MELODIC_CRAFT_PRACTICE.md

music/MELODIC_CRAFT_PRACTICE.md

MELODIC CRAFT PRACTICE — THE MEASURABLE PROPERTIES OF MELODIES PEOPLE REMEMBER

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

Canon served: the composition floor at docs/spine/DECISIONS_PENDING_JOSH.md — MUSIC ROUND 1

GRADED + THE COMPOSITION FLOOR and its ROUND 2 GRADED successor; the melody-first / 30-year-bar

music direction (memory: music-direction-melody-first-30-year-bar); the theme rows of

registries/T0_Theme_Registry [DRAFT v0.1].

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

Tier: RESEARCH SYNTHESIS, proposal-tier. It contains no canon and rules on nothing.

Every claim about published craft carries a URL or a book, author and chapter. Every claim about

our own stack carries a file and a function, field or line. Claims are labelled **CRAFT

CONSENSUS** where practitioners and theorists agree without having measured anything, and

EVIDENCE where somebody measured something and published the number. The two are not

interchangeable and this document does not dress one as the other. A third label, **MEASURED

HERE**, marks numbers this document produced by running our own scorers — those are reproducible

from §5 and should be re-derived, not trusted.

CITATION AUDIT PASSED WITH REPAIRS — 2026-08-08. See §11 before citing anything from §4.

An independent pass fetched and re-read the source set. The published-corpus numbers in §3 came

through clean, cell for cell. Two fabricated claims and three misattributions were found — all of

them in the citation apparatus or in §4's game-practice prose — and all are repaired in place with

the original error stated rather than quietly overwritten.

Sibling volumes. MOTIVIC_DEVELOPMENT_AND_PHRASE_GRAMMAR.md covers how a tune develops;

HOOK_AND_TENSION_ARCHITECTURE.md covers where the hook sits in a cue;

ORNAMENT_AND_MELODIC_SURFACE.md covers the divided surface. This volume covers the one question

those three assume answered: what does a good melodic line actually look like, in numbers.

---

0. The defect this was written against

Josh graded our melodies as *a few notes slowly played back to back*. That is not a vague

complaint — it is a precise one, and it names four measurable properties at once:

• a few — small note count and small pitch-class alphabet inside the thematic unit;
• notes — undifferentiated events, not a line with a shape;
• slowly — low onset density;
• back to back — sequential adjacency with no phrase, no arrival, no return.

Our two named melodic files are harness/music_gen/pass3_grammar.py (which builds the phrase

objects) and harness/music_gen/instr_melodic_intelligence.py (which scores the result). Between

them they measure thirteen things — METRIC_KEYS at instr_melodic_intelligence.py:1767-1772 —

and not one of them is note density, pitch range, or climax placement. The grade Josh gave

names three properties our instrument cannot see. That is the gap this document exists to close,

and §8 states it as findings the audit phase can consume.

The headline synthesis, stated before the evidence so the evidence can be checked against it:

**The cure for "a few notes slowly played back to back" is not more notes in the structural

line. It is a two-level line** — a slow structural skeleton carrying a fast divided surface —

plus a placed climax and a cadence. Schoenberg says exactly this (§2.3), the corpus

numbers are consistent with it (§3.1), and pass3_grammar.divide() already builds it while

instr_melodic_intelligence never checks whether it happened (§8, GAP-03).

---

1. SOURCES

Read-status is declared for every item, because a source consulted through a secondary summary is

weaker evidence than one read in full, and a research document that hides the difference is

lying by formatting.

1.1 Books and industry manuals

1.2 Peer-reviewed papers and corpus studies

1.3 Talks, studio practice, and published analyses of game themes

1.4 A declared hole in the source set

No published, numerically-grounded melodic analysis of a canonical game theme was obtainable.

Hooktheory (S23) has the note data and blocks automated access; the academic ludomusicology

literature (S7, S20) analyses harmony, mode and function rather than melodic statistics. Everything

this document says about Zelda, Final Fantasy and Dragon Quest is therefore **qualitative and

sourced to prose analysis, and §5 substitutes a public-domain reference set** — including one

melody that *is* a beloved game theme — so that the numeric side of the argument rests on melodies

this document could actually measure. A future pass with legitimate access to symbolic

transcriptions should redo §5 on the real targets.

