NOSTALGIA_RUBRIC.md

music/NOSTALGIA_RUBRIC.md

THE NOSTALGIA RUBRIC — the exemplar-anchored bar, as scored axes

CANON SUBORDINATION — this document is PROPOSAL-TIER: it serves canon and never outranks it.

The canon: the CVD · the T1 foundation docs · the T0 registries (registries/) · the spine

(docs/spine/CH_*.md) · the region pages (_source/02_Tier_2_Region_Pages/). Authority order: docs/DOC_MAP.md § 0.

Canon served: NONE FOUND — this document cites no canon anchor anywhere in its body.

Do not build from it alone: open the spine node (docs/spine/CH_NN.md), the region page

(_source/02_Tier_2_Region_Pages/), and the T0 registry your work targets, and derive from those.

READ THAT CANON FIRST — open it and derive from it before you build anything from this document.

If this document disagrees with canon, CANON WINS and this document is the defect — fix the

document, never the canon. Nothing here is applied until it is ratified into canon.

Status: ACTIVE (2026-08-05). Instrument: harness/music_gen/nostalgia_score.py (self-test

exit-code-bearing, 25 controls). Census: harness/music_gen/nostalgia_census.md. Refusal

fingerprints: harness/music_gen/exemplar_hook_signatures.json.

---

0. The ruling, and the defect it named

Josh, 2026-08-05, verbatim and binding:

"Youve failed hard on the music from what I heard on the website. This is nothing like all the

major nostalgia music tracks from Super Nintendo and Nintendo 64, legend of Zelda oot, super

Mario 64, and the hundreds of other nostalgia tracks."

The fourteen authored themes on the site were selected on proxy metrics — distinctness, motif

recurrence, a hummability proxy — with no comparator anchor against the exemplar class. That is

the defect, and it is a measurement defect before it is a composition defect. feature_rig.py

measures whether a signal carries a melody *at all*; it never once asked whether that melody is in

the dialect. A rig that scores a theme against itself will always report success, and it did.

This rubric is the gate-51 pattern applied to music: the mechanism stated verbatim from the school

grammars, the measurable check that mechanism implies, tied to an instrument that exists, and the

honest declaration wherever no instrument exists.

THE SCHOOL GRAMMARS, AND A CORRECTION TO THIS SECTION (2026-08-05, verifier finding F3). The

sentence above was true about the METHOD and false about the CORPUS at the moment it was written.

Every rule id cited in the tables below — MELODY-01, ECON-04, PHRASE-04, HARMONY-05,

FORM-05, LOOP-06, RHYTHM-01, VOICE-01 and the rest — resolved to exactly one file in this

repository, and that file was this one. The grammars themselves existed only inside a workflow run

journal, which is agent scratch that no later reader can open: the rubric was quoting an authority

layer that was not on disk. A cold verifier found it by positive-controlled grep, and it is the same

class as the invented-rule defect this repo withdrew in commit 7e5494d6.

WHAT CHANGED: nothing in the rules, the bands or the instrument. The three research returns were

extracted from the run journal VERBATIM — mechanically, by script, unedited — and landed as

proposal-tier documents, so every citation now resolves to a readable rule:

• docs/proposals/music/school_grammars/KONDO_SCHOOL.md — MELODY-*, PHRASE-*, HARMONY-*,

FORM-*, RHYTHM-*, VOICE-*, PROCESS-* (43 rules, 20 sources, 43 features).

• docs/proposals/music/school_grammars/SNES_JRPG_SCHOOL.md — LEIT-*, ECON-*, HARM-*,

AFF-*, LOOP-*, TOOTH-* (44 rules, 25 sources, 50 features).

• docs/proposals/music/school_grammars/EARWORM_LOOP_SCIENCE.md — R1–R14 (14 rules, 22

sources, 33 features).

