# Loudness conformance, and what it hides (2026-09-10)

Evidence for `182.C.18.05`: "Verify loudness and true peak against delivery
profiles **without hiding provider speech defects through aggressive
processing**."

The second half of that sentence is the whole item, and it is measurable.

## Method

The twelve `182.C.18.04` fixtures
(`tools/human-video/evaluator/make_defect_fixtures.py`) measured with
`ffmpeg -af ebur128=peak=true`, then passed through
`loudnorm=I=-16:TP=-1:LRA=11` — the repo's `DEFAULT_DELIVERY_PROFILE` target —
and measured again, both for loudness and with
`tools/human-video/evaluator/audio_defects.py`.

## The finding

| Clip        | As received (I / LRA / TP / crest / clip run / SNR) | After loudnorm                           |
| ----------- | --------------------------------------------------- | ---------------------------------------- |
| clean       | −23.9 / 1.9 / −5.4 / 20.8 / 0 / 81.3                | −16.2 / 1.6 / −1.0 / 17.3 / 0 / 89.2     |
| **clipped** | **−9.5 / 1.9 / +0.9 / 11.6 / 15 / 96.0**            | **−14.9 / 2.4 / −3.2 / 13.2 / 0 / 90.5** |
| noisy       | −23.7 / 1.0 / −6.2 / 18.4 / 0 / 11.9                | −16.3 / 1.0 / −0.9 / 16.5 / 0 / 11.7     |
| very-noisy  | −14.9 / 0.1 / −3.0 / 14.2 / 0 / 3.0                 | −16.1 / 0.2 / −4.1 / 14.3 / 0 / 3.1      |
| pumping     | −33.2 / 2.4 / −7.1 / 26.7 / 0 / 17.4                | −17.6 / 1.6 / −0.7 / 17.6 / 0 / 16.6     |
| duplicated  | −24.4 / 1.3 / −5.4 / 20.7 / 0 / 81.3                | −16.7 / 1.4 / −1.0 / 17.4 / 0 / 89.2     |

**Loudness normalisation erases the evidence of clipping without repairing it.**
The clipped fixture arrives with a 15-sample run of samples pinned at full scale
and a true peak of +0.9 dBTP. After the conform, the clipping detector measures
a longest run of **zero**. The waveform's tops are still cut off — scaling down
does not restore what a converter discarded — but nothing downstream can see it.
A conform step launders a clipped take into one that passes the clipping gate.

**Noise is not hidden the same way.** The very-noisy fixture measured 2.98 dB
estimated SNR before and 3.06 dB after, because scaling moves speech and noise
together. Saying which defects survive matters as much as saying which do not: a
rule that cries wolf about all nine gets turned off.

**Dynamic range is the cost every take pays.** The clean fixture's crest factor
fell 20.8 → 17.3 dB and its LRA 1.9 → 1.6 LU purely from meeting the target, and
the pumping fixture's crest fell 26.7 → 17.6 dB under `loudnorm`'s own limiter.

## What the module does about it

- `assertDefectsMeasuredFirst` refuses a defect report produced after any
  processing step. A report from that point is a report about the conform, not
  about the delivery.
- `LOUDNESS_HIDES` names `clipping` as hidden and the other seven as surviving,
  with the measurement quoted alongside.
- `planLoudness` reports the gain a take needs **and** whether the target is
  reachable by gain alone. Gain is transparent; limiting changes the waveform's
  shape, and the shape is what the detectors read.
- `assertConformableWithoutLimiting` refuses a take that only reaches the target
  through limiting. The profile is a delivery specification, not a licence to
  reshape a bad take until it fits. On these fixtures every clip fails this —
  they are all quiet enough that the gain to reach −16 LUFS breaches the ceiling
  — which is the correct answer: they go back, they do not go into a limiter.
- `assertNotAlreadyOverProcessed` refuses a take that ARRIVES squashed, at a
  provisional 13 dB crest floor. That sits under every clean measurement above
  and over the clipped fixture's 11.6 dB, and under our own conform output's
  17.3 dB so it never rejects what this pipeline produced.

## Per-item status this evidence supports

| Item          | Was           | Is                                                                                |
| ------------- | ------------- | --------------------------------------------------------------------------------- |
| `182.C.18.05` | `not-started` | done; conformance plus the ordering rule that keeps `182.C.18.04` findings honest |
