Surfaced: Build 2493 WAR Room — R83 (the Listener Seat).
The anti-pattern this names
When the proof sprint is unstarted, every new pillar evaluation is procrastination dressed as productivity. Strategy debates feel like work; cold DMs feel like exposure. The build-mode infrastructure (analytics, dashboards, CTAs, landing pages, retention emails) accumulates because it’s the path of less psychological resistance. The Phase 1 gate signals — runs, share rate, paid conversions — only populate from outreach. Infrastructure does not produce them.
The check
Before starting any new pillar / scaffold / dashboard build, ask: what would change about this decision if I had 30 days of real proof-sprint data? If the answer is "I’d know whether to ship this at all," then the data is the prerequisite, not the deliverable.
Surfaced: Build 2544 WAR Room of 100 Creatives — R93 (the Listener Seat again).
The anti-pattern this names
There is a permanent gap between what an AI lyric tool can do and what a human collaborator does. The gap is presence. Every other gap closes with more compute. This one does not. The product can write the song, grade the song, map the user onto the Atlas, generate the TikTok. The thing it cannot do is be IN the room when the user is writing at 2am. Every other AI tool overclaims; this product names the gap. The modesty is the moat.
The check
Before any marketing copy ships that promises completeness, companionship, or replacement of a human collaborator, ask: am I trying to fill the gap SA#12 names? If yes, the copy is wrong. The product is one tool in a craft toolkit; it is not the entire toolkit.
Surfaced: Build 2662 WAR Room of the Memorial + Wedding song-creation pipeline — Round 83 (Panel groups B + F unanimous, refined R88, R93).
The anti-pattern this names
Memorial + wedding songs are usually heard exactly once in the context they were made for. Most listeners will not hear them attentively from start to finish. The radio-pop convention puts the most-cuttable line in the bridge or the final chorus — that convention assumes attentive listening from beginning to end. Funerals don't work that way. Weddings don't work that way. By the time the bridge arrives at a graveside, half the audience is already wiping eyes; by the time a wedding song's final chorus hits, guests are watching the couple, not the song. The lyric's strongest moment has to land BEFORE the audience's attention has drifted.
The check
The inside-detail MUST surface by the end of the FIRST CHORUS. The highest-emotional line MUST appear in the first 90 seconds (~first 24 lines). The memorial/wedding vertical that first surfaced this rule was retired; the front-loading discipline it named remains a brand truth.
Surfaced: Build 2662 WAR Room of the Memorial + Wedding song-creation pipeline — Round 91 (Panel groups D + E unanimous).
The anti-pattern this names
When a buyer's intake answers are generic ("she was kind", "we love each other"), the failure mode is to write a generic song from generic source material. Most failed memorial / wedding songs are downstream of failed source material. The pipeline used to take the buyer's input verbatim and produce a sentiment-shaped output. Mary Oliver does not let "she was kind" stand. Neither do we. A song is the friction between specifics, not the average of universalities.
The check
The writing room must invent specific details that COULD be true of the subject, then ground them in something the prompt DID say, however small. "She was kind" + "red boots every day" → the song uses red boots, not kind. The memorial/wedding vertical that first surfaced this rule was retired; the specificity discipline it named is applied across every genre arc.
Surfaced: Build 2673 (operator-surfaced discipline — during a routine review the operator asked whether using named artists in the voice-roster panels carried any liability risk; the discussion crystallized the rule that had been implicit since B2635 but never named).
The anti-pattern this names
Every craft discipline operates by reference. Writers study writers; painters study painters; songwriters study songwriters. The MFA syllabus says read Mary Oliver because that's how craft is taught. SongForgeAI's writing room does the same — but the moment a reference becomes an imitation, or the moment a reference name surfaces in delivered customer-facing output, the discipline breaks. The boundary is mechanical, not philosophical: names in the prompt, never in the artifact.
The check
Names live in the prompt as internal craft references, never in delivered output. A display-time scrubber (src/lib/artist-name-scrubber.ts) strips "in the style of [artist]" framing from public song surfaces before delivery. Published as the "No artist-identity forgery" structural refusal at /standards/ethics under CC BY 4.0.
Surfaced: Build 2733 (100-song batch audit, May 2026 — WAR ROOM × 100 rounds × 10 expert panels). A mechanical motif audit on the 232 KB May 2026 100-song export found unnaturally high motif counts ("voice" 155×, "breath" 140×, "hands" 101×, "crack" 96×, "whispered" 93×, "chest" 66×); the system was reaching for the same emotional anatomy + bridge architecture + therapy-pivot resolution in 40+ of 100 songs.
The anti-pattern this names
Single-song discipline (every other Sacred Accident) governs what each song does well. A single excellent song doesn't prove the system; 100 songs that all reach for "something in my chest opening" prove the system is one writer with seven feelings doing 100 takes. The category-of-one moat ("we publish craft standards") only works if the standards apply to the BATCH, not just to individual songs. Without cross-song discipline, every song defends itself perfectly while the catalog reads as a single voice on repeat.
The check
Four-wave cure: Wave 1 (B2730-B2733) installs per-song pressure (forge-prompt prohibition of 6 body-anatomy phrasings + 3 bridge-architecture tells; banned-terms expansion to 32 motif-cluster phrases; scoring rubric penalties for motif over-clustering, weak choruses, broken syntax; this discipline named). Wave 2 (~8 hr queued): src/lib/claude/batch-motif-ledger.ts + bridge-architecture picker + chorus-compression second pass. Wave 3 (~6 hr queued): check-motif-saturation CI ratchet + /admin/system-health surface. Wave 4 (longer-term): voice-roster rebalance (JOY + HUMOR + NARRATIVE + SENSUALITY + FRIENDSHIP clusters) + per-batch emotional-mode quotas. A catalog has 100 voices; a system has one voice repeated 100 times. SA#16 commits to remaining a catalog.
Surfaced: Build 2763 (5-song stress-test WAR Room × 100 rounds, May 2026). The operator ran a 5-song stress test with difficult prompts; the reviewer's response listed 10 concrete recommendations. Across them, ONE pattern emerged: 9 of 10 recommendations were special cases of the same parent failure mode — the model optimizes for craft and forgets the brief.
The anti-pattern this names
Quality measures "is this a good song"; fidelity measures "did this answer the question I asked?" They are independent axes. A song can be A+ quality + F fidelity (gorgeous, wrong prompt) or C+ quality + A+ fidelity (dutiful but bland). Both happen routinely. The 12-metric rubric has been measuring one axis for two years; the next leap is measuring both. Without fidelity scored separately, the system can ship "good" songs that drift from the user's brief and the rubric will reward them. A working writer who hands the system a specific brief and gets back a beautiful song about something else will (rightly) churn.
The check
21-build Constraint-Aware Forge (CAF) roadmap across 3 phases (Phase 1 = brief extractor + structure/thesis/specificity audits + pipeline wiring; Phase 2 = prompt amendments + chorus evolution planner + earned transcendence + conditional sensory rewrite + premise-match Haiku judgment + dashboard two-grade chip; Phase 3 = leaderboard composite + public docs at /scoring/standard/fidelity + RFC-0010 + ratchet CI gate + Rap Mode + npm package). All 21 items shipped between B2763–B2820. Trigger phrase: `Continue Constraint-Aware Forge`. Canonical phrasing: `BRAND.sacredAccident17`.
Surfaced: Build 2827 (Deep Audit 2026-05-20 follow-up). The Quality Council ritual is mandated every 3 days, with a 7-day overdue threshold enforced by scripts/check-cadence-health.ts. The audit found the last Quality Council entry was 2026-05-04 (Build 2027) — 16 days stale, more than 2× the overdue threshold. During the same 16-day window 44 builds shipped (B2783-B2826), including 21 CAF Phase 2 + Phase 3 ship-throughs, a public RFC, an npm package build, and three nightly ratchet hardenings. None of it was synthesized; none of the across-build patterns were named.
The anti-pattern this names
Every prior Sacred Accident governs SONGS or SYSTEMS that produce songs. SA#18 governs the discipline THAT PRODUCES the disciplines. The Quality Council surfaces drift in scoring + product surfaces; the Trust Decay Audit surfaces drift in public claims vs. implementation; the Bet Review surfaces drift in resource allocation. Each ritual is a class-of-bug detector. If the detector itself drifts, every class-of-bug it catches starts accumulating silently until an external audit happens. 44 builds of unanalyzed velocity is exactly the window in which the system's gravity fights the user's intent (SA#17) without anyone naming it. The meta-finding was that 16 days of cadence-data was missing — a single class-of-bug that hides every other class.