---

2. THE CRAFT: what the composition manuals actually require of a melody

2.1 Schoenberg's five conditions on a melodic line (S1, ch. IV, "Connecting Motive-Forms")

Schoenberg states the requirements of a well-balanced melody in a single compact passage. Rendered

as conditions rather than quoted:

1. It progresses in waves — every rise is answered by a fall.

2. It approaches its high point through lesser high points, interrupted by recessions. The

climax is *arrived at*, never simply present.

3. Upward motion is balanced by downward motion.

4. Large intervals are compensated for by conjunct motion in the opposite direction. This is

the manual's version of what psychology later called post-skip reversal (§3.2).

5. It stays within a reasonable compass, not straying far from a central range.

Conditions 1–3 are contour conditions; 4 is an interval-order condition; 5 is an ambitus condition.

Our instrument measures none of the five directly. contour_depth

(instr_melodic_intelligence.py:480) measures hierarchy of a contour, not its shape; nothing

measures ambitus or wave balance at all. This is GAP-01 and GAP-02.

2.2 The motive is a shape, and repetition without variation is the named failure

Schoenberg (S1, ch. III) defines the motive as intervals and rhythms combined into a memorable

shape or contour, which usually implies an inherent harmony. He then states the mechanism that

matters to us: a motive appears constantly and is therefore repeated; **repetition alone gives rise

to monotony, and monotony is overcome only by variation.** Variation means changing the

less-important features while preserving the more-important ones — change everything and you get

something foreign and incoherent; change nothing and you get monotony. He names the preservation of

rhythmic features as the cheapest way to buy coherence.

The corollary he draws in ch. VII is the operational one, and it is stated in capitals in the

source: preserving the rhythm permits extensive change to the melodic contour. In slow tempo

this allows far-reaching contour change; in fast tempo it protects comprehensibility.

CRAFT CONSENSUS. This is the same doctrine as pass3_grammar.OPERATIONS' spine_changing

flag, from the opposite direction: our flag asks whether the interval spine changed; Schoenberg

asks whether the *rhythm* survived. A variation that changes the pitches and keeps the rhythm is

the canonical developing variation. A variation that keeps the pitches and changes only the

register or instrument — six of round 2's ten transforms — is what he means by repetition.

2.3 Density rises into the climax — the passage that answers Josh's grade directly

In his analysis of Beethoven's Op. 2/1-II (Adagio) (S1, ch. VI, "Analysis of Periods from

Beethoven's Piano Sonatas"), Schoenberg makes an observation that is directly measurable and

directly on the point of Josh's complaint: the melodic approach to the climax in bar 6 is supported

by more frequent changes of harmony and an increasing use of small notes, and the downward

movement after the climax is equally significant. *(Correction 2026-08-08: this passage was

previously attributed here to Op. 2/2-IV. It is not — in the source the Op. 2/2-IV paragraph begins

immediately after this one and says something else entirely, about a caesura on V and a cadence in

the dominant. The observation itself is verbatim as stated.)*

And in ch. VII, on the cadence contour: if there is a climax, the melody is likely to recede from

it, balancing the compass by returning to the middle register — this decline, combined with

harmonic concentration and the liquidation of motivic obligations, is what delimits the structure.

This is the two-level line. The structural pitches may move slowly; the *surface* accelerates

into the climax and thins after it. "A few notes slowly played back to back" is precisely a line

that has the structural level and never grew the surface level. Schoenberg's remedy is not "write

more structural notes" — it is "divide the surface, and place the division where the climax is."

pass3_grammar.divide() (pass3_grammar.py:542) and _nct_figure() (:591) already implement

exactly this surface. **Nothing schedules the division against the climax, and nothing scores

whether it happened.** GAP-03.

2.4 Intelligibility requires repetition; the classical masters sought a popular touch

Two further Schoenberg positions matter for a game score, both from ch. V (verified by chapter

offset in the scan; this section previously said ch. IV):

• Rapidity obstructs grasp. Pieces in rapid tempo exhibit *less* variety, deliberately; the

controls are delimitation, subdivision and simple repetition. A fast battle cue should carry

*fewer* distinct ideas than a slow one, stated more often — the opposite of the instinct that

says a busy cue needs busy material.