All 49 rule ids cited in this document resolve into that set, each to exactly one file, and the

resolver was positive-controlled (an id that is deliberately absent reports absent). What is still

NOT closed is verifier finding F4: several sources in those documents are bare author-year strings

that a reader cannot resolve from this repo. The grammars are now readable; their scholarship is

still unaudited, and each one says so on its own face.

What did NOT change with the ruling. PIPE_AUDIO_MUSIC_2026-07-29.md remains the lane's law for

stack, licence and records. The 30-year bar, the scoring law (per-culture period instrumentation,

mostly wordless) and the melody-first doctrine all stand. This rubric does not replace them; it is

the ruler that was missing underneath them.

---

1. The three structural laws

LAW 1 — EVERY BAND NAMES ITS EXEMPLAR. A band with no named track behind it is a proxy, and a

proxy is what shipped fourteen themes. Every row in every table below carries the rule it enforces

and the exemplar the number came from. Where a floor was *derived* rather than cited (there is

exactly one, I0), it says so in its own row rather than borrowing someone else's citation.

LAW 2 — NO SINGLE SCALAR, ANYWHERE. PROCESS-01: *"REPORT PER FEATURE, never as a single scalar —

the collapse to one score is what let 14 unanchored themes ship."* The artifact schema has no

overall key and the emitter refuses to mint one; --self-test check S14 asserts its absence. A

cue is described by a seven-vector plus a hard-fail list. A reader who wants one number has to decide

for themselves what to trade away, in the open.

LAW 3 — WHAT CANNOT BE MEASURED IS DECLARED, NOT DROPPED. Nine features are NOT_COMPUTABLE from

the symbolic head-cell representation, each carrying the reason. They are **excluded from the axis

denominator** rather than silently counted as passes — check S15 asserts exactly that. This is the

same discipline feature_rig.py already applies to the doctrine's human-listener gates.

---

2. The anchor set

Every band is grounded against this named track set, and the census reports distance per feature

rather than a centroid distance:

SMB Ground Theme · SMB3 Overworld · SMW Athletic · SM64 Main Theme · SM64 Dire Dire Docks ·

LoZ Overworld · ALttP Hyrule Castle · ALttP Kakariko Village · OoT Title · Zelda's Lullaby ·

Saria's Song · Song of Storms · Great Fairy's Fountain · Hyrule Field · Gerudo Valley ·

MM Astral Observatory · MM Song of Healing · WW Grandma's Theme.

---

3. How an axis is scored

• Each feature returns PASS, FAIL, or NOT_COMPUTABLE.
• Axis pass rate = passes / (passes + fails). NOT_COMPUTABLE never enters the denominator.
• Axis verdict: EXCEEDS at >= 0.85, MEETS at >= 0.65, otherwise BELOW.
• Any HARD feature failing forces the whole axis to BELOW, whatever the rate. Seven features are

hard: H1 H4 C5 C6 P1 Y1 L1 L2.

• Axes are never averaged against each other. A theme that EXCEEDS on hook and is BELOW on loop form

is exactly that, and the recompose gets told which one to go fix.

---

4. The seven axes

Tables are the instrument's own feature set. If a row here and a row in

harness/music_gen/nostalgia_score.py disagree, the code is right and the drift is a defect —

self-test check S16 asserts the two sets are identical in both directions.

AXIS 1 — HOOK PLACEMENT + STRENGTH

*Does the identity arrive, early, as a thing you could carry out of the room?* The Library of

Congress essay locates the SMB Ground Theme's identity in "the first six notes that greet the

player"; Ferguson's whole catalogue indexes Kondo cues by opening interval, because the opening

is the identity. A four-bar atmospheric run-up before the tune is a defect in this dialect.

AXIS 2 — CONTOUR SINGABILITY

*Is the shape the ordinary one, and can a person reproduce it with no notation?* Jakubowski et al.

(2017) found earworm melodies carry more common global contours than non-earworms against a large

pop corpus — the plain rise-then-fall arch is a positive predictor. Novelty belongs in the interval

detail and the harmony, never in the contour. C5 is the hard hummability test that replaces the

failed proxy.