The check
Three-wave cure: Wave 1 (B2827, this build) — the missing Quality Council entry written; SA#18 named in docs + brand.ts + data.ts. Wave 2 (B2828-B2830) — promote check:cadence-health to BLOCKING (currently warn-only); add overdue badge to /admin/system-health. Wave 3 (queued, longer-term) — auto-fire actionable prompt on overdue in CLAUDE.md "Pre-push ritual"; 15-minute Quality Council variant (synthesize last 5 commits, not last 50 builds) to remove the activation barrier that caused the 16-day skip. Canonical phrasing: `BRAND.sacredAccident18`.
Surfaced: Build 2860 (Rap Excellence WAR Room §8 Round 99, ratified at the close of the 20-build WAR Room arc B2838-B2859). Operator brief at the WAR Room's open: "Our system does not create rap lyrics all that well. I would like you to do deep research to achieve a massive upgrade for the rap genre." Round 99 asked: why did rap quality slip below default-genre quality for so long? Because audit primitives existed for the genres the panel could feel (Americana, folk, conscious pop) but rap shipped with a single mode + single rubric-weight override. The composite was correctly weighted; the PER-BAR DIAGNOSIS of why a rap verse fails (dead bar, slant at peak, hook obscuration, run-on) was never encoded.
The anti-pattern this names
SongForgeAI scores all genres against the 12-metric quality rubric. The rubric's pass/fail signal is sufficient for genres the panel can feel intuitively (the panel's literary anchors carry the diagnostic load). For genres the panel CANNOT feel — genres with their own technical vocabulary, conventions, and failure modes — the rubric is necessary but insufficient. Saying "M3 = 62" tells the operator the song is bad at rhyme intelligence; it does NOT tell them WHICH bar leaned on a forced rhyme, or that bar 7 hook-obscures, or that the verse violates the subgenre's syllable band. Without per-bar diagnosis, the operator + the forge cannot improve the song iteratively — they can only re-roll.
The check
The 20-build Rap Excellence WAR Room (B2838-B2859) is the proof-of-concept implementation. Phonology layer (B2838) + phoneme-aware audit (B2839) + 5 subgenre profiles (B2840) + 9-dimension critic loop (B2842) + per-bar validator (B2844) + corpus calibration loop (B2848-B2850) + public doc surface (B2857). When future genres need this depth, the pattern is inherited: domain-specific diagnostic substrate → audit primitives → subgenre profiles → diagnosis layer → calibration → public docs. Canonical phrasing: `BRAND.sacredAccident19`.
Surfaced: Build 2910 (Country Excellence WAR Room §4 closing argument, ratified at the close of the 15-build C1-C13 arc B2896-B2909). WAR Room Panel D — country authenticity scholars (Bill C. Malone, Diane Pecknold, Aaron Fox, Nadine Hubbs, Evan Malone, Charles Hughes) — surfaced this when arguing what should fail the genre-eval test. Most generic LLM "country" output fails the C3 Forbidden Archive #2 (Object List Verse) and #3 (Could-Be-Any-City Verse) — they collect country-coded vocabulary (truck, beer, tailgate, bonfire, sundress, dirt road) without grounding the narrator in coherent material logic. A great country song with NONE of those signals still reads as country if the narrator inhabits a specific world; a mediocre country song with ALL of them still reads as cosplay if the narrator could be subbed into a city apartment without breaking 80% of the lyric.
The anti-pattern this names
Genre-coded vocabulary is treated as the authenticity signal. Object List Verse + Could-Be-Any-City Verse + Demographic Cosplay (fake "y'all" / "fixin' to" / phonetic spellings) all stack the country-keyword density without producing a coherent narrator. The "sub the narrator into a city apartment" test fails — the lyric works in any setting because it has no setting. Brett (Nashville pro per the B2861 deep-audit) flagged this exact pattern as "the rubric scored a verified country hit as Collapsed because the rubric was reading country-keyword density and missing inhabited-narrator signal." This is the genre-evaluation gap SA#19 named, specifically applied to country.
The check
The 15-build Country Excellence WAR Room (B2896-B2909) ships the inhabited-vs-inherited test as system enforcement. The C3 Country Forbidden Archive (B2898) names 3 specific cosplay-failure modes (#1 Demographic Cosplay, #2 Object List Verse, #3 Could-Be-Any-City Verse). The C9 country critic (B2905) flags these by canonical name. The C11.5 corpus calibration (B2908) enforces the rule via 35 verified-hit credential-validated anchors — any rubric mis-scoring becomes the documented "Brett Collapsed" failure mode. The pattern inherits to future genres needing authenticity tests: name the cosplay failure modes for that genre, encode them in audit primitives, validate via credential-validated corpus. Canonical phrasing: `BRAND.sacredAccident20`.
Surfaced: Build 2925 (Pop Excellence WAR Room §1 closing argument, ratified at the close of the 13-build P1-P12 arc B2911-B2924). WAR Room Panel C — phonetics + singability researchers (channeling Andy Bennett pop phonetics, Nate Komaniecki melodic-math research, Pattison-on-pop) — surfaced this when arguing what makes a pop chorus survive streaming compression + earbud delivery. Max Martin's craft moat is not melodic invention; it is engineered phonetic mass. Every Top-40 chorus he has produced for 25 years lands 3-6 hook centers per chorus on Atlantic Vowels (AA / AO / AW / EY / IY / OW / UW). The chorus is engineered as a phonetic structure; semantics inhabits the structure, not the other way around.
The anti-pattern this names
Generic LLM pop output optimizes for semantic precision first — finding the meaning, then fitting the words around it. The result is choruses with closed-vowel landings, dense onset clusters, and pickup-syllable run-ups that the streaming-era listener swipes away from before the second line. The lyric on the page is not the lyric the singer can sing — the singer reshapes the words to make them singable, which means craft control passes from writer to performer. Every audited Diamond pop song (Rolling in the Deep, bad guy, Blinding Lights, Shake It Off) shows extreme Atlantic Vowel density on chorus stress positions. Pop's most-canonical "I am" affirmation choruses (Stronger / Brave / Roar) all land their stresses on AY / OW / IY — not by accident, by craft. The LLM-default pop chorus does not know this rule + cannot enforce it without the audit substrate.
The check
The 13-build Pop Excellence WAR Room (B2911-B2924) ships the phonetic-mass test as system enforcement. The P2 phonology layer (B2912) measures PVR (Pop Vowel Resonance) + SSR (Stress-Singability Ratio). The P6 CHR primitive (B2916) counts hook centers per chorus block. The P12 forge amendment (B2924) gates output with the CHORUS PHONETIC SHAPE: declaration — the forge sketches the chorus phonetic structure BEFORE the chorus words exist. The P9 critic (B2920) flags Phonetic Pileup + Centerless Chorus by canonical name. The P11.5 corpus calibration (B2923) enforces the rule via 30 verified-hit credential-validated anchors. The pattern inherits to future genres where singability is load-bearing (Latin Pop, Worship Anthem, R&B). Canonical phrasing: `BRAND.sacredAccident21`.
Surfaced: Build 2940 (Latin Excellence WAR Room §1 closing argument, ratified at the close of the 15-build L1-L13 arc B2926-B2939). WAR Room Panels E + I — linguists/dialectologists + adversarial/competitive — surfaced this when arguing what makes Latin lyric generation distinctive. Spanish has three major lyric-engine traditions sharing one grammar but diverging on particles, pronouns, syntax, phonological reduction, and slang. Mixing markers across regions is the genre's distinctive failure mode; native listeners hear region mismatch instantly.
The anti-pattern this names
Generic LLM Spanish-language output mixes dialect markers across regions because the model treats Spanish as one register with vocabulary substitutions. A line with Mexican "ándele" + Puerto Rican "¿qué tú quieres?" + Iberian "chaval" is grammatically correct Spanish but reads as tourism — written by someone who studied the regions from outside rather than by someone who inhabits one of them. The opposite failure (Tourist Spanish — grammatically correct but identity-less, region-free Duolingo Latin) is equally a dialect-ownership failure. Both are caught by the same principle: coherence beats coverage.
