Acquisition risk memo
Prepared for the acquiring party · 16 assets reviewed
$171,710
6 of 16 tracks (37.5%) are suspect, carrying $8,365 of $48,000 annual revenue (17.4%). Recommended escrow $171,710 against an asking price of $720,000.
Suspect revenue of $8,365 held in full plus 50% of contested revenue of $6,166, capitalised at the 15.0x acquisition multiple.
| Title | Artist | AI score | Annual | Tier |
|---|---|---|---|---|
| Sleigh Ride | Kamakaze Kitten | 1 | $12,175 | clean |
| Pour Quoi Royale | Zipp | 0 | $5,013 | clean |
| sonics__fake_00105_suno_1 | Generated | 100 | $4,974 | suspect Suno |
| Sleigh Ride | Bubblyfish | 2 | $4,820 | clean |
| Reading 2 Volume 3 | Lemoness | 5 | $3,551 | contested |
| Footloose (feat. Paul Whiteman) | Dog On Springs | 0 | $3,074 | clean |
| 01_hiliv003_onsundaysidream_hypnagogic | mccandless | 1 | $2,727 | clean |
| Aydio_-_01_-_Track | Aydio | 5 | $2,614 | contested |
| Dance With The Democrat | Boogie Belgique | 0 | $2,105 | clean |
| The Night Before Bit Mess | c-c | 0 | $1,866 | clean |
| sonics__fake_00214_udio_0 | Generated | 100 | $1,698 | suspect |
| Animosity | Intoxicated Piano | 0 | $1,690 | clean |
| sonics__fake_00001_suno_0 | Generated | 100 | $517 | suspect |
| sonics__fake_00524_udio_1 | Generated | 71 | $484 | suspect |
| sonics__fake_00416_udio_1 | Generated | 100 | $357 | suspect |
| sonics__fake_00327_udio_0 | Generated | 100 | $335 | suspect |
These could not be settled from audio alone. They are excluded from the clean base and only partially weighted in the escrow figure, and they are listed by what they earn so a reviewer with limited time starts where the money is.
| Asset | Score | Annual | Why |
|---|---|---|---|
| 2014-02-12Bigwords__Reading2Volume3.mp3 | 5 | $3,551 | Split verdict — press-safe human, human-safe human, recall ai. Borderline by the detector's own account. Risk peaks at 100% from 0:20. similarity map |
| DWK031__Aydio_-_01_-_Track_64kb.mp3 | 5 | $2,614 | Split verdict — press-safe human, human-safe human, recall ai. Borderline by the detector's own account. Risk peaks at 68% from 0:33. similarity map |
Against 16 labelled tracks: 6 of 6 planted AI tracks were caught, 0 missed, and 0 landed in the contested band. No human track was wrongly flagged as suspect. Precision 1.00. Recall 1.00. Of 6 correctly flagged AI tracks whose true generator is known, 1 was attributed to the right one, 5 carried no attribution. Only 6 AI tracks were planted, so these are counts rather than measured rates — the confidence interval on a sample this size is wide enough that the figures should be read as indicative.
Knowing a recording is synthetic is one claim. Knowing which model made it is a second, harder one, and it is the claim a seller will be asked to answer. Scored only on assets correctly identified as AI whose true generator is recorded.
A declined attribution is counted separately from a wrong one. The detector saying "I cannot place this" is the behaviour this audit is built to reward, not an error to be penalised.
This catalog was constructed with known labels so the error rate could be measured. A real acquisition has no answer key, which is why the contested band exists.
In July 2026 IFPI, RIAA, A2IM, WIN and IMPALA adopted a unified track-level standard distinguishing AI-Generated from AI-Assisted recordings. HumanStandard reports each verdict against those definitions, so this is the label the catalog would carry on a platform that adopts the standard — not a private score.
The labelling programme is voluntary and self-declared. A seller who has not declared these is not necessarily concealing them, but the buyer inherits the declaration either way.
HumanStandard places every recording against reference populations — verified human music, and the known generators (Suno, Udio, ElevenLabs, Lyria, Treblo). Where a track sits clearly inside one of those populations, it is named. 1 asset sit nearest a generator rather than the human set.
| Nearest population | Tracks | Annual | Basis |
|---|---|---|---|
| Suno | 1 | $4,974 | 23 of its 25 nearest reference recordings are Suno generations. |
| Verified human recordings | 11 | $40,120 | 24 of its 25 nearest reference recordings are verified human recordings. |
Attribution is neighbourhood evidence, not the verdict itself. It tells a buyer which supplier the material resembles — which is the question a seller will be asked to answer.
Every asset was submitted to the HumanStandard detection API, which returns a three-valued verdict — AI, human, or uncertain — at three calibrated operating points with their false-positive rates published. Assets are tiered on those operating points rather than on thresholds chosen here: suspect when HumanStandard's human-safe tier (~1-2% FPR) says ai | clean when its recall tier (~5-10% FPR) says human | contested when the tiers disagree. Where the operating points disagree, the recording is borderline by the detector's own account and is routed to a person. Anything below the detector's confidence floor is routed to review regardless of verdict, and anything that could not be scored is excluded from the clean base rather than assumed safe. Suspect revenue is escrowed in full; contested revenue is weighted at 50% because uncertainty is not a finding.