• Intelligibility is impossible without repetition, and only as much variation as character,

length and tempo require should be admitted. He explicitly notes that discretion of this kind is

most necessary where immediate intelligibility is the goal, and that the classical masters

themselves sought a "popular touch" in their themes — citing Mattheson's 1739 formulation that a

theme should contain something the whole world already knows.

That last position is the theoretical bridge to §3.5. Schoenberg in 1948 and van Balen's

regression in 2015 say the same thing: a memorable theme is partly conventional on purpose.

Our melodic_complexity axis has a variety half that rewards the opposite. GAP-05.

2.5 Caplin's phrase functions, as our grammar already implements them

pass3_grammar.py implements Caplin (S2) faithfully — FUNCTION_NAMES at :131, THEME_TYPES at

:134, the cadence-strength ladder at :125-130, and builders sentence() (:992), period()

(:1093), hybrid_ant_cont() (:1193). The doctrine, restated only where §8 needs it:

• Sentence = presentation (basic idea + its repetition) → continuation → cadential. The

continuation is defined by fragmentation (units get shorter) and harmonic acceleration

(chords change faster).

• Period = antecedent ending on a weak cadence, answered by a consequent ending on a strong

one. The antecedent's basic idea is juxtaposed against a *contrasting* idea; in a sentence it is

*repeated*.

• Liquidation (S1 ch. VIII; S24) = gradually eliminating characteristic features until only

uncharacteristic ones remain, which no longer demand continuation. It is generally supported by a

shortening of the phrase.

The grammar has all of this. The scorer has none of it — it re-derives "phrases" from rest gaps.

GAP-06 and GAP-07.

---

3. THE EVIDENCE: what people have actually measured

3.1 Note density — the number Josh's grade is about

**EVIDENCE (S15, BiMMuDa, 366 manually transcribed lead-vocal melodies from the US top-5 of each

year 1950–2022).** Onset density, defined as notes per second, by era:

The paper also reports that in the most recent era **two-thirds of pitches lie within a

5.5-semitone range** — i.e. a pitch standard deviation of roughly 2.75 semitones — and that both

pitch-related and rhythmic information-theoretic complexity fell across the period, with revolutions

identified at 1975–76 and 2000–01 (tier-1 evidence) and 1996 (tier-2).

Two things follow, and the second is the one that stops this becoming a bad gate.

• A pop-idiom melodic line lives around 1.8–2.8 notes per second. Our voice-detection floor,

VOICE_MIN_RATE = 0.30 notes/s (instr_melodic_intelligence.py:153), sits **six to nine times

below** that. A line at 0.4 notes/s — unmistakably "a few notes slowly played back to back" —

passes our voice gate and is then never scored on density anywhere, because density is not in

METRIC_KEYS. GAP-01.

• Density is genre-conditional and a flat floor would be wrong. §5 measures Londonderry Air at

0.74 notes/s and Scarborough Fair at 0.66 — both far below the pop norm, both unimpeachable

melodies. A lyrical theme legitimately runs slow. The rule must therefore be **conditioned on the

cue's own idiom and stated at two levels** (structural vs surface), which is exactly rule R-03.

3.2 Steps, leaps, and step inertia — with real corpus counts

EVIDENCE (S11, Chiu & Temperley 2024, Table 1). Probability that a step is followed by another

step *in the same direction*, with a step defined as ≤ 2 semitones, pitch repetitions removed:

With the step size widened to 3 semitones to accommodate pentatonic pop, the folk and classical

figures fall to 60% and 59% while pop is unchanged at 47% and 42%.

This is a style discriminator, not a quality bar, and that is the important reading. Folk,

classical-theme and hymn melodies *continue* through steps; rock and pop melodies *reverse* through

them. A scorer that hard-codes one target imposes a style. Our

melodic_complexity does exactly that with its step-to-leap term (§3.3).