The check
The 15-build Latin Excellence WAR Room (B2926-B2940 / L1-L13) ships the dialect-ownership test as system enforcement. The L6 DCS (Dialect Consistency Score) deterministically detects cross-region marker conflicts. The L8 DCT (Dialect Coherence Test, Haiku-judged) catches the subtler Tourist Spanish failure that DCS misses (grammatically correct but identity-less). The L9 critic flags Region Salad + Tourist Spanish + Slang Spam + Translated English Underneath by canonical name. The L11.5 corpus calibration enforces the rule via 40 verified-hit credential-validated anchors across 4 lanes. The L12 forge amendment gates output with the LANE-FIRST commitment — forge declares lane + dialect + persona BEFORE writing line one. Future genres where dialect coherence matters (Worship Anthem regional varieties, Latin-American sub-regions, African-language pop) inherit this pattern. Canonical phrasing: `BRAND.sacredAccident22`.
Surfaced: Build 2955 (R&B Excellence WAR Room §1 closing argument, ratified at the close of the 13-build R1-R13 arc B2941-B2954). WAR Room Panel D — R&B songwriters + craft researchers (channeling Babyface, Ne-Yo, The-Dream, Carvin Haggins, James Fauntleroy, Tracey Thorn, Pat Pattison on showing-vs-telling) — surfaced this when arguing what makes R&B distinct from pop confession-theater. R&B is the genre where emotional truth is the product. The genre's craft moat is not melodic invention (pop owns that) or phonetic mass (pop owns that too) — it is the discipline of grounding emotional claims in concrete, witnessable evidence. The AID rule names what every great R&B lyricist already does and what every generic LLM R&B output fails: pair each "I miss you" with one of (a) an action the speaker took, (b) an image the speaker saw, (c) a named detail. Without the receipts, the lyric collapses into the Hallmark register.
The anti-pattern this names
Generic LLM R&B output optimizes for emotional vocabulary first — finding feeling-words and stacking them. The result is verses full of "I miss you / I need you / you're everything / without you I'm broken / my heart is shattered" — grammatically correct, emotionally generic, indistinguishable from one writer to another. The four AID-Violation failure modes (Telling Not Showing / Generic Romance Mush / Cliché Rhyme Pair / Performative Vulnerability) are all special cases of the same parent failure: making emotional claims without supplying receipts. Every audited Diamond R&B song (End of the Road / We Belong Together / Confessions Part II / Earned It) grounds its claims in concrete actions, images, and named details — "Although we've come to the end of the road" pairs with the bridge's "I can't let go." "Pyramids" pairs every claim with a specific sensory image. The LLM-default R&B verse does not know this rule + cannot enforce it without the audit substrate.
The check
The 13-build R&B Excellence WAR Room (B2941-B2955) ships the AID-rule test as system enforcement. The R2 VDA (Vowel Downbeat Alignment) measures open-vowel discipline on chorus peaks. The R6 NCD (Narrative Concrete Density) deterministically counts concrete nouns + active verbs + color words + numeric/time details per verse — the AID rule operationalized. The R8 HVT (Hook Vulnerability Test, Haiku-judged) catches the subtler failure where the verse provides receipts but the chorus collapses to bare claims. The R9 critic's aidRuleAdherence axis weights this most heavily of the 6 dimensions. The R10 plugin composes NCD at 35% (highest weight) + VDA at 20% in the R&B craft score. The R11.5 corpus calibration enforces the rule via 30 verified-hit credential-validated anchors across 5 substyles + 4 paradigms. The R12 forge amendment gates output with paradigm-first declaration + verse-by-verse AID-self-audit BEFORE the lyric ships. Future genres where emotional-claim density is load-bearing (Singer-Songwriter, Adult Contemporary ballad, Christian Worship, Soul-Pop crossover) inherit this pattern. Canonical phrasing: `BRAND.sacredAccident23`.
Surfaced: Build 2970 (Worship Excellence WAR Room §1 closing argument, ratified at the close of the 13-build W1-W13 arc B2956-B2969). WAR Room Panels D + I + J — worship pastors + liturgists + adversarial — surfaced this when arguing what makes worship distinct from CCM-flavored pop. Worship is the FIRST genre deep-dive where the load-bearing failure mode is NOT a craft gap (rap rhyme / country authenticity / pop phonetic / Latin dialect / R&B vulnerability) — it's a PASTORAL gap. Vince Wright's Berean Test (2008), which drives worship-evaluation discipline across mainstream evangelical, Reformed, and seeker-sensitive review processes, ratified this in a single mechanical rule: any second-person ("You / Your") pronoun without an explicit divine identifier in the same section receives an automatic -2 point penalty out of 10. The penalty exists because ambiguous-pronoun worship songs are indistinguishable from love songs to a human lover — the "boyfriend-style worship" failure mode.
The anti-pattern this names
Generic LLM worship output optimizes for emotional vocabulary first — finding feeling-words and stacking them into ambiguous "You" addresses. The result is choruses that could ship unchanged to a human lover: "You're all I want / You're all I need / You hold me close / You're my everything." Grammatically religious-coded, semantically indistinguishable from pop love song. The four Ambiguity-Violation failure modes in the W3 Worship Forbidden Archive (Boyfriend-Style Worship / Anthropocentric Drift / Doctrinal Vagueness / Implied Subject) are all special cases of the same parent failure: addressing or referencing the divine subject without naming Him. Every audited canonical worship song (In Christ Alone / What A Beautiful Name / Holy Forever / Goodness of God / Holy Holy Holy / How Great Thou Art) names the divine subject explicitly in every chorus — by proper noun (Christ / Jesus / God / Lord / Father / Spirit) or by unique divine attribute (cross / resurrection / blood of the lamb / made the heavens). The LLM-default worship chorus does not know this rule + cannot enforce it without the audit substrate.
The check
The 13-build Worship Excellence WAR Room (B2956-B2970) ships the Name-the-Subject rule as system enforcement. The W2 CSR (Congregational Singing Range) measures vocal compass + open-vowel discipline on chorus peaks. The W3 Forbidden Archive ratifies 10 canonical failure modes split 4 Ambiguity-Violation + 6 Structural. The W6 APR (Ambiguous Pronoun Ratio) deterministically counts second-person pronouns lacking a divine identifier in the same section — the Berean Test mechanically operationalized. The W7 TFI / DGS / PRT / SAI primitives quantify the other four worship-craft dimensions. The W8 BVT (Berean Vetting Test, Haiku-judged) applies the classical 4-category Berean evaluation (Lyrical Message 20% / Scriptural Alignment 40% / Outsider Accessibility 20% / Doxological Glorification 20%). The W9 critic's worshipFidelity axis is weighted highest of the 6 dimensions. The W10 plugin composes APR at 35% (highest weight) + CSR / TFI / DGS at 15% each + PRT / SAI at 10%. The W11.5 corpus calibration enforces the rule via 30 verified-hit credential-validated anchors across 5 substyles + 4 paradigms + 5 traditions. The W12 forge amendment gates output with paradigm+tradition-first declaration + NAME-THE-SUBJECT self-audit BEFORE the lyric ships. Worship is also the FIRST genre to introduce a tradition policy layer (broadly-evangelical / reformed-confessional / pentecostal-charismatic / catholic-liturgical / gospel-participatory) — denominational fit is a first-class concern absent in the prior five genre deep-dives. Canonical phrasing: `BRAND.sacredAccident24`.
Surfaced: Build 2985 (Indie Excellence WAR Room §1 closing argument, ratified at the close of the 13-build I1-I13 arc B2971-B2984). WAR Room Panels C (linguists) + D (indie songwriters) + I (adversarial) surfaced this paradox when arguing what distinguishes canonical indie lyrics from generic alt-pop. Across the cited computational musicology + 30-hit indie corpus benchmark (Pavement / Phoebe Bridgers / Bon Iver / Sufjan / Strokes / Smiths / MBV / Pixies), one finding converges: indie's craft moat is NOT melody, NOT production, NOT instrumentation — it is the hyper-specific local detail that grounds emotional claims in objective physical reality.
The anti-pattern this names
Generic LLM indie output stacks emotional vocabulary first. The result is verses full of "I am so lonely / my heart is broken / I miss you so much / I cry myself to sleep" — grammatically indie-shaped, semantically indistinguishable from any pop song. The four Generic-Fog failure modes (Generic Fog / Cliché Trap / Persona Drift / Authenticity Theater) are all special cases of the same parent failure: emotional declarations without concrete receipts. Every audited canonical indie song (This Charming Man / Maps / Casimir Pulaski Day / Holocene / Where Is My Mind? / There Is a Light) grounds its claims in specific, named, physical anchors — a double-decker bus / a Casimir Pulaski Day / a Holocene revelation / a Caribbean diving anecdote.