**FINDING 2026-08-07 — C3 AND THE CARD'S the_hook.range_semitones ARE NOT THE SAME OBJECT, AND

A CUE CAN PASS ONE WHILE FAILING THE OTHER FOR NO MUSICAL REASON.** C3 bands the AUTHORED HEAD

CELL at 7 to 17 semitones — a singable span for a monophonic tune a player could whistle, grounded

in the ocarina-song class. The formula card's the_hook.range_semitones measures something else

entirely: the span of the PREDOMINANT PITCH TRACK extracted from a finished polyphonic mix over the

card's hook window, which wanders across whichever voice is loudest. Its median across the 30 corpus

rows is 22 semitones, which is outside C3's band — so a literal reading would fail most of the

music this rubric was built from, on a measurement C3 never intended to make.

Neither number is wrong. They are different objects wearing one word, and nothing in the rubric or

the card said so, which is how a future lane would have "fixed" a passing cue to satisfy a band that

does not apply to it. **The rule going forward: C3 is scored on the AUTHORED head cell only, via

harness/music_gen/author_head_cell.py, and is never read off a card.** The card's

the_hook.range_semitones is a descriptive mix statistic and is not a rubric feature; where it is

printed beside a rubric score it must be labelled as the extracted-line span. Note C18 in

theme_compositions.py already checks the head-cell law's own words against a ninth, on the

authored notes, which is the correct site for this class of check and the model C3 should follow.

Evidence and the corpus medians: docs/proposals/music/MUSIC_CRAFT_DOCTORATE.md §4.

AXIS 3 — INTERVAL VOCABULARY

*Are the intervals the dialect's intervals, and is there exactly one fingerprint leap?* Uematsu

centres the FF7 theme on a major 7th that recurs across the soundtrack — one characteristic wide

interval, deliberately placed, inside an otherwise conjunct line. I1's band is the **declared

intersection** of two rules that disagree: MELODY-05 asks for 0.25-0.45 and ECON-05 implies <= 0.40.

The rubric takes 0.25-0.40 and says so rather than picking a favourite silently.

AXIS 4 — PHRASE ARCHITECTURE

*Does the 8-bar unit parse as a Caplin sentence — basic idea, near-literal repeat, continuation by

fragmentation, cadence?* Zelda's Theme repeats its basic idea at mm. 3-4 and again at mm. 11-12;

"Serenade of Water" repeats mm. 1-2 in mm. 3-4. **Development belongs in the continuation, never in

the statement**, and P1 is hard for that reason. P1b exists because the landed cohort answers its

head with a *mirror* rather than a repeat: the rubric prints that as a named deviation instead of

letting it read as a silent miss.

AXIS 5 — HARMONIC SIGNATURE

*Does the harmony do the four things this school does — prolong tonic over a held bass, borrow from

the parallel mode, substitute for the dominant by voice leading, and refuse to close?* Gervais's

four-step phrase model is the spine: tonic prolongation → subdominant → a chromatically altered

chord where the dominant is expected → a weakened cadence on a major tonic. Y1 is the single

most load-bearing hard gate in the rubric: a fully-closed cadence at the wrap point is the most

common way a 90-second loop becomes unbearable.

AXIS 6 — LOOP FORM

*Can the loop survive being heard a hundred times?* This is Kondo's own answer, verbatim, to the

question of how he avoids sounding repetitive: **"there are four motifs to the music and each of

those motifs is unique. They don't sound like the other motifs, and that's why it doesn't sound too

repetitious."** A loop that rides one motif for 90 seconds fails by construction, no matter how good

the motif is — and L1 says a themed cue cannot carry three or four contrasting motifs inside 8 or

16 bars, so the length is a hard gate too. L4's band is the only band in this document taken from a

corpus statistic rather than a named track: Hunt's 21,391-piece survey, mean 62.79%, SD 3.49, flat

across eleven console generations.

