Every track that comes back from generation must go through this skill. This is the final check of this library.
Core thesis: AI-tell is not mysticism. Generative models fall back on defaults when no spec is given, and these defaults are enumerable. This skill is that enumeration, plus "how to check each one and who owns it."
Which sections to read by task
| Task | Read |
|---|---|
| Just finished generating, need acceptance | §2 Fourteen items → §3 Verdict |
| "Sounds like AI but can't say where" | §2（★ Check item by item, don't rely on intuition） |
| High compliance rate but poor results | §4 |
| Need to set thresholds | §3.2 Zero-point calibration |
| Found problems, need to fix | §5 Remediation paths |
| How to check the four vocal items | §6 |
Boundary
| Not handled here | Where it goes |
|---|---|
| SpecWrittencheck at that time | mc-workflow §3.0's 20-item self-check table |
| How to calculate compliance rate | This skill §3.1 (definition and calibration); for how the three honors tiers affect scoring see mc-render-compile §6 |
| How specifically to fix | Each L1/L2 skill (§5 gives routing) |
| What to do about fields the backend can't do | mc-render-compile §6、§7 |
★ Division of labor with the 20-item self-check： The self-check asks "is the spec written," this skill asks "does the generated output look like it was made by a human." Only when both have been checked is it done. A spec that passes all 17 items can still generate audio full of AI-tell.
1. Must fill in before starting
| Required | Wording |
|---|---|
| ① This version's spec and seed | No record means no reproducibility,and the audit result is useless |
| ② The backend's honors table | Which items it can't do——those don't count against it |
| ③ Are there stems | The four vocal items (§6) require separating the vocal track first |
2. ★ Fourteen-item enumeration
Check off item by item. Don't rely on overall impression.
| # | AI-tell source (backend default) | Spec field to counter it | Auto-measurable | Who owns it |
|---|---|---|---|---|
| 1 | Sections are all integer multiples of 8 | form[].bars |
✅ Section boundary detection | mc-arrangement-arch §4 |
| 2 | Energy curve monotonically increasing, never decreasing | energy_curve、subtraction_events |
✅ Segmented RMS | mc-arrangement-arch §2、§3 |
| 3 | Instrumentation unchanged from start to finish | roster[].entry/exit |
✅ Stem onset density | mc-arrangement-arch §5 |
| 4 | One harmonic loop for the whole song | bridge must modulate or key-change | ✅ Chord recognition | mc-harmony、mc-modulation |
| 5 | Fills stereotypically land on multiples of 4 | fill_policy |
✅ onset peak position | mc-rhythm-section §8 |
| 6 | No non-vocal hook | hooks.arrangement_hook |
⚠️ Semi-automatic (repeated section similarity) | mc-arrangement-arch §7 |
| 7 | Dynamics crushed, spectral centroid constant | dynamics、width_map |
✅ DR/LRA, centroid temporal variance | mc-mix-intent |
| 8 | Gridded, no push/pull | groove.push_pull |
✅ onset offset relative to grid | mc-rhythm-section §3 |
| 9 | Duration always 3:30, no tail | target_duration + ending treatment |
✅ | mc-arrangement-arch §8 |
| 10 | Every sentence has the same dynamics and articulation, no breathiness, no vocal break, no rubato | vocal.delivery、dynamics_by_section |
⚠️ Requires stems to measure short-term loudness variance | mc-vocal-direction §2、§4 |
| 11 | The chorus is sung exactly the same all three times | vocal.chorus_variation |
✅ Vocal track segment similarity | mc-vocal-direction §1④ |
| 12 | Breath points are unnatural or breaths are inaudible | vocal.breath_points |
⚠️ Semi-automatic | mc-vocal-direction |
| 13 | Ad-libs scattered evenly or simply absent | vocal.ad_libs |
⚠️ Semi-automatic | mc-vocal-direction |
| 14 | Pitch accuracy is perfect | vocal.pitch_policy |
⚠️ Semi-automatic (measure pitch offset distribution after stemming) | mc-vocal-direction §4 |
★ This table serves one purpose in three ways: the skeleton of this skill, the skeleton of the AB experiment evaluation table, and the source of the 20 self-check items.Change one place and you must sync three places.