The check
The 13-build Indie Excellence WAR Room (B2971-B2985) ships the hyper-specificity rule as system enforcement. The I2 SAR (Slant-Assonance Rhyme) measures loose-rhyme discipline (indie REJECTS strict AABB). The I3 Forbidden Archive ratifies 10 canonical failure modes split 4 Generic-Fog + 6 Structural. The I6 HSI (Hyper-Specificity Index) deterministically counts verse lines containing at least one concrete-specific anchor (brand / place / body / time / object / color / numeric detail) — SA#25 mechanically operationalized via 5 lexicons + proper-noun detection + digit-time-string anchors. The I7 LDI / GFI / NVC / MAR primitives quantify the other four indie-craft dimensions. The I8 APT (Authenticity-Posture Test, Haiku-judged) applies the 3-axis Indie Authenticity Circuit (nostalgic / originality / iconic) + post-authentic posture detection — the FIRST genre arc to operationalize the post-authentic-era trap (Authenticity Theater). The I9 critic's hyperSpecificity axis is weighted highest of the 6 critic dimensions. The I10 plugin composes HSI at 35% + SAR/LDI/GFI at 15% each + NVC/MAR at 10%. The I11.5 corpus calibration enforces the rule via 30 canonized indie anchors across 5 substyles + 4 narrative shapes + 4 POV modes. The I12 forge amendment gates output with persona-card + narrative-shape declaration + verse-by-verse HSI-self-audit before output. Indie also introduces the NARRATIVE-SHAPE-FIRST discipline (scene-chain / apostrophe / memory-collage / title-thesis) as a structural commitment BEFORE word choice. Canonical phrasing: `BRAND.sacredAccident25`.
Surfaced: Build 3000 (Folk Excellence WAR Room §1 closing argument, ratified at the close of the 13-build F1-F13 arc B2986-B2999). WAR Room Panels A (Pat Pattison + Berklee songwriting tradition) + D (Dylan / Mitchell / Cohen / Springsteen / Prine craft lineage) + I (adversarial) converged on this: folk and singer-songwriter craft's most distinctive load-bearing principle is STRUCTURAL-EMOTIONAL PROSODY. The cited research states explicitly: "listeners do not just hear words; they physically and emotionally feel structure." Folk has minimal production overlay — the lyric + voice + acoustic accompaniment must carry all emotional weight, and the structure is the primary somatic vehicle.
The anti-pattern this names
Generic LLM folk output writes every section as 4-line AABB quatrains regardless of emotional content. The result is heartbreak / grief / panic / uncertainty paired with rigid symmetric stanzas — the structure tells the listener "you'll get over it," reducing trauma to a recitation of facts. Additionally, function words land on strong-beat downbeats (greedy spots), forcing the singer to reshape words mid-delivery and breaking the conversational folk illusion. Every audited canonical folk + singer-songwriter song (Like a Rolling Stone / Both Sides Now / Hallelujah / Casimir Pulaski Day / Thunder Road / Fast Car / Holocene / Stick Season) uses ASYMMETRIC stanza shapes (3/5/6-line) when content is unstable, RESERVES symmetric 4-line quatrains for resolution sections, and avoids ALL greedy spots (Pat Pattison rule).
The check
The 13-build Folk Excellence WAR Room (B2986-B3000) ships the structure-IS-feeling rule as system enforcement. The F2 SDV (Structural Deviancy Value) deterministically counts 3/5/6-line asymmetric stanzas + line-length variance + delayed-rhyme position. The F6 PGI (Prosody-Greedy-Index) operationalizes Pat Pattison's greedy-spot rule via ~70 function-word lexicon + syllable-stress proxy. The F3 Forbidden Archive ratifies 10 canonical failure modes split 4 Structural-Mismatch + 6 Craft. The F7 NAP / SLC / RSV / STC primitives quantify the other four folk-craft dimensions. The F8 SMM (Specificity-Memory-Mirror Test, Haiku-judged) tests the Specificity Paradox: hyper-localized detail triggering universal listener projection. The F9 critic's structuralProsody axis is weighted highest of the 6 critic dimensions. The F10 plugin composes PGI at 30% + SDV at 20% + NAP/SLC at 15% each + RSV/STC at 10%. The F11.5 corpus calibration enforces the rule via 30 canonized folk anchors across 5 substyles + 4 craft paradigms + 4 POV modes. The F12 forge amendment gates output with paradigm + POV + EMOTIONAL ARC SHAPE declaration + stanza-by-stanza structure-self-audit before output. Folk is the FIRST genre arc to encode STRUCTURE as the load-bearing primary axis (prior arcs encoded vocabulary / phonetics / dialect / specificity / pastoral fit / authenticity). Future genres where structure-emotion alignment matters (Spoken Word, Slam Poetry, Conceptual Album-Track, Theatre-Song) inherit this pattern. Canonical phrasing: `BRAND.sacredAccident26`.
Surfaced: Build 3015 (Rock Excellence WAR Room §1 closing argument, ratified at the close of the 13-build Rk1-Rk13 arc B3001-B3014). WAR Room Panels A (acoustic-physics + vocal coaches) + C (Mutt Lange / Desmond Child stadium production lineage) + D (Bruce Springsteen / Neil Peart / Pete Townshend craft lineage) converged on this: rock craft's most distinctive load-bearing principle is ACOUSTIC-PHYSICS DISCIPLINE at chorus altitude. The cited research grounds the rule in measurable formant physics, not craft tradition alone — closed-vowel F1 sits ~280-360 Hz, open-vowel F1 sits ~600-850 Hz, and chest-voice F0 ceiling sits ~392-440 Hz (G4-A4). The collision is structural, not stylistic.
The anti-pattern this names
Generic LLM rock output writes choruses that end on closed-vowel words ("see", "free", "you", "true", "love") because closed-vowel rhymes are easier to find and feel "poetic." The result is choruses that demand mid-note vowel reshaping at chest-voice peaks — singer-distorting, listener-jarring, payoff-killing. Every audited canonical rock anthem (Stairway to Heaven / Don't Stop Believin' / Livin' on a Prayer / Highway to Hell / Born to Run / Free Fallin' / Tom Sawyer) terminates its chorus peaks on open vowels. The few apparent exceptions (We Will Rock You's "you", We Are the Champions' final climb) demonstrably reshape mid-note — and pay the price in conversational illusion.
The check
The 13-build Rock Excellence WAR Room (B3001-B3015) ships the vowel-yield-to-pitch rule as system enforcement. The Rk2 RVL (Register-Vowel Landing) measures chorus-wide open-vowel discipline. Open vs closed vowel taxonomy single-sourced via ROCK_OPEN_VOWEL_BASES and ROCK_CLOSED_VOWEL_BASES sets. The Rk3 Forbidden Archive ratifies 10 canonical failure modes split 4 Acoustic-Mismatch (Closed-Vowel Cliff / Greedy Spot On Peak / Tension-Less Title / Hook Shadowing) + 6 Craft. The Rk6 VPI (Vowel-Pitch Integrity) — THE LOAD-BEARING SA#27 PRIMITIVE — mechanically detects closed-vowel chorus peaks via CMU phoneme analysis + stress detection at line-end positions; substyle ceilings are arena-anthem + hard-rock-blues ≤ 0.10 (strictest), classic-rock-radio + roots-heartland ≤ 0.15, prog-rock-literary ≤ 0.20. The Rk7 TTI / RHC / CWM / GDD primitives quantify title-tension, rhythmic hook contrast, concreteness density, and gap design. The Rk8 SPT (Stadium Participation Test, Haiku-judged) tests whether 50,000 strangers would sing the chorus back on first listen. The Rk9 critic's vowelPitchIntegrity axis is the SA#27 axis weighted alongside titleTension / hookContrast / stadiumIntegration / substyleAdherence / authenticityCredibility. The Rk10 ROCK_CRAFT_PLUGIN composes VPI at 30% (load-bearing) + RVL 15% + RHC 20% + CWM 20% + GDD 15%. The Rk11.5 corpus calibration enforces the rule via 30 canonized rock anchors across 5 substyles + 4 craft paradigms + 5 POV modes + per-axis expected bands. The Rk12 forge amendment gates output with paradigm + POV + OPEN-VOWEL CHORUS RHYME POOL declaration + chorus-line-end VOWEL-PITCH SELF-AUDIT before output. Rock is the FIRST genre arc to encode ACOUSTIC PHYSICS as the load-bearing primary axis. Canonical phrasing: `BRAND.sacredAccident27`.