A NAMED GAP, carried rather than guessed. No source gives a measured distribution of loop lengths

in seconds for the exemplar class; the only figures available are practitioner conventions of

non-citable quality. L1 is therefore stated in bars, grounded on Zelda's Theme's 24-bar binary.

Any seconds-based gate must be derived by direct measurement of the named exemplar tracks, and until

that measurement exists this rubric does not have one. The census reports loop seconds as data, not

as a scored feature.

AXIS 7 — VOICE ECONOMY

*Is the texture thin on purpose?* The SMB Ground Theme was written for two pulse channels, one

triangle and one noise channel, with Kondo alternating melodies between the three melodic

channels rather than doubling them — and Ocarina of Time's engine had four sequence players, sixteen

channels each and sixty-four preallocated layers, and the writing still stays thin. **The thinness is

an aesthetic choice, not a hardware ceiling**, and modern orchestral density is the fastest way to

lose this dialect. Five of this axis's nine features cannot be seen from a symbolic sketch, and they

say so.

---

5. The anti-plagiarism tooth

TOOTH-01, which is the whole licence position of this lane in one sentence: **grammar is learned;

melodies are never reproduced.** Cadence formulas (bVI-bVII-I, I-bVII-bVI-V), modes, ostinato

figures, sentence structure and loop forms are not protectable and are the intended borrowings.

Actual tunes are never copied, and the screen is what makes that a measurement instead of a promise.

How it works. Every candidate's melody is reduced to an interval sequence — transposition

invariant by construction, so a quote in any key is visible — and screened against a reference set of

exemplar hook signatures held **as interval sequences only: no audio, no absolute pitches, no

rhythms, no lyrics.** It is a refusal fingerprint in exactly the sense a plagiarism checker's hash

database is, and it exists to REFUSE. It is never a generation input, and nothing in the pipeline

reads it as one.

Two refusal rules, both needed:

• LENGTH RULE — a contiguous run of >= 6 matching intervals (7 notes) against any reference.
• COMPLETE-SIGNATURE RULE — a match covering 100% of a reference signature of length >= 4. The

diegetic player-performable class is six notes long, so its entire identity is five intervals; a

length rule alone could never see it, **and that class is precisely the class this game also

builds.**

Inversion and retrograde hits are advisory — printed, never a refusal. The false-positive cost of

refusing an inverted five-interval run inside a stepwise dialect is real, and a noisy tooth conceals.

MUTATION-PROVEN BOTH DIRECTIONS (--mutation-proof, and the same checks run inside --self-test):

SEED-SET HONESTY. The reference set is 13 signatures and a declared gap list, not a census.

Every entry carries a confidence and a derivation. Fourteen anchor-set exemplars are **declared

absent** — LoZ Overworld's main phrase, SM64's main theme, Dire Dire Docks, SMW Athletic, SMB

Underwater, SMB3 Overworld, both ALttP cues, OoT Title, Hyrule Field, Gerudo Valley, Great Fairy's

Fountain, MM Astral Observatory, WW Grandma's Theme — because no reliable interval transcription was

available and a fingerprint invented from memory is a false negative wearing a green tick.

Therefore: CLEAR means "no seed signature matched", and never "clean". Expanding the seed set by

direct transcription of the named exemplars is the next measured rung, not a formality.

---

6. The census — the baseline the recompose must beat

harness/music_gen/nostalgia_census.md (script-emitted; re-run and read the diff). Scored

2026-08-05, the first time the landed fourteen have been measured against the exemplar class rather

than against themselves. Reported as found:

Best in cohort: COMPANION_HEALERS_CHILD at 4 of 7 axes. Eleven of the fourteen sit at 1 or 2.

Six features fail across the entire cohort or all but one of it, and they are the recompose's

worklist rather than a scolding:

• L1 loop_length_bars — 0/14. Every landed theme is an 8- or 16-bar sentence. The exemplar bar

is 24-32. This one feature makes L2 unreachable by construction: three to four mutually-unlike

motifs do not fit in eight bars.

