# Synthetic negative fixtures, with measured ground truth (182.C.33.06)

Generated 2026-09-11 on the Linux dev server, ffmpeg 6.1.1-3ubuntu5. Reproduce
with `bash generate.sh <outdir>`; the machine-readable result is
`manifest.json`. The media itself is **not committed** — it is regenerated,
deterministic, and there is no reason to carry binaries for it.

## Why these exist

The real corpus has no labels. Nobody knows the true audiovisual offset of a
clip a customer complained about, so an evaluator measured only against real
failures cannot be told how sensitive it is. A fixture with a **known** defect
is the only way to ask "does the detector find a 120 ms offset, and does it find
a 40 ms one".

## The finding: every fixture's achieved defect differs from the requested one

| Fixture                      | Asked for                 | Measured (audio − video duration)  | Note                                                            |
| ---------------------------- | ------------------------- | ---------------------------------- | --------------------------------------------------------------- |
| `offset-filter-120ms.mp4`    | 120 ms audio lag          | **+137 ms**                        | `adelay` is sample-accurate; AAC adds ~17 ms of encoder padding |
| `offset-container-120ms.mp4` | 120 ms audio lag          | **+23 ms**, `start_time` **96 ms** | `-itsoffset` is quantised to an AAC frame                       |
| `drift-1pct.mp4`             | audio at 1.01×            | −24 ms over 4 s                    |                                                                 |
| `dropped-300ms.mp4`          | 300 ms removed            | −262 ms                            |                                                                 |
| `inserted-300ms.mp4`         | 300 ms duplicated         | +286 ms                            |                                                                 |
| `silent-mouth.mp4`           | audio replaced by silence | −6 ms                              | the defect is content, not duration                             |
| `frozen-picture.mp4`         | picture frozen            | 0 ms                               | likewise                                                        |
| `stretch-1.05x.mp4`          | both streams at 1.05×     | −54 ms                             |                                                                 |
| `edit-boundary.mp4`          | 0.5 s cut at 2.0 s        | +20 ms                             |                                                                 |

**Two ways to build "a 120 ms offset" produce two different fixtures, and
neither is 120 ms.** A fixture labelled with the value that was _requested_
teaches an evaluator the wrong number, and the error is systematic rather than
random — every AAC fixture carries the same encoder padding in the same
direction. So the manifest records what was measured, and the requested value is
kept only as provenance.

## Two of the nine have no duration signature

`silent-mouth` and `frozen-picture` are defects in _content_: the durations
match and the file is wrong anyway. An evaluator that checks stream lengths
finds seven of these nine and reports a clean bill of health for the two where a
person is visibly speaking silence.

## What is not here

Real speech. These are a test pattern and a sine tone, which is enough to
measure an offset detector and not enough to measure an aligner — word-level
defects (`dropped-speech`, `inserted-speech`) are _durations_ here rather than
words. Labelled speech needs the consented corpus of `182.C.33.01`.

## Addendum, 2026-09-11 — the two timing fixtures (182.C.35.06)

`generate-timing.sh` builds the two kinds the original nine did not cover, and
`timing-manifest.json` is its measured output. Both are synthesisable, which is
exactly why they belong here and why wrong-speaker, face-loss and identity-swap
do not: those are content signatures of a real person.

| fixture                 | duration       | r_frame_rate | avg_frame_rate | frames | bytes   |
| ----------------------- | -------------- | ------------ | -------------- | ------ | ------- |
| `cfr-source.mp4`        | 4.000 s        | 25/1         | 25/1           | 100    | 400,672 |
| `vfr.mp4`               | 3.040 s        | **50/1**     | **625/19**     | 100    | 320,789 |
| `vfr-attempt-vsync.mp4` | 3.080 s        | 25/1         | 25/1           | 77     | 284,340 |
| `reencoded.mp4`         | **4.017007 s** | 25/1         | 25/1           | 100    | 118,731 |

Three findings, all measured rather than reasoned:

1. **`-vsync vfr` does not produce a variable frame rate.** It drops frames to
   fit a constant one: 77 frames in 3.08 s, reported as an ordinary 25/1 file. A
   fixture suite built on that recipe would report "VFR handled correctly"
   forever. `-fps_mode passthrough` keeps every frame with its modified
   timestamp and produces the real thing — 100 frames in 3.04 s, `r_frame_rate`
   50/1 against `avg_frame_rate` 625/19. Both are built, and the failed one is
   asserted **not** to be variable, so the difference is a test rather than a
   comment. Same shape as this directory's `-itsoffset` finding, on another
   flag.

2. **The asset validator cannot see a variable frame rate.**
   `VideoStreamFacts.frameRate` is read from `avg_frame_rate`, and
   `r_frame_rate` is not in the ffprobe shape at all. The VFR file averages
   ~32.9 fps, comfortably inside the 12–120 policy, and is admitted cleanly.
   Reading `r_frame_rate` instead would not help either — 50 fps is also inside
   the policy. The gap is not the range; it is that the probe carries nothing
   that says _variable_.

3. **A lossy re-encode moves the duration by 17.007 ms** (4.000 s → 4.017007 s,
   same 100 frames). Against the 45 ms ATSC sync window that is **38% of the
   budget**, spent by an operation nobody thinks of as a timing change, before a
   single frame of drift.

### Reproducing

```bash
bash docs/domains/isis/human-video/evidence/2026-09-11-negative-fixtures/generate-timing.sh /tmp/fx
```

The `.mp4` files are not committed, for the same reason the original nine are
not: the manifest records the measured facts, and the script rebuilds the bytes.