The related Huron universals (S4, via S11 and S10) are CRAFT CONSENSUS supported by corpus work:

pitch proximity (melodies favour small intervals), step declination (descending steps outnumber

ascending), step inertia, the melodic arch, and post-skip reversal / gap fill (a large interval

tends to be followed by a change of direction). Chiu & Temperley note Huron's own finding that step

inertia may hold mainly for *descending* steps. Schoenberg's condition 4 (§2.1) is post-skip

reversal stated as a compositional instruction two decades before it was measured.

3.3 What our own step-to-leap term does with those numbers

MEASURED HERE. instr_melodic_intelligence.melodic_complexity (:496) counts steps as

absolute intervals of 1–2 semitones (n_step) and leaps as ≥ 5 semitones (n_leap), forms the

ratio, and scores it as 1 − |log10(S:L) − log10(3.0)| / 1.3 — a target of S:L = 3.0, i.e. 75%

steps among counted motion. Evaluating that term:

Measured S:L on real melodies (§5): Londonderry Air 8.5, Korobeiniki 11.0, Greensleeves 19.0,

Ode to Joy 20.0. Meanwhile the deliberately-bad control — eight slow notes, the Josh failure

mode — scores S:L 4.0 and therefore 0.904 on this term.

**Our instrument gives "a few notes slowly played back to back" a step-to-leap score of 0.904 and

gives Ode to Joy 0.366.** That is not a calibration wobble; the term is inverted relative to

practice over the range real melodies occupy. GAP-04.

Two causes, both fixable:

1. The dead zone. Steps are 1–2 semitones and leaps are ≥ 5, so **thirds (3–4 semitones) are

counted in neither.** Thirds are 17–30% of intervals in the §5 reference melodies. Excluding the

most common non-step interval from both sides of a ratio makes the ratio meaningless.

2. The target. 3.0 was chosen as "the one classical rule about order rather than variety"

(the function's own docstring). No corpus supports 3.0. The Essen/classical figure that *is*

supported is a step-inertia probability, not a step-to-leap ratio, and it is style-dependent

(§3.2).

3.4 Contour, and what actually distinguishes an earworm

EVIDENCE (S9 via S10; S10's own experiments). Huron reduced each Essen phrase to three points —

first note, mean of the interior, last note — and found the **convex (arch) contour most common,

followed by descending, ascending, and concave.** Goldstein et al. (S10) replicated without

assuming shapes in advance, clustering 35,793 European Essen phrases, and recovered the same four

families: convex, concave, descending, ascending. Arch and descending contours are common

cross-culturally and even in birdsong (Tierney et al. 2011, via S10). Complete short melodies tend

to describe an arch regardless of the shapes of their constituent phrases — the arch is a

multi-level property.

EVIDENCE (S8, Jakubowski et al. 2017). 100 tunes named as involuntary musical imagery matched

against 100 never so named, on 82 FANTASTIC melodic features plus tempo — 83 predictors in all

(this section previously said "83 FANTASTIC features"), second-order features computed against a

reference corpus of 14,063 commercial pop MIDI transcriptions (Geerdes). Random forest, then a

reduced three-predictor model. Results:

• Global contour conventionality is the strongest single predictor. Where the second-order

feature dens.step.cont.glob.dir exceeded 0.326 — i.e. the tune's overall melodic direction

profile was *common* relative to the pop corpus — approximately 80% of those tunes were

earworms.

• Local gradient unconventionality is the second. Among tunes below that contour threshold,

those whose average gradient between melodic turning points was *less* like the corpus

(dens.int.cont.grad.mean ≤ 0.421) were more likely to be earworms.

• Earworms are faster: M = 124.10 bpm (SD 28.73) versus 115.79 bpm (SD 25.39). This was

the only tempo finding in the literature at the time and the authors flag it as novel.

• Cross-validated classification rate: 62.5%. Chart success and recency were separately

predictive.

The shape of that result is the design instruction: **a conventional large-scale contour carrying

an unusual local gradient.** Familiar at the level you hum; distinctive at the level you notice.

Müllensiefen & Halpern (S17, via S8) independently found simple contours predictive of implicit

melodic memory.