• Y3 mixture_chord_rate — 0/14 and Y8 bVII-for-V — 0/14. The cohort's harmony is

diatomically clean, which is precisely the thing HARMONY-05 and HARMONY-11 say this dialect

never is. Y7 (a house cadence present) is 1/14.

• I1 leap_fraction — 0/14 and I3 signature leaps — 1/14. The heads are either purely

stepwise or triad-outlining; the one licensed wide interval the briefs specify is not landing as a

measurable fingerprint in the statement.

• Y6 tonic_prolongation_first4 — 1/14. The opening four bars move functionally instead of

holding a tonic field over a common bass.

• P1 basic-idea literal repeat — 3/14 (HARD). The cohort answers its head with a *mirror*

(P1b is high where P1 is low). That is a coherent compositional choice and it is not this

dialect's — the exemplar restates before it develops.

• C6 rest_density — 3/14 (HARD). The lines are close to wall-to-wall. Kondo's own stated

method makes the silence structural.

Where the cohort is genuinely strong, and should not be thrown away in the recompose: **hook

placement (12 of 14 at MEETS or better) and voice economy (11 of 14)**. The heads arrive early,

sit in small spans, and the textures are thin and openly voiced. The failure is not the head cell —

it is everything that was supposed to happen after it: length, motif count, harmonic colour,

and the answering phrase.

Anti-plagiarism: 14 of 14 CLEAR, against a 13-signature seed set with 14 declared gaps.

---

7. Consumers

• harness/music_gen/nostalgia_score.py — the instrument. Self-test is exit-code-bearing and wired

into the standing suite as gate nostalgia_rubric.

• harness/music_gen/feature_rig.py — the audio-side rig. It measures whether a signal carries a

melody; this rubric measures whether that melody is in the dialect. **Neither one alone is the

bar** and the two are reported side by side, never merged.

• harness/music_gen/theme_compositions.py + author_head_cell.py — every future head cell is

scored here before it is rendered, and a candidate that cannot clear the hard features does not

get orchestration spend.

• harness/music_gen/ab_judge.py / emit_authored_picks.py — TOOTH-02: every A/B runs the current

theme as the losing control plus an exemplar-class anchor. A candidate that beats only the

incumbent has not cleared the bar.

• docs/proposals/music/COMPOSER_BRIEF_PACK.md and LEITMOTIF_ARCHITECTURE.md — the briefs specify

the shape; this rubric is what checks the shape arrived.

• docs/pipeline_review/tech_research/PIPE_AUDIO_MUSIC_2026-07-29.md — remains the lane's law for

stack, licence and records. This rubric sits under it, not over it.

• docs/proposals/MUSIC_COMPOSITION_DOCTRINE.md — the hummability checklist's computable half now

has an exemplar anchor; the human-listener gates are unchanged and still unbought.

---

8. How to run

python harness/music_gen/nostalgia_score.py --self-test        # 25 controls; exit 2 on any failure
python harness/music_gen/nostalgia_score.py --mutation-proof   # both teeth, both directions
python harness/music_gen/nostalgia_score.py --census           # score the landed 14, emit both files
python harness/music_gen/nostalgia_score.py --theme PROTAGONIST_THEME
python harness/music_gen/nostalgia_score.py --control CONTROL_HOOKLESS
python harness/music_gen/nostalgia_score.py --screen-intervals 0,0,-4,4,3,-12

The self-test is the claim. Three controls carry it: a composition built to the rubric (must

score 7/7 with no hard fail), a texture control (a held pad — must go BELOW on 6 of 7), and a

deliberately hook-less control (a real melody with real harmony and a real cadence that is out of

the dialect on every axis — must go BELOW on 6 of 7 and trip at least three hard gates). **A rubric

that cannot fail a bad melody is not a rubric, it is a formality**, and the hook-less control is the

one that proves it is not.

Generated by harness/site/structure_site.py — the URL path is the repo path. review root