Surfaced: Build 3114 (Sacred Accident codification — surfaced by the 19-build B3094-B3113 session that closed FOUR distinct instances of the same root pattern in one week). Unlike SA#19-SA#27 which were ratified at the close of per-genre WAR Rooms, SA#28 was surfaced by the operator-noticed pattern across CONTAMINATION BUGS rather than by a craft audit. It is closer in shape to SA#18 (cadence ritual meta-drift) than to the genre-craft SAs — both describe meta-patterns where the system's behavior diverges from what the visible surface implies. The four canonical examples: (1) Vault corpus contamination (B3104) — corpus had no arc filter; worship + Christian-hip-hop songs dominated SuperStyle embedding mean across every arc. (2) Customize panel in batch mode (B3104 investigation) — ForgeV2BatchPanel hardcoded cleanMode='' overriding visible Customize panel state. (3) Splice handoff degeneration (B3106) — use-forge-handoffs auto-enabled splice with the SAME genre on both sides when user deep-linked from /genres/[arc]. (4) Surprise Mode visual lie (B3105) — button rendered in active/selected style even when the user had not toggled it.
The anti-pattern this names
Every generic LLM-assisted codebase produces this pattern eventually: a subsystem reads state from a source the operator did not just touch (an inherited handoff, an ambient corpus filter, a hardcoded panel default, a visual style that implies selection), and the gap between "what the visible surface implies the system will do" and "what the system actually does" becomes a contamination class. The B3094-B3113 instances all followed the same forensic shape: operator saw output X, predicted output Y based on visible surface state, the gap was an implicit-state channel the surface did not expose. Single-layer fixes leave the class alive in adjacent code paths; the cure is multi-layer + structural — name the state in the comment, render it visibly in the surface, make the visible render reflect the ACTUAL value not the default.
The check
The four shipped fixes operationalize the rule: (a) B3104 Vault filter — explicit `arc_slug` filter on every Vault corpus query used by SuperStyle generation; contamination eliminated at the query layer. (b) B3105 Surprise-button visual — default render flipped to the dark/inactive style so visible state matches actual state. (c) B3106 Splice handoff — same-genre splice handoffs collapse to no-splice at three layers (handoff seeding, body-builder server-side, SuperStyle re-derivation) so the regression is visible the moment any future change re-opens any one of them. (d) B3114 Catalog source-arc badge — added per-card arc chips to PromptBatchPicker (/genres/[slug]/prompts) and GenreLibraryBatchPreview (forge batch mode); arc identity no longer inferred from URL context, every card carries its own source-arc identity. The triple-locking discipline from B3106-B3108 generalizes: when a contamination class is identified, fix it at every layer where the implicit channel exists. SA#28 is the most operational of the meta-SAs because every canonical example shipped a code-level fix. Relationship to SA#18: cadence-ritual drift IS implicit-state drift at the team-process layer — the cure for both is identical (surface the implicit state at the point of action). Canonical phrasing: `BRAND.sacredAccident28`.
Surfaced: Build 3280 (Genre Routing WAR ROOM — operator-reported 2026-05-25). User submitted Surprise Mode prompt "A funk rock song inspired by Red Hot Chili Peppers"; system produced "Intimate folk ballad with fingerpicked acoustic guitar, male baritone vocals, 70 BPM in 4/4 time, key of G major" — a complete genre-routing collapse. A 100-expert forensic WAR ROOM with four parallel agents identified four converging failure modes: (1) Surprise Mode at use-surprise-flow.ts:275 dispatched only the composed prompt text, discarding the structured palette.artistBrief that already held funk-rock genre intent. (2) SuperPrompt at superprompt/route.ts:91 INSTRUCTED Sonnet to preserve the artist verbatim in paragraph 1 but had zero programmatic enforcement — Sonnet routinely paraphrased "Red Hot Chili Peppers" into "California funk-rock pioneers", silently scrubbing the INSPIRED_BY regex window. (3) post-forge-pipeline.ts:241 populated the SuperStyle "rawPrompt" field with the already-enhanced text, so SuperStyle's artist-extraction also saw the paraphrased artifact, fell through to Haiku reading the lyrics, and chose folk-ballad to match the lyric mood. (4) genre-profile.ts:335 used substring `.includes("rap")` matching that false-positive-fired on common words like "wrap", "rapper", "therapy" — a third independent detector with a different algorithm than the word-boundary detector at genre-modes.ts:3825.
The anti-pattern this names
Every multi-stage LLM pipeline that uses an enhancement step (SuperPrompt, query rewriter, summarizer, translator) eventually produces this pattern: the enhancement is allowed to paraphrase the structured signal (artist name, genre token, locale, persona), and downstream consumers that re-extract via regex silently no-op when the paraphrase succeeds. The structured truth exists upstream in EVERY case (database lookup, curated library, user dropdown selection, deep-link query param) — the system just chose not to plumb it through. The audit shape is consistent: the operator submits a prompt naming a specific artist or genre, the model rewrite step turns the proper noun into a descriptor phrase, the downstream artist-lock returns null, and the default attractor (folk-ballad for slow lyrics, hip-hop for fast verses, pop for repeated hooks) wins. Single-layer fixes leave the class alive in adjacent rewrite stages; the cure is structural — promote the structured signal to a typed argument, demote text-parsing to a fallback.
The check
The four shipped fixes operationalize the rule: (a) B3276 Surprise Mode persona handoff — use-surprise-flow.ts:275 now passes `personaName + personaBrief` alongside the composed prompt; the forge's B2557 persona short-circuit at forge-stream.ts:362-372 bypasses regex extraction entirely when structured state is present. (b) B3277 rawPrompt end-to-end — added `rawPrompt?: string` to SongSeed (types.ts:21), threaded through forge-effects.ts → forge-v2-effect-deps.ts → buildForgeBody → request-preflight → forge-stream-session, so downstream extraction reads the original user paste, not the SuperPrompt-paraphrased artifact. (c) B3278 word-boundary matching — genre-profile.ts:335 migrated from substring `.includes()` to word-boundary `\b…\b` regex; "therapy / wrapping / lollipops / rocky" no longer false-positive-fire. (d) B3279 integrity ratchet — scripts/check-genre-routing-integrity.ts ships 15 canonical fixtures (8 happy-path genre arcs + 4 false-positive-resistance cases + 3 artist-extraction patterns) wired into check:all; every fixture is a regression that cannot silently recur. Relationship to SA#28: SA#28 is the UI-side discipline (surface implicit state at the point of action); SA#29 is the source-side discipline (plumb structured state directly to the consumer). Together they close the implicit-state-as-attack-surface class at both ends. Canonical phrasing: `BRAND.sacredAccident29`.
Surfaced: Build 3298 (The One Line WAR ROOM — 100-round audit). The product committed to a bet: detect ONE line per forged song that carries the unrepeatable signature of top-0.5% human writing. The first move every council voted unanimously for was NOT "build the detector" but "build the corpus the detector calibrates against." Operationalized across B3299-B3305 — corpus skeleton + 7-feature taxonomy + 9 structural invariants in CI (B3299), detector + 200 operator-original entries written by the operator themselves as ground truth (B3300), +50 operator-originals + ratio-based calibration that stays stable across corpus growth (B3301), OneLineGenerator with anti-optimization prompt + 250-line few-shot anchor pool (B3303), Haiku final-pass judge for the three features deterministic heuristics could not score (B3304), first public surface in the dossier where buyers see the detector's top candidate One Line + per-feature breakdown (B3305). 75% of the 336-entry corpus is operator-owned — no licensing surface, no third-party clearance, the calibration data is the product's own intellectual asset.
The anti-pattern this names
Detector built first, calibration handwaved. The shape is: someone codes up heuristic feature scorers, picks plausible thresholds by intuition, ships the detector as a public surface, and only later discovers that the thresholds are guesses because there is no inspectable ground truth to test against. When the system later misfires — flags lines that aren't canonical, misses lines that are — there is no trail to walk back. The fix invariably becomes "raise this threshold by 5" or "add another lexicon word", small patches that paper over the absence of foundational calibration data. The B3300 calibration test forced honesty here: three of the seven features failed the ratio test on the heuristic alone, which forced the Haiku judge in B3304. Without the corpus, the heuristic stubs would have shipped as if they worked + the misfires would have been invisible. The corpus made the limits of the heuristic legible.
