Public craft-metrics report
The Real Data Behind SongForgeAI
Every song forged by the pipeline logs its composite score, prosody warnings, detail balance, emotional arc shape, and fire-line peaks into forge_metrics. This page is the aggregate — updated hourly, suppressed below 30-sample cells, no user-identifying data.
Report generated Tue, 21 Jul 2026 12:36:48 GMT. Window: 5/3/2026 to 7/21/2026.
Corpus summary
Songs forged
1,000
Avg composite
81.1
Median composite
83.0
Arc variety
98% non-flat
Composite score distribution
0-49
0.2%
50-59
0.1%
60-69
0.8%
70-79
36.4%
80-89
62.3%
90+
0.2%
Emotional arc shape distribution
flat
2%
single-axis
61.7%
shaped
35.5%
volatile
0.7%
By genre
Each slice shows at least 30 songs. Genres below that floor are suppressed to protect small-cell identity.
| Genre | Songs | Avg composite | Prosody/song | Fire-line peak |
|---|---|---|---|---|
| folk | 164 | 83.2 | 6.73 | 64.1 |
| alternative | 116 | 79.7 | 6.69 | 62.6 |
| contemporary | 92 | 82.2 | 6.46 | 64.6 |
| indie | 64 | 82.6 | 6.86 | 62.8 |
| 1990s | 53 | 82.1 | 7.36 | 61.6 |
| christian | 42 | 79.1 | 7.26 | 63.5 |
| pop | 37 | 77.5 | 7.32 | 63.7 |
| country | 32 | 81.5 | 8.13 | 64.0 |
By quarter
| Period | Songs | Avg composite | Prosody/song | % gauntlet improved | % enriched |
|---|---|---|---|---|---|
| 2026-Q2 | 964 | 81.3 | 6.67 | — | 41.3% |
| 2026-Q3 | 36 | 75.2 | 12.64 | — | 83.3% |
Glossary
- composite score
- 0-100 output of the 12-metric rubric at /scoring/standard. Eight-voice panel with anti-inflation rules (gravity, burden of proof, antagonist ceiling). 50 = the gravity default; 80+ is genuinely strong; 90+ is rare territory.
- prosody warnings
- Count of Pattison-rule violations per song from the deterministic singability scanner. Fewer is better. Catches stress mismatches, weak hook-word placement, vowel-on-low-note-with-cluster, etc.
- arc shape
- Classification of the song’s emotional contour, computed per-section from valence + arousal:
- flat — valence and arousal both sit in a narrow band the whole song. Reads as monotone; usually a wound, not a stylistic choice.
- single-axis — one of valence or arousal moves meaningfully but the other is flat. The song goes dark or goes loud, but not both.
- shaped — both valence and arousal change section to section in a coherent direction (rising, falling, or rising-then-falling). The intended healthy shape.
- volatile — both axes move but the direction is incoherent: a section jumps brighter then darker then brighter again with no accumulating arc. Reads as confused emotional intent.
- fire-line peak
- Highest line-level Ritter memetic score across the lyric. Tries to surface the most quotable / shareable single line.
- % gauntlet improved
- Of songs the gauntlet ran on (mode=‘gauntlet’ rows), fraction whose composite score went UP after the gauntlet rewrite. Build 1183 fix: the prior aggregator counted initial forges in the denominator, biasing this toward 0%. The current number is “of gauntlet attempts, what fraction lifted the score.” Reverted gauntlets aren’t logged to this table at all (separate code path), so they don’t pull this down.
- % enriched
- Fraction of forges that had the SuperPrompt enrichment step run. Caveat: as of Build 1183 only the V2 forge UI sets this flag; the dominant V1 path doesn’t log it even when SuperPrompt fires. So a low number here (e.g. 2%) reflects V2 routing share, NOT how often enrichment actually ran. Tracking under-coverage; will be fixed when V1 logging is wired up or V1 is retired.
- Berklee batch lift
- Composite + prosody comparison before vs after Build 935, when Pattison prosody rules and Stolpe destination-writing priors were pushed directly into the forge + gauntlet system prompts. A positive composite delta + a negative prosody delta is the target.
- arc variety
- The percentage of songs whose arc shape is NOT ‘flat’ (i.e. everything that’s single-axis, shaped, or volatile combined). High = the pipeline is producing songs with motion; low = monotone output.
Methodology + disclosure
- All data sourced from the
forge_metricstable, populated fire-and-forget after every completed forge. - Any slice below 30 samples is suppressed. No user, song, or lyric identifiers are included in this report.
- This report is updated hourly via ISR. Refresh to pull the latest snapshot. The open rubric is at /scoring/standard.