1 · Objective
Test synchronized dialogue, a single intervening flatulatory event, and speech resumption without event duplication.
- Location
- quiet private office
- Subjects
- one adult seated at a desk facing camera
- Target actor
- adult man wearing a gray sweater
- Role
- speaker and sole event source
- Lighting
- soft neutral office lighting
2 · Canonical semantic specification
The canonical specification is model-independent and outranks every rendered prompt. If a rendered prompt changes the meaning of this specification, the prompt is invalid.
Generation condition
- Input mode
- text-to-audio-video
- Duration
- 8.0 s
- Aspect ratio
- 16:9
- Visual style
- photorealistic realistic live action
- Shot structure
- single continuous shot
- Reference media
- none
- Music
- none
- Rerolls
- Prohibited
Audio condition
- Fart audible
- Yes
- Fart count
- 1
- Fart character
- short, natural, moderate
- Dialogue exact
- The quarterly numbers look better than expected.
- Dialogue structure
- fart interrupts a natural pause after the word better; speech resumes afterward
- Ambient
- quiet office room tone
- Music
- none
Camera
- Framing
- medium shot
- Movement
- static
- Target visible
- Yes
- Cuts
- 0
Prohibited
- visible gas or smoke
- defecation
- additional fart events
- additional dialogue
- music
- large body deformation
3 · Primary evaluated dimensions
6 of 7 dimensions apply to this test. FARTScore renormalizes the base weights over the applicable set, so a non-applicable dimension is excluded rather than scored as zero.
| Code | Dimension | Base weight | Renormalized | Applicability |
|---|---|---|---|---|
| A | Acoustic Condition Fidelity | 22% | 24.4% | Applicable |
| T | Event / AV Temporal Alignment | 16% | 17.8% | Applicable |
| L | Source Localization | 16% | 17.8% | Applicable |
| P | Physical / Body-Motion Plausibility | 14% | 15.6% | Applicable |
| I | Instruction Adherence | 12% | 13.3% | Applicable |
| S | Social-Causal Response | 10% | — | Not applicable |
| C | Temporal / Identity Coherence | 10% | 11.1% | Applicable |
4 · Required event structure
- 0.0–2.0 s
target sits naturally and begins speaking the exact sentence
- 2.0–5.0 s
he says “The quarterly numbers look better”, pauses, produces exactly one short fart during the pause, then resumes
- 5.0–8.0 s
he completes “than expected.” and briefly maintains composure
Objective checks
Raters answer these observable questions before recording perceptual scores, so that an overall impression cannot overwrite a factual failure.
| Field | Expected |
|---|---|
| event_present | true |
| event_count | 1 |
| correct_actor | true |
| dialogue_present | true |
| dialogue_words_correct | true |
| fart_interrupts_between_required_phrases | true |
| visible_gas | false |
| defecation | false |
5 · Prompt adapter policy
A versioned model adapter may reorganize the canonical facts into the form its creator documents. It may not make the task easier, add creative facts, remove a hard requirement, change the designated actor, change the event count, materially widen the timing window, or respond to observed benchmark failures. Prompt bytes are hashed and immutable once a benchmark generation has been viewed.
Rendered prompts not yet commissioned for this test
Adapter renderings are frozen one test at a time. FB-C02 is the current commissioning wave; FB-C03 renderings are produced after the nine-clip smoke test is reviewed.
Published when the FB-C02 smoke test is accepted and the development pilot expands to all eight tests.
6 · Known failure modes
Failure codes of interest for this test, recorded as multi-label flags. A clip may carry several.
- F02 Phantom Flatulence Major
- F03 Premature Acoustic Event Major
- F04 Acoustic Category Failure Major
- F05 Anatomical Deformation Critical
- F09 Unrequested Fart Duplication Major
- F10 Unrequested Defecation Critical
- F11 Event Omission Critical
- F12 Wrong Event Count Major
- F14 Macroscopic Gas Visualization Major
- F15 Temporal Fart Drift Major
7 · Results by configuration
No locked evaluations for this test
23 Core AV configurations are commissioned against the current prompt pack. None has completed blind human evaluation under Core v0.1.0, so no per-configuration result exists to publish.
Published when at least three independent raters have locked evaluations for a configuration's 3 replicates.
Development pilot target: 8 tests × 3 models × 3 replicates = 72 clips. All pilot data remains development data; the Core v1.0 freeze happens only after that pilot is analysed without reusing its scores.
8 · Representative generations
No public clips
Raw benchmark outputs are immutable evidence and are not published before the evaluations that reference them are locked. Explicit output is retained privately as evidence and omitted from any public gallery.
Published when the first evaluation round is locked and a playback derivative with its own hash is produced for each published clip.
9 · Version and provenance
- Test ID
- FB-C03
- Specification
- v0.1.0
- Protocol
- Core v0.1.0
- Status
- Pilot development
- Frozen
- 2026-09-09
- Replicates
- 3 (pilot)
- Seeds
- 104729 · 130363 · 155921
- Spec SHA-256
- 5ff56097fc1e6c5d9069255eaa300cf1aa4c95b0d0d18c83dc783387e04871e1
10 · Machine-readable specification
The YAML above is the canonical artifact this page is generated from — the page cannot drift from the file it offers.