The check
Before any future detector claim ships as a public surface, six steps in order: (1) Assemble a corpus of canonical examples + counter-examples, operator-owned where possible. (2) Tag the corpus against the taxonomy the detector will use, with inspectable + auditable tags — no opaque LLM grading at the calibration layer. (3) Validate the corpus with structural invariants in CI (sufficient examples per feature, balanced provenance, no orphan tags). (4) Calibrate the detector via a RANKING invariant (tagged > untagged) rather than absolute thresholds — ratio-based assertions stay stable as the corpus grows. (5) Surface the corpus' limits in detector output — when a feature can't be scored deterministically, fall back to an LLM judge with a clear isHeuristic flag in the result. (6) Never ship the detector as a public claim until steps 1-5 are green. The One Line arc B3299-B3305 IS the canonical implementation; future arcs (B3308 vault batch rank, future quality-of-craft surfaces) consume this ladder, they do not invent it. Relationship to SA#19 ("A genre we cannot evaluate cannot be a genre we can serve"): SA#19 is the genre-specific instance of SA#30. SA#30 is the parent discipline; SA#19 is the first place we put it to work. Canonical phrasing: `BRAND.sacredAccident30`. Operator-facing summary of the arc: docs/ONE-LINE.md.
Surfaced: Build 3331 (One Line pairwise battle calibration — 200-battle ratification). The B3325 substrate (one_line_battles + one_line_elo tables + Haiku judge + Elo K=32 update) ran two 100-battle rounds against the operator-owned 296-anchor corpus. The top Elo winners after 200 battles were moderate-composite (67-78) lines anchored in concrete imagery: "Tea ring where I've been setting down the same cold cup" / "Earned every breath like rent was due" / "My throat knows the weight / Of words I'll never say" / "Throat full of years I couldn't speak" / "Now the vein knows the needle better than it ever knew her prayers." The high-composite Elo losers were abstract-paradox / meta-aware / religious-poetic / smart-aphorism lines: "Every morning I construct me / Every night I come apart" (composite 86, Elo delta -32) / "I've optimized myself out of existence" (composite 77, Elo delta -30) / "Your hands become communion / And I taste salvation whole" (composite 75, Elo delta -46 across 3 battles). The Haiku judge's per-battle rationale named the discriminator each time: "uses concrete self-destruction to anchor abstract loss" / "uses fluorescent hum, engine idle to SHOW emotion not tell it" / "achieves concrete-abstract anchoring via hands as the specific anchor in a named room."
The anti-pattern this names
Composite scorer ships as the only quality signal for a craft detector. Composite (mean-of-top-N over a feature taxonomy) cannot distinguish "smart, abstract, well-assembled" from "concrete, embodied, located" — both trip many of the features. Without pairwise, the high-composite/low-concrete lines get treated as canonical when they are actually the cleverness-tax: lines that LOOK like the move but are not. The pattern shape is consistent (abstract paradox / meta-aware self-commentary / religious-poetic without anchor / smart-aphorism without specific room) and it is invisible to composite alone. Every future quality-of-craft detector built per SA#30 that ships without a pairwise counterpart will recapitulate this bias. The bias is not noise — it is a class — and the class is measurable: across 200 battles in the B3331 ratification round, 13 lines with composite >=75 lost head-to-head by Elo deltas of -100 or more.
The check
Every future quality-of-craft detector that ships per SA#30 must also ship: (1) A pairwise substrate — two tables: comparison records + Elo per item. Schema like one_line_battles + one_line_elo. K=32 standard Elo update. Initial rating 1500. (2) A judge — LLM-based per-battle adjudicator that returns winner + confidence + rationale. The rationale field is load-bearing — it is where the discriminator becomes legible to operator review. (3) A pair-pick strategy — similar-composite windows (+/-5 points works; +/-10 for cold-start) so the judge isn't asked predictable questions. Bias toward pairs with low Elo confidence. (4) A disagreement view — admin query that surfaces lines where composite says "great" and Elo says "average" (or vice-versa). This is the calibration-signal-as-dashboard. (5) A coexistence policy — when composite + Elo disagree by >100 Elo points (~12 composite points), pairwise wins for downstream surfaces (forge anchors, public ranking, public claims). Composite remains the cheap pre-filter; pairwise is the expensive corrective. Relationship to SA#30: SA#30 builds the corpus before the detector. SA#31 names the second axis the corpus makes visible — companion pair. Relationship to SA#25 ("The personal detail is the universal door"): SA#25 is the indie-specific genre instance; SA#31 generalizes it across all genres via pairwise calibration. Canonical phrasing: BRAND.sacredAccident31. Operator-facing arc: docs/ONE-LINE.md + docs/SACRED-ACCIDENTS.md (SA#31 section).
Surfaced: Build 3372 (Main Character Energy WAR ROOM). Operator surfaced a real-world failure: a "main character energy" prompt asking for upbeat, joyful, energetic, sudden bold living returned three meditative-introspective songs about gradual healing, identity loss, and confidence rehearsal. Quantified deficits: 4 active verbs across all 3 songs vs 15+ requested; ~3/10 valence vs 8+/10 requested; meditative ~65 BPM vs energetic ~120+ BPM requested; private rehearsal vs public bold acts; 2-year/6-month gradual healing vs "suddenly feeling like." The cultural-register diagnosis: the user asked for the TikTok / IG "main character energy" phenomenon (Lizzo "About Damn Time" — joyful, energetic, bold); the system delivered singer-songwriter "quiet resilience" (Phoebe Bridgers "Moon Song" — bittersweet, reflective, melancholic). This validated the B3366 corpus-audit prediction empirically: the system's introspective-register dialect overrides explicit prompt instructions. The B3371 CI ratchet had already confirmed the dialect lives in the system's BEST output (top500 anchors carry more dialect density than bottom500). SA#32 names the lesson: applying one register's rubric to every prompt produces register-blind output.
The anti-pattern this names
A single craft rubric (anti-inflation rules, specificity floors, restraint-rewarding critic voices) ships as universal. The rubric is calibrated for ONE emotional register — typically the register the operator-original anchors live in. When the user prompts for a different register (joy, swagger, rage, lust, awe), the system runs the prompt through the same rubric, optimizes toward the same metrics, and produces a song in the original register with the new prompt's surface details. A pop banger that needs "high active-verb density, present-tense dominance, public scene-setting, body-in-motion, repetition as celebration" gets refined toward "specificity through restraint, body as witness, object as relic, silence as resolution." The user reads the result, thinks the system misunderstood the prompt — but the deeper failure is that the system was register-deaf from the rubric-design moment. Symptoms across registers: joy prompts return melancholic ballads; swagger prompts return self-doubt; rage prompts return resigned acceptance; lust prompts return wistful longing. Each is the home-register output dressed in the prompt's subject matter.
The check
Every quality-of-craft system that scores or refines lyric output must operate per-register, not universally. The discipline: (1) Brief schema carries a typed ToneRegister field (joy / grief / rage / lust / swagger / awe / tenderness / melancholy / defiance / playfulness). The brief extractor surfaces it from prompt text; the default is unset (no register lean). (2) Forge prompt includes a per-register block when the field is set — explicit instructions for what the target register requires (active-verb density floor, tense dominance, scene-setting, body register, repetition discipline). (3) Negative anchor pool per non-default register — explicit "do NOT sound like this when register X is set" exemplars. The current positive-anchor corpus may itself be a negative anchor for non-introspective registers. (4) Critic voices reweighted per register — anti-cliché reader and prosodist get muted in joy/swagger mode; hook architect and vocal coach get boosted. (5) Anti-inflation rules per register — Gravity Rule / Burden of Proof / Antagonist Ceiling all carry register modifiers. The right rubric for grief is the wrong rubric for joy. (6) Per-register audit primitives — Active Verb Density (AVD), Body-In-Motion (BIM), Public Scene Density (PSD) — calibrated against per-register corpora per SA#30. (7) Per-register corpora — SA#30 ladder applies: corpus precedes calibration; calibration precedes detection. A joy-register positive anchor corpus is the calibration anchor for any "did the output land in joy register" detector. (8) Gauntlet register-aware — the refinement target shifts based on declared register; otherwise the gauntlet undoes the register. (9) Dashboard surface — the declared register is visible per song + the AVD or equivalent register-specific audit score is visible. When delivered output violated declared register, the gap is operator-visible. Relationship to SA#19: SA#19 ("A genre we cannot evaluate cannot be a genre we can serve") generalizes to register — "A register we cannot evaluate cannot be a register we can serve." Same discipline, different axis. Relationship to SA#28 / SA#29: emotional register is signal that must be plumbed as STATE per song, not parsed from prompt text on every forge call. Relationship to SA#30 / SA#31: per-register corpora (SA#30) + per-register pairwise calibration (SA#31) operationalize the discipline. Canonical phrasing: BRAND.sacredAccident32. Operator-facing arc: docs/WAR-ROOM-MAIN-CHARACTER-ENERGY-2026-05-27.md + docs/SACRED-ACCIDENTS.md (SA#32 section).