2.1 ★ The three most easily missed and also most fatal
Not because they are the hardest to check, but because they are not unpleasant to hear—they just "don't sound human-made."
| # | Why fatal |
|---|---|
| 2 (energy only rises, never falls) | Measured 62/62 Real arrangements all have drops.This is the strong rule with the highest hit rate, and also the one that sounds most like AI |
| 8 (everything aligned to the grid) | Measured 62/62. Note the reading is "not every part in the whole song may be on the grid"—absolute quantization of the drum kit is allowed (→ mc-rhythm-section §2.0） |
| 6 (no non-vocal hook) | Its symptom is "sounds okay but not memorable,"easiest to pass all other checks. A song missing it is forgotten right after listening |
3. Two numbers, never merge them
3.1 Definitions
| What it asks | How to obtain | Range | |
|---|---|---|---|
| Compliance rate | How much the backend followed the spec | Field-by-field comparison against the spec | 0–100 |
| AI-flavor flag count | How many machine fingerprints the final product has | Check off item by item per §2 of this skill | 0–14 |
★ Absolutely never combine them into a single score. They diagnose different diseases:
照做率高 + 旗标少 → 好，接受
照做率高 + 旗标多 → ★ 规格本身写得像 AI（见 §4）
照做率低 + 旗标少 → 后端自己干得不错，但你不可复现（见 §4.2）
照做率低 + 旗标多 → 编译或后端选择有问题，回 mc-render-compile
3.2 ★ Zero-point calibration: why the threshold is 82, not 60
Unrelated, well-produced finished tracks score 63–66 on the compliance rate table.
This is because any normal song will "accidentally" satisfy some items (has tempo, has sections, has dynamics). So:
| Line | Value |
|---|---|
| Zero point(unrelated finished tracks) | 63–66 |
| Pass | 70 |
| Qualified | 82 |
★ The above three numbers arereference values: the zero point was measured by taking a City Pop spec (112 BPM, 10 sections, 114 bars) and comparing it with three well-produced Chinese-language finished tracks unrelated to it, scoring by weighted fields (63.4–66.4; the same spec compared with a pure MIDI-rendered demo scored only 39.5). Used to understand magnitude, not a delivery threshold; when comparing field by field, the compliance rate is just one judgment after comparison.
★★ A pitfall I've stepped into: moving weight fromdiscriminativeitems (boundaries) to
non-discriminativeitems (energy_shape, irrelevant tracks can also get 0.8),
willraise everyone's score, including the zero baseline (66→72).
Weighting is important; put it on discriminative items.
3.3 What the backend can't do isn't its fault
honors: none 's fieldsare excluded from scoring, not given 0.
See mc-render-compile §6.1——This is a correctness issue, not a criterion issue。
4. ★ High compliance rate but still looks like AI
This is the main reason this skill exists.
The compliance rate only measures "whether the backend complied,"not "whether the spec itself is good". If the spec is written like AI and the backend faithfully executes it, the result is an AI song with a 95 compliance rate.
4.1 Four signs a spec looks like AI
| Sign | Where to check |
|---|---|
intent.one_thing isfilled in just to pass lint, and doesn't read like a human sentence |
mc-workflow S0 of §1 |
reference_pair 's borrow is written very vaguely ("borrow its vibe") |
Same as above."Only borrow the emptiness of the intro" is what counts as written |
Each section's bars chosen just to avoid multiples of 8odd number |
mc-arrangement-arch §4——Asymmetry needs a reason |
| The progression / modulation used hitscliché warning lineand no reason is given | mc-progressions §3、mc-modulation §6 |
★ Common point：Rules were treated as conditions to satisfy, not choices to make. This is exactly the definition of an AI flavor—no one made a choice。
4.2 Low compliance rate but good to listen to
means the backend improvised well on its own. You can accept this result, but you must realize:
- Youcan't reproduceit
- Youdon't knowwhich decision made it good
- the next one has to rely on luck from scratch
★ Handling method: reverse-engineer this result field by field into a measured ARR-SPEC (fill it in while listening, mark uncertain ones as "estimate"), and see where it differs from your original spec.That difference is what you learned.
5. Remediation path
Follow this order; don't jump by intuition.