3.5 Catchiness: repetitive and typical beat varied and novel

EVIDENCE (S12, van Balen et al. 2015). Recognition data from the *Hooked!* game — 321 songs,

1,715 fifteen-second segments, catchiness operationalised as drift rate (reciprocal of median

recognition time under a linear-ballistic-accumulation model). Mixed-effects regression, stepwise

selection by Satterthwaite-adjusted F-tests at α = .005, coefficients on standardised component

scores.

Read the provenance column before the numbers. These coefficients are not one model. The

audio components come from the audio-only fit on the full 321 songs / 1,715 segments; the symbolic

components come from fits on the reduced set of 99 transcribed songs / 536 segments, because

transcriptions were unavailable for the rest. An earlier version of this table presented all rows

as a single regression, which overstated the symbolic side's sample.

Marginal R² = .10 for the audio-only model on 321 songs and .10 for the combined model;

the audio-only and symbolic-only fits on the reduced 99-song set reach .06 and .07. Earlier

this table gave Melodic Range Conventionality as "0.05–0.07 [0.01, 0.13]", silently unioning three

different columns; the single-model value is now given. The authors' own reading of the symbolic side:

**melodic entropy and productivity both load negatively — recognisable melodies are more

repetitive** — and higher document frequency with lower second-order productivity means

recognisable melodies contain more typical motives.

Every melodic term that predicts catchiness points the same way: **less entropy, more repetition,

more conventional motives, more conventional range.** Our melodic_complexity axis is a

geometric mean of a five-term *variety* score and a three-term *structure* score

(instr_melodic_intelligence.py:496-560). There is no conventionality term anywhere in the

instrument, and the variety half rewards the exact quantity the catchiness literature finds

negative. GAP-05.

This is not an argument for blandness and must not be read as one. Schoenberg's own position (§2.4)

is that classical themes deliberately sought a popular touch; the earworm result (§3.4) pairs a

*conventional* global contour with an *unconventional* local gradient. The correct design is

conventional at the level of shape, distinctive at the level of detail — which is a two-band

requirement, not a single slider, and cannot be expressed by a scorer with only a variety axis.

3.6 Phrase and climax placement

CRAFT CONSENSUS, converging from three directions:

• Schoenberg (§2.3): the climax is approached through lesser high points, supported by increasing

small-note activity, and the line recedes to the middle register afterwards.

• Practitioner teaching: the highest note should be the climax and should generally fall roughly

two-thirds to three-quarters of the way through — as taught in songwriting curricula and

restated in accessible summaries (<https://hackmusictheory.com/blogs/theory/posts/6959928/melody-interval-rule>,

<https://champaignschoolofmusic.com/how-is-the-golden-ratio-present-in-music/>). Labelled

consensus, not evidence: no corpus study measuring peak position was obtainable in this pass.

• Contour research (§3.4): the arch is the dominant single-phrase and whole-melody shape, which is

the same statement in aggregate form.

MEASURED HERE (§5): Londonderry Air places its peak at 0.75 of the way through the strain —

squarely on the consensus figure. Ode to Joy places it at 0.10, and is a descending-family melody

with a repeated-period structure, which is a legitimate different design. So peak placement is a

per-theme design target to be declared, not a universal gate — see R-06.

3.7 Phrase-boundary evidence

EVIDENCE (S10) for the first half; CRAFT CONSENSUS restated by S10 for the second. In the Essen

database a phrase is *defined* as a unit marked by metrical quality, musical rests, and musical

syntax — that is the annotation scheme itself, read verbatim in S10. Separately, S10 lists as a

convention of Western melody, citing prior literature rather than measuring it here, that **notes at

the ends of phrases tend to be relatively long**. The two were previously presented as one

measurement; only the first is an Essen fact. That agogic cue is one of the two our

segment_phrases uses. The other class — a phrase that closes

harmonically inside an unbroken rhythmic surface — is invisible to a rest-and-agogic segmenter, and

segment_phrases' own docstring says so honestly (instr_melodic_intelligence.py:762-778). §5

shows what that costs in practice.