Surfaced: Build 3409 (Genre Catalog Clustering WAR ROOM). Operator submitted screenshots showing the /genres/indie/prompts page rendering 83 indie prompts essentially all medical / health themed ("Dream-Pop Radiation Therapy Planning", "Reverb-Heavy Hormone Replacement", "Bedroom Sports Medicine Rehabilitation"). The smoking gun was in src/lib/genre-prompt-catalog/arc-markers.ts: indie's INDIE_PACKAGE.mustInclude array literally contained the string "medical term" as a vocabulary anchor. The Sonnet generator was told this was a valid lane-lock anchor and dutifully used medical vocabulary in ~84% of generated titles. Quantified across 9 arcs: country 94% single-axis (Nashville / honky-tonk / pickup-truck / dirt-road cluster); worship 89% (Ancient / Sacred / Holy / Cathedral cluster); folk 86% (Stanza Study / Mechanics / Prosody Workshop instructional vocabulary cluster); rap 67% (corner-store / cypher / drill-sergeant cluster); rnb 60% (artist namedrops + Quiet Storm / Fender Rhodes cluster); rock 36% (highway / Glory Days / Rust Belt cluster); pop 31% (bathroom mirror / dance floor cluster — mildest); latin 21% (place-name cluster with regional dispersion). Each arc had its own clustering shape; the same architectural failure was responsible.
The anti-pattern this names
A single vocabulary list does double duty as both LANE LOCK ("the prompt MUST mention at least one of these") and SCENE ANCHOR ("the prompt should be vivid"). The generator dutifully satisfies the lane-lock rule by mentioning ANY of the list's entries — and since concrete scene anchors are easier to write than abstract lane-locks, the generator piles up on the same scene anchors across the batch. The indie generator chose "medical term" not because medical themes were the intent but because the operator had used the string "medical term" as a stand-in for "hyper-specificity" in the mustInclude list — and one literal string in a 28-entry list captured 84% of generated outputs. The same shape recurs across every arc: any vocabulary that does double duty BECOMES the clustering axis. Country's mustInclude was a country-cliché vault that became a country-cliché output cluster. Folk's mustInclude was Pat Pattison instructional vocabulary that became chorus-as-writing-workshop output. Worship's mustInclude was CCM-cliché vocabulary that became CCM-cliché output. Each arc trained its own dialect into its own catalog.
The check
Every system that generates content under a lane constraint must enforce the lane-vs-scene split structurally: (1) Lane-lock vocabulary (laneMarkers) is narrow + true — production markers, canonical artists, substyle names, structural craft vocabulary. ≥1 required per generated instance. List should not contain any "creative suggestion" phrasings — those leak to scene anchors. (2) Scene-anchor vocabulary (sceneRotationPool) is wide + diverse — ≥30 entries per arc spanning multiple domains (domestic / transit / work / school / public / nature / relationship / memory / body / identity). Generator picks ONE per instance and rotates across the batch so no single domain dominates more than 8%. (3) The two lists must not overlap (CI ratchet enforced via arc-markers.test.ts). (4) The coherence checker reads the LANE list, not the combined list. A prompt that anchors only on a scene without including a genuine lane lock fails the floor. (5) The generator's system prompt renders both as distinct sections, with explicit rotation discipline on the scene list ("never let one entry dominate > 8% of the catalog"). Operationalized at B3410 (ArcCraftPackage required new laneMarkers + sceneRotationPool fields; all 9 arcs populated; checkPromptCoherence switched to prefer laneMarkers) and B3411 (8 catalogs regenerated against the new generator in a single ~21 min Sonnet batch). Empirical dominant-axis clustering deltas: country 94% → 55% (-39pp); pop 31% → 19%; rap 67% → 52%; folk 86% → 84%; rock 36% → 32%; latin 21% → 7%; worship 89% → 80%. All 9 arcs at 100% coherence post-regen. The architectural split BROKE the clustering at the data layer. Relationship to SA#16: SA#16 catches the OUTPUT-side variant — the same single-axis clustering when the forge over-uses a single motif across many songs. SA#33 catches the INPUT-side variant — single-axis clustering in the seed-prompts that the forge is asked to generate from. Same craft truth at two pipeline positions. Relationship to SA#29: sister disciplines for content-generation pipelines. SA#29 says the genre signal must be plumbed as typed state, not re-parsed from text. SA#33 says the lane vocabulary must be split from the scene vocabulary, not piled into one list. Both are about preserving the distinction between structural constraints and surface coloration. Canonical phrasing: BRAND.sacredAccident33. Operator-facing arc: docs/WAR-ROOM-GENRE-CATALOG-CLUSTERING-2026-05-28.md.
Surfaced: Build 3412 (Concrete-to-Abstract Drift WAR ROOM). Operator pasted a multi-AI craft critique of three actual SongForgeAI outputs ("Speed Limit Thirty-Five" / "Coffee Stains and Crooked Lipstick" / "What I Keep"). The reviewer's diagnosis was unusually precise: the system observes brilliantly in verses, then thesis-summarizes in choruses and bridges. Concrete examples: "Henderson's mailbox leans" (strong) vs "Every reason that I had / Couldn't make me understand" (weak). "Called my sister from the parking lot" (strong) vs "What if this is everything" (weak). "The sweater fits in the donation bag" (strong) vs "I need to breathe more than remember" (weak). Same songs, verses succeed where choruses fail. The pattern appeared across all three songs in three sub-shapes: thesis-line openers ("Every reason", "What if this", "I'm done X-ing"); told-not-shown at emotional peaks ("rationing my happiness like it might run out" vs the observed slowing-from-25-to-20); mixed metaphors that sound right but fail close-reading ("the bruise you called my name" — a bruise can't be called a name; the abusive name LEAVES a bruise). The WAR ROOM identified SA#34 as the structural-level generalization of three single-arc disciplines already shipped (SA#23 line-level R&B AID rule; SA#25 indie hyper-specificity; SA#21 pop phonetic mass).
The anti-pattern this names
A craft system that has primitives for chorus SONIC quality (pop CHR / PVR, rock VPI) but no primitive for chorus SEMANTIC anchor — does the chorus OBSERVE or DECLARE? Without the second axis, choruses can sing well on the page and still feel like AI commentary on the song instead of the song itself. The failure is invisible to phonetic-mass scoring because the chorus achieves phonetic mass — it just doesn't inhabit a scene. The symptom shapes across registers: thesis-line openers in the chorus ("Every reason that I had", "What if this is everything", "I'm done X-ing"); told-not-shown emotional-peak declarations instead of action/image/detail ("Been rationing my happiness like it might run out" instead of the observed "slowing the car from twenty-five to twenty"); mixed metaphors that sound right but fail close-reading ("survival over shelter", "the bruise you called my name"). Each is the same craft failure dressed in different surface vocabulary: the chorus summarizes when it should observe.
The check
Every quality-of-craft system that scores or refines lyric output must operate at TWO chorus layers, not one: (1) PHONETIC MASS (SA#21) — the chorus must sound inevitable, ride the open-vowel structure, be singable for a non-native speaker on first listen. (2) SEMANTIC ANCHOR (SA#34) — the chorus must observe something specific the verse earned: a named action, image, or sensed detail. Operationalized as three primitives shipped B3413-B3414: (A) ATL (Abstract Thesis Line detector — heuristic stage flags chorus/bridge lines that match thesis-opener patterns ("Every X", "What if this", "I'm done X-ing", "I need to X", "I choose to X", "More than X / Instead of X", "Couldn't make me") AND have concreteness < 20%. False positives tolerable because R5-B1.2 Haiku judge will adjudicate the heuristic's output; (B) STDD (Show-don't-Tell Density — cross-arc generalization of R&B's NCD/AID rule. Runs on every chorus + bridge regardless of arc. Single floor at 0.30 cross-arc; substyle-specific weights queued as R5-B5+. Same algorithm as NCD but the load-bearing structural sections, not just verses); (C) EFI (Emotional Friction Index — Haiku-judged song-level question: does the song carry a SECONDARY emotion that complicates the primary? Joy with no fear of impermanence: fail. Departure with no lingering love: fail. Newfound maturity with no grief for time lost: fail.). The three primitives together close the failure-mode surface: ATL catches the chorus-line thesis pattern, STDD catches the cross-arc told-not-shown drift, EFI catches the one-note emotional arc. Relationship to SA#23: SA#34 is the structural-level generalization. SA#23 governs the LINE level (every emotional claim ships with action + image + detail). SA#34 governs the SECTION level (the chorus and bridge — the structural positions where AI summarization surfaces). Relationship to SA#25: SA#34 is the systemic statement of which SA#25 is the indie-specific instance. SA#25 was indie's name for the chorus-must-observe truth; SA#34 generalizes it across all 9 arcs as the cross-arc craft discipline. Relationship to SA#21: complementary, NOT competing. SA#21 governs the chorus as a SONIC object; SA#34 governs the chorus as a SEMANTIC-ANCHOR object. The two axes can both pass or both fail; a chorus is great when both are tight. Canonical phrasing: BRAND.sacredAccident34. Operator-facing arc: docs/WAR-ROOM-CONCRETE-TO-ABSTRACT-DRIFT-2026-05-28.md.
