ab48a12431878896b27ffd5cef552c17af3317d9623b939a7f1e` | **DIFFERENT — new build** |

The 2026-08-12 digests are read from the frozen kit's own recorded receipts.

Two consequences for how those rows read:

- **`nemotron-3.5-lightning` is not "a new build of a chaired tag."** The bare
  `:30b` tag and the chaired `:30b-a3b` tag resolve to the same manifest digest,
  byte for byte. A nemotron row is a *harness-and-hardware* reproducibility
  check — same weights, different bench — not a model comparison.
- **`qwen3.6:27b` genuinely is a new build.** Its 2026-08-12 figures were
  produced by different weights under the same tag, so any difference against
  them is a build-to-build comparison and is labelled as one.

## 3. ⚠ An unresolved size-vs-digest anomaly — recorded, not chased

**Read this before quoting any size in this file.** Two tags show the same
oddity: an **identical manifest digest reporting a different `size`** across
readings, purely across runtime versions.

- `nemotron-3.5-lightning:30b` reads 25,430,749,387 bytes on both runtimes used
  in this battery's boxes (ollama 0.32.13 and 0.32.15) but 25,336,532,172 in the
  2026-08-12 census — a **94 MB** gap.
- `muse-glimmer:30b` is the louder case: its digest matches the 2026-08-12
  record exactly, yet reads 18.16 GB against 16,750,089,337 bytes (16.75 GB) in
  that record — an **8 % gap on byte-identical manifests**.

Because the two current runtimes agree with each other and only the older
reading differs, the most likely explanation is that the runtime's `size`
summation changed between versions, not that any weights changed. **The manifest
digest is the authoritative identity and it is stable in both cases.** The bench
recorded this as a standing flag to resolve before anything is published that
quotes a model size, and it is unresolved as this kit ships.

## 4. What a fit receipt's `size_mib` is, and is not

`fit-<tag>.json` reports `size_bytes` / `size_mib` as the runtime reported them
**with a 32,768-token context already loaded**. That is a *loaded footprint*, not
a shipped weight file, and the two are different numbers — nemotron's loaded
total of 24,434 MiB is `size_vram_bytes` plus the 4,136 MiB it spilled to CPU.
The shipped weight figure is the tag listing's own size, in the table above.