---

4. GAME-THEME PRACTICE

4.1 Koji Kondo (S19)

Kondo's GDC 2007 talk — its actual title is "Painting an Interactive Musical Landscape",

confirmed on the archive.org item — names three essentials for gameplay music: **rhythm, balance,

and interactivity.** That triad is directly confirmed in the Nintendo World Report write-up, as is

the contrast between Mario's music, which announces the action, and Zelda's, which is written for

ambience.

*Sourcing correction 2026-08-08:* this section previously offered *Super Mario Bros.*' tempo

increase as time runs out as Kondo's own illustration of interactivity. The tempo speed-up is a real

and well-known property of that game, but neither cited report attributes it to this talk, and

the talk itself was not watched. It is retained here as a general illustration of state-driven

music, not as something Kondo said.

CRAFT CONSENSUS with direct pipeline consequence: an "internal rhythm" the melody must ride is

a statement about onset density relative to the player's motion, not about tempo. A traversal

cue and a battle cue at the same BPM should not have the same note rate. That is R-03's second

level.

Ludomusicological analysis (S20 — Dominic Everitt, 2023) names Kondo's recurring devices as

modes, chromatic harmony, parallel motion and proximal voice leading; that phrase is verbatim in

the abstract. Two claims previously in this paragraph have been cut as unsupported:

• *"the NES's three available timbres forced melodic and rhythmic ingenuity to carry the identity"*

— removed. S20 was read at abstract level only, and its abstract says merely "limitations of

the 1990s era game console hardware". Worse, S20 analyses *Ocarina of Time*, which is an N64

title, so an NES voice-count was never the right hardware claim for that source at all.

• *"The Zelda overworld theme is consistently analysed as B♭-centred with a Mixolydian inflection

(S22)"* — corrected. S22 (Harrison Shimazu, Splice, 20 September 2019), now fetched in full

rather than read through a search summary, places the theme in B♭ major and explicitly

*rejects* the Mixolydian reading in favour of chords borrowed from the parallel minor. "B♭-centred"

survives; "Mixolydian inflection" and "consistently analysed" do not.

4.2 Winifred Phillips (S18)

The most directly actionable studio practice found in this pass, and it is about **restatement

count**, which nothing in our stack currently plans:

• A hook is a thematic element that catches the mind, and the craft rule for it is repetition.

Her worked example is the Star Wars main title: a four-bar melody with internal repetition,

alternating full statements with a contrast or bridge. Verbatim in Pt 1: this pattern happens

five times over the course of the theme.

• A second restatement count, from Pt 2: in *Assassin's Creed Liberation* the four-chord signature

pattern is heard 14 times in the main theme alone.

• When a full statement will not fit, fragment it: Pt 3 tracks a five-note excerpt of the

*Spore Hero* theme through the game's opening menu and its most frequently-used stinger.

• Variation is what prevents fatigue under repetition: the same melody returns in a dramatically

different arrangement rather than being replaced.

FABRICATION REMOVED 2026-08-08. This section previously asserted that for *The Dark Eye: Book

of Heroes* Phillips "restated the theme in six variations, and used a roughly four-minute

dedicated main-theme track to assert the identity before propagating it." **Both numbers are

unsupported.** Pts 2 and 3 were fetched in full for this audit; the game is named in each as one of

her projects, but neither states a variation count and neither describes a main-theme track of any

duration. The doc's own §1 read-status for Pt 2 was "read via listing" — i.e. not read — and a

specific number was nonetheless attributed to it. Nothing downstream may cite "six variations".

**This is Schoenberg's monotony-and-variation doctrine (§2.2) restated by a working game composer,

plus two verified numbers: five full statements across a main theme (Star Wars), 14 restatements of

a four-chord signature (Liberation), and a five-note fragment as the smallest transportable unit.**

4.3 Sugiyama and the eight-melodies template (S21) — with a correction

The Dragon Quest template — Overture, Castle, Town, Field, Dungeon, Battle, Final Battle, Ending —

is real and is correctly described as the template most RPG soundtracks inherited, with per-slot

stylistic commitments (Baroque contrapuntal castle music, romantic field music, deliberately

frantic and dissonant battle music), reused and expanded across the series.

