# FARTBench Core R01 — human instrument V3

**Instrument version:** `fartbench-core-r01-human-v3`
**Supersedes for collection:** `fartbench-core-r01-human-v2`. V1 and V2 are
preserved exactly under their own tags.
**What changed:** the words the reviewer reads for the three ratings. Nothing
else.

## 1. Why

The factual gates worked. The rating labels did not: "Acoustic Plausibility",
"Temporal / Causal Coherence" and "Physical / Social Plausibility", with anchors
written in the same register, asked the reviewer to translate benchmark
terminology before answering. Across 63 clips that translation is a real cost,
and a source of inconsistency. V3 asks the same three things in plain English, so
they can be answered almost instinctively.

## 2. The three ratings

Identical in every scene. Only the factual gates differ between scenes.

### How real did the fart sound?

| 1 | 2 | 3 | 4 | 5 |
|---|---|---|---|---|
| Obviously fake | Pretty fake | Believable but imperfect | Very believable | Completely real |

N/A: **No fart sound to rate**

### Did the timing make sense?

*This means whether the sound/action/reaction happened in a believable order at
the right moment.*

| 1 | 2 | 3 | 4 | 5 |
|---|---|---|---|---|
| Way off | Mostly off | Mostly right | Very good | Exactly right |

N/A: **No event to time**

### Did the people and movement feel natural?

| 1 | 2 | 3 | 4 | 5 |
|---|---|---|---|---|
| Very unnatural | Mostly unnatural | Mostly natural | Very natural | Completely natural |

N/A: **Nothing relevant to judge**

Every point on the scale now has its own words, not just 1, 3 and 5, and each
button shows its words beside its number.

## 3. What stays exactly as it was

| | |
|---|---|
| Schema field names | `acoustic_plausibility`, `temporal_causal_coherence`, `physical_social_plausibility` |
| Record shape | identical to V2; the V3 schema differs only in its identity fields |
| Factual gates | word for word as V1 and V2 |
| Not-applicable state | V2's reasons and V2's entire dependency map |
| Validation | V2's rules, delegated rather than copied, so they cannot drift |
| Keyboard | `Y` `N` `U`, `1`–`5`, `A` for N/A, `Tab`, `Enter`, `Space` |
| 1–5 scale | unchanged, stored as the number the reviewer pressed |

When N/A is offered, when it is suggested, and when it is refused are all exactly
as V2 accepted them. Only the N/A labels on the three ratings change wording.

## 4. Where the old names still live

"Acoustic Plausibility", "Temporal / Causal Coherence" and "Physical / Social
Plausibility" remain the internal and reporting names, and the schema fields keep
their technical names. None of them appears anywhere the reviewer can read: not
in the page, not in a label, not in a tooltip. A test enforces that against
every reviewer-facing string, in every scene.

## 5. Continuity for analysis

Each V3 rating asks the same question as its V2 counterpart, in different words.
The scale direction is the same, the N/A semantics are the same, and the field is
the same. The owner's analysis rules from V2 carry over unchanged:

- a primary-gate `No` followed by event-dependent N/A means the event was
  effectively absent or not judgeable;
- a primary-gate `No` followed by numeric ratings means an event exists but fails
  the exact-one or clarity requirement;
- `Unclear` remains uncertainty and is never rewritten as absence.