# Where the 2026-09-09 model snapshot actually lives in the tree

**182.C.34.10.** The item says: use benchmark evidence to choose defaults, and
do not enshrine the 2026-09-09 snapshot as architecture. "Do not enshrine" is
unfalsifiable until you name what enshrinement looks like. It takes four shapes
here, and exactly one of them is visible to a grep:

1. **A model id above the adapter layer** — measurable, and measured below.
2. **A capability assumed rather than read from the live catalog** — not
   greppable; the rule is enforced at runtime in `human-video-lane-adapters.ts`.
3. **A threshold fitted to one model's output distribution** — not greppable,
   and the one nobody sees: a sync window tuned on Wan's takes is a Wan constant
   wearing a physics name.
4. **A default with no benchmark behind it, or one from a year ago** — not
   greppable; handled by `defaultStatus` in
   `apps/oshun/bff/src/generation/human-video-snapshot-boundary.ts`.

So this evidence covers form 1 only, and says so.

## Method

`measure.sh`, run from the repository root on **2026-09-11**. It greps `apps/`
and `libs/` for the five model ids named by the phase (`wan/wan-2*`,
`bytedance/seedance*`, `google/veo-3*`, `heygen/avatar-iv`), excluding
`.spec.ts` and `.test.ts`, and splits each file's hits into code and
comment/prose. `results.txt` is its output.

## Result

| measure                                                         | value |
| --------------------------------------------------------------- | ----- |
| source files naming a model id                                  | 15    |
| of those, in `apps/oshun/web` or `apps/oshun/admin`             | **0** |
| occurrences in `libs/contracts`, all inside description strings | 2     |
| of those, carrying the date they were true                      | 1     |
| of those, carrying no date                                      | **1** |

The fifteen files are the adapter table, the provider-slug table
(`human-video-provider-endpoints.ts`), the provider-constraints table, the
Seedance direct client, one contract vocabulary file, and five files where the
id appears only in a comment. Every one of those except the contract file is a
layer where a model id **is the data** — a table of ids is what the adapter
layer is for, and deleting a model means deleting a row.

## The one finding

`libs/contracts/src/human-video/capability-vocabulary.ts` carries two model ids
in description strings, and they are not the same case:

- line 355 — `Null for 3 models on 2026-09-10, including heygen/avatar-iv` —
  **dated**. A reader in 2027 knows what it is.
- line 369 — `google/veo-3.1 also exposes aspectRatio as a passthrough` —
  **undated**. Nothing in it goes stale.

The second is a small live instance of form 1: a statement about one model, in
the shared contract, that will read as current forever. It is recorded here
rather than quietly edited, because the census is the thing the rule gets
checked against later, and a census adjusted to look clean measures nothing.

## What this does not show

Forms 2, 3 and 4 are invisible to this method. A capability hardcoded three
call-sites deep, a threshold fitted against one model's distribution, and a
default nobody benchmarked all pass this grep. The zero in the web/admin row is
a real result about one shape of the problem and not a statement about the
others.

## Reproducing

```bash
bash docs/domains/isis/human-video/evidence/2026-09-11-snapshot-boundary/measure.sh
```

One caveat the first run found: `grep -rl` exits 1 when it matches nothing, and
under `set -o pipefail` that ends the script exactly when the answer is the good
one — the web/admin check printed `0` and then killed the run before the last
section. The pipeline is wrapped in `|| true` for that reason.