Correction to the domain brief: this pass found no primary source for a Sugiyama doctrine

about eight-bar loops or composing specifically for repeated listening. The RPGFan feature contains

no such quotation. **That claim should not be repeated in our docs until a primary interview is

found.** What *is* supportable is the structural claim: a small fixed set of strongly-characterised

slot themes, each reused and varied across a long series.

4.4 The honest state of the game-theme evidence

Everything in §4 is craft consensus and prose analysis. **No number in this document about a

commercial game theme has been verified against its notes**, for the access reason declared in

§1.4. The numeric backbone of this document is §3 (published corpora) and §5 (our own scorers on

public-domain melodies). Any future doc that wants to assert "beloved game themes have property X

with value Y" must first obtain symbolic transcriptions and measure them.

---

5. REFERENCE MEASUREMENT: our own scorers, on melodies we can legally measure

MEASURED HERE. Five public-domain melodies were encoded as (MIDI pitch, duration in beats) with

a stated tempo, realised into feature_rig.Note objects, and passed through the *actual* functions

in harness/music_gen/instr_melodic_intelligence.py — melodic_complexity, contour_depth,

segment_phrases, phrase_sophistication. A sixth entry is a **deliberate control encoding Josh's

grade**: eight notes, one per bar at 70 bpm.

One of the five is a beloved game theme in its own right: *Korobeiniki* is the Russian folk song

(1861) used as the Tetris Type-A theme, and is public domain.

Reproduce with the script preserved at

C:/Users/joshu/AppData/Local/Temp/claude/C--dev-humanity-forgotten/c3c5f839-9982-4e1c-a28d-8fb8163ec031/scratchpad/probe.py

(transient; the encodings are restated in §5.3 so the measurement can be rebuilt anywhere).

5.1 The table

Columns: n note count; notes/s onset density; range ambitus in semitones; step%/3rd%/leap%

share of intervals at 1–2 / 3–4 / ≥5 semitones (our own definitions); S:L our step-to-leap ratio;

inertia% computed by the Chiu & Temperley method (§3.2); peak@ position of the highest note as a

fraction of the line; contour Huron's three-point reduction; depth, H1 and complexity are

contour_depth, interval_entropy and melodic_complexity as our instrument returns them.

5.2 What the table shows

Six findings, each of which becomes a rule in §6 or a gap in §8.

1. melodic_complexity barely separates the failure mode from real melodies. The control scores

0.384; the Tetris theme scores 0.461; Londonderry Air scores 0.512. A 0.08 margin

between the thing Josh rejected and one of the most-recognised melodies of the last century is

not a working discriminator. Greensleeves at 0.699 does separate — but Greensleeves is the most

*ornamented* of the set, which is the two-level-line point (§2.3) arriving through the back door.

2. The step-to-leap term is inverted over the real range (§3.3). The control gets 0.904; Ode to

Joy gets 0.366.

3. contour_depth does not discriminate at all. The control scores 4.00 — identical to

Korobeiniki and Ode to Joy, and above Greensleeves' 4.50 only by rounding. Windowed

Douglas-Peucker depth saturates on any line that wanders.

4. phrase_sophistication was unavailable for every single melody in the set, including Ode to

Joy, which is the textbook example of a four-phrase double period. Cause in §5.4.

5. Density is genre-conditional. 0.66–1.88 notes/s across five unimpeachable melodies, against

a pop-corpus norm of 1.8–2.8 (§3.1). A flat floor would fail Scarborough Fair. The control's 0.29

is nonetheless below every real melody by a factor of 2.3.

6. Ambitus separates cleanly where nothing else does. Real melodies span 7–14 semitones; the

control spans 5. Range is currently measured nowhere in the instrument.

5.3 Encodings, for reproducibility

Stated as (MIDI pitch, beats), tempo in bpm. These are standard published forms of public-domain

melodies; they are given so the measurement can be rebuilt and challenged.

• Korobeiniki (Tetris A), 150 bpm: 76·2, 71·1, 72·1, 74·2, 72·1, 71·1, 69·2, 69·1, 72·1, 76·2,

74·1, 72·1, 71·3, 72·1, 74·2, 76·2, 72·2, 69·2, 69·2.

• Ode to Joy, 120 bpm: 76, 76, 77