听着"像 AI"
└─ ① 逐字段对照规格，拿照做率
└─ ② 逐条过 §2 的十四条，数出旗标数
├─ 照做率低 → 查 honors
│ ├─ 这个字段该后端是 none → 不是它的错。换后端（YuE2 控制力最强）或接受
│ └─ 不是 none → 编译问题 → mc-render-compile §3
└─ 照做率高但旗标多 → ★ 规格本身的问题 → §4.1
5.1 Symptom → skill (quick reference)
| Symptom | Check first |
|---|---|
| flat, no ups and downs | mc-arrangement-arch §2 |
| muddy, murky | mc-texture-layering §4（★ Check sustained notes first, not the drums first） |
| The protagonist isn't prominent | mc-texture-layering §3's four deep clues |
| mechanical, too precise | mc-rhythm-section §2 ＋ mc-vocal-direction §4 |
| Sounds okay but can't be remembered | mc-arrangement-arch §7（arrangement_hook is it empty) |
| The second chorus is the same as the first | mc-arrangement-arch §6 ＋ mc-vocal-direction §1④ |
| Vocals buried in the accompaniment | Check pitch first（mc-vocal-direction §4.1), then check the frequency band |
| Bass disappears on mobile | mc-sound-design §6 difference tones |
| The ending is abrupt / fades out to be done with it | mc-arrangement-arch §8; ★ Check exclude whether it excludes fade out ending |
| technically clean but emotionally off | mc-mix-intent §3.1（mood was not written) |
| harmony loops the same all the way through | mc-harmony §7、mc-modulation |
6. Four vocal items (#10, #12, #13, #14)
All four require first separating the vocal track.
6.1 Stem separation
Any stem separation tool can be used to separate the vocal track (demucs and the like, or the stem separation that comes with the user-generated carrier, usually split into vocals/bass/drums/other). This skill does not prescribe which one.
★ This is one of the few cases in this library where stem separation is necessary: Multi-track corpora (Cambridge-MT) cannot provide "vocals as sung by Suno"; they can only be separated from the generated result.
6.2 What to measure after stem separation
| # | What to measure | Criterion |
|---|---|---|
| 10 dynamics | The vocal track'sshort-term loudness variance | variance near 0 = one dynamic level throughout |
| 11 choruses identical | Each chorus vocal track'ssegment similarity | similarity too high = copied verbatim |
| 12 breath | low-energy segments between phrases | ⚠️ Semi-automatic; currently relies on listening |
| 13 ad-lib | distribution of ornamental vocalizations | ⚠️ Semi-automatic.even = sprinkled in, not sung |
| 14 pitch accuracy | distribution of pitch deviation | ⚠️ Semi-automatic.all falling within ±5 cents = over-corrected/generated |
6.3 ⚠ Current status
#12, #13, #14 rely on human listening; this is a deliberate choice, not a missed step.
These three measure the energy troughs of breaths, the distribution of ornaments, and the distribution of pitch deviation; their automatic criteria all require thresholds calibrated per generative model, and baselines differ greatly across models, so forcing a single uniform number easily flags good singing as problematic. Listening through these three once is quick anyway, and it is exactly where the ear is naturally sharper than a machine, so it is not an extra burden.
7. Acceptance checklist
Every generated result must pass.
- recordedspec version and seed
- went through the spec field by field, with a clear sense of the compliance rate
- item by itempassed the fourteen items of §2,counted the number of flags(not by impression)
-
honors: noneitemsexcluded from scoring, not given 0 - if compliance rate is high but there are many flags, checked the four signs of §4.1
- the four vocal items at leastlistened through manually
- ★ The ears intervene last: shuffle A/B → pairwise comparison → listen to the same pair twice to see if the conclusion is stable
- conclusions inconsistent across two listens → this difference does not exist, don't change it
Appendix: Sources
- Measured data (62/62 in §2.1, the 22% and 5/7 below): statistics on 62 naturally sampled tracks from the Cambridge-MT multi-track set; the 5/7 are the ground truth where 7 of them had segment boundaries manually annotated
- Zero points 63–66: see §3.2
- All fourteen items of this skill can be judged manually
- Among the fourteen items, 1–5, 7–9, and 11 have audio analysis, so they can be computed by machine,6, 10, 12, 13, and 14 are manual items
★ A methodological reminder(pitfalls I've stepped into): When a metric measures inaccurately,don't use the metric to overturn the textbook; first ask whether the metric is measuring that very thing. Example: "There should be an energy falloff before the climax" holds in only 22% of all 62 tracks, but among the 7 manually annotated ground-truth tracks it is 5/7— the gap comes from the algorithm using energy peaks to locate the "climax," whereas the human ear does not. That 22% figure is not measuring this rule.