Surfaced: Build 3525 (KIDS-MUSIC "Tiny Genius Songs" WAR ROOM). Operator asked for a joyful children's educational album (ages 3-8, playful pop / ukulele, each song teaches one idea); the system produced adult literary singer-songwriter confessionals — a burnt-out teacher's existential crisis (#3 The Shape Zoo), a church-meeting scene (#9 Feelings Traffic Light), a stroke-rehab narrative (#5 Left/Right Shoe), a parent with "pressure charts and wind maps" (#6). It wrote ABOUT teaching children instead of FOR children. The Substitution Test quantified it: 9 of 10 tracks survived swapping the child-listener for a melancholy adult. Fixed B3526-B3530; ratified B3532 after the operator re-ran the album 2026-06-02 and confirmed "major deficits fixed!"
The anti-pattern this names
A quality-of-craft system that hardcodes a single LISTENER (here: an introspective literary adult) delivers that listener's song regardless of the brief — even overriding an explicit audience instruction. The failure is invisible to register obedience: the system can honor "joy" and still write for the wrong listener, because it renders the SUBJECT (children learning) as material for adult reflection rather than writing FOR the child who will hear it. The signature: an adult narrator observing children; adult interiority + settings (PT, church, pressure-charts) inside a kids' song; the lesson taught through adult metaphor (left/right via a stroke-rehab scene) instead of as the thing itself.
The check
Before declaring a quality-of-craft system general, ask: does the rubric know WHO the listener is — and would the song BREAK if you swapped them? (The Substitution Test, mirror of SA#20's city-apartment test.) Audience is the PARENT axis; register varies WITHIN an audience. Operationalized for the children's case across B3526-B3530: (1) register/audience telemetry on both forge paths so the signal is visible; (2) a children's forge discipline — "write FOR a child, not ABOUT one" + the 5 playable-action elements (repeatable hook, call-and-response, one silly image, one physical movement, one learning target) + a HARD VETO on adult interiority/settings/wounds; (3) the audience plumbed as STATE through the album path (concept-level detection forced onto every track) so it can't be re-parsed away from the bible's prose (SA#29); (4) a deterministic audience-appropriateness detector that makes the Substitution Test a measurement. Relationship to SA#32 ("register defines craft"): SA#35 is the PARENT axis — SA#32 says quality is per-register; SA#35 says the prior question is per-LISTENER, and register only varies once the listener is fixed. Relationship to SA#19 ("a genre we cannot evaluate we cannot serve"): the audience-specific instance — an audience we cannot evaluate we cannot serve. Canonical phrasing: BRAND.sacredAccident35. Full WAR ROOM: docs/WAR-ROOM-KIDS-MUSIC-2026-06-01.md.
Surfaced: Builds 3539-3543 (Constraint Fidelity + Opera/Language WAR ROOMs, 2026-06-02). Across one session the same gap recurred FOUR times: the album fan-out — which skips preforge — silently lacked banned-word extraction (B3539), structural-form variety (B3540), target-language capability (B3541), and a shared world canon (B3542), each a capability the single interactive Forge either had or gained. The operator's all-Italian opera "La Nona Campana" came back as English pop songs with anachronistic, renamed characters because none of language / world-canon / form / bans reached the album generator. Each fix was the same shape: extract the signal at the bible stage, thread it bible → generation fan-out → refine. Ratified B3545 with the check:album-capability-parity gate.
The anti-pattern this names
A pipeline with two entry points to the same core — here the single interactive Forge and the headless album fan-out — where the second is a stripped-down path that skips the first's preprocessing (preforge / brief extraction). Every capability added to the rich path is invisible to the stripped path by default. The gap is SILENT because the stripped path still produces output; it just quietly ignores the constraint (writes English instead of Italian, pop instead of opera, renames the cast). The signature: a capability demonstrably works on /forge but is absent on the album, and the absence is discovered only by an operator running a real album — never by a test, because no test asserted parity.
The check
Before declaring a Forge capability shipped, ask: does the ALBUM path carry it too — through the bible (extraction + plan_json STATE), the generation fan-out (the forgeAlbumTrack call), AND the refine re-forge? Capability parity is enforced by check:album-capability-parity (B3545): a canonical list of album-inherited capabilities (forbiddenLanguage, language, form, setting, characters) each asserted present at all three plumbing sites (album-track-forge.ts, album-generate-fn.ts, album-track-refine.ts). Adding a capability at one site without the others fails the gate; adding a new album capability means adding it to the canonical list + all sites. Relationship to SA#29 ("genre signal is plumbed as STATE, not parsed from TEXT"): SA#36 is the META-level sibling. SA#29 governs HOW a signal travels (typed state, not prose) — it is per-signal, within one path. SA#36 governs WHETHER a capability reaches the second entry point at all — it is per-capability, across entry points. SA#29 is the plumbing technique; SA#36 is the parity obligation. Canonical phrasing: BRAND.sacredAccident36. Full WAR ROOMs: docs/WAR-ROOM-CONSTRAINT-FIDELITY-2026-06-02.md + docs/WAR-ROOM-OPERA-LANGUAGE-2026-06-02.md.
Surfaced: Build 4040 (Song-Novel, operator-caught, 2026-06-29). A song-novel forge produced beautiful prose with >=4 hard continuity contradictions (narrator "fourteen" in T1 vs "twelve" in T10; at the willow vs "back at Mama's" in T7, violating the bible's own forbiddenContradictions; a mis-voiced T9). The operator ran the coherence check and got "Coherence 100/100, verdict na." The `na` was the tell: the verdict can only be `na` when the audit DID NOT RUN — the 20s Haiku timeout fell through to the hard-coded no-penalty default of score:100, and the UI rendered that non-result as "the tracks hold together as one work." A FAILED check wore the costume of a clean pass — and a clean pass tells the operator to ship.
The anti-pattern this names
A detector that returns a benign default (score 100, "looks fine") on error/timeout/parse-failure, displayed identically to a real PASS. Every detector has three outcomes, not two: PASS, FAIL, and DID-NOT-RUN. The third is the dangerous one, because the safe internal default ("don't penalize a thing we couldn't judge" -> score 100) is indistinguishable at the UI from a genuine pass unless the code deliberately branches on a state flag. The failure is silent and self-concealing: the worse the detector's reliability, the MORE often it shows a perfect score (every timeout becomes a 100) — the exact inverse of what a quality gate should do.
The check
Every detector must (1) carry a distinct STATE, not just a score (judged / na / error), and the UI must BRANCH on it — a non-judged result renders "NOT CHECKED" + the reason, never a score, never a verdict, never a reassuring sentence; (2) keep the no-penalty default (score 100) as an INTERNAL non-blocking value only, never a user-facing signal; (3) size reliability (timeout, retries) so the COMMON case actually judges — at B4040 the coherence timeout went 20s->45s, maxRetries 1->2, + one parse-retry; (4) add a DETERMINISTIC backstop where a fact is mechanically checkable, so there is nothing to time out (B4041 auditFactContradictions catches the age contradiction with no model); (5) anchor a ground-truth corpus so the false result cannot silently return (B4042 coherence-corpus.ts, the album that exposed this is anchor #1). Relationship to SA#30 ("calibration precedes detection"): direct sibling — SA#30 governs whether a detector's judgment can be TRUSTED; SA#37 governs whether a detector's NON-judgment can be MISTAKEN for one. SA#19 ("a genre we cannot evaluate cannot be a genre we can serve") is the upstream principle; SA#37 is what "cannot evaluate" must LOOK like in the UI at runtime. Canonical phrasing: BRAND.sacredAccident37. Full arc: docs/SONG-NOVEL-WAR-ROOM-2026-06-28.md (B4040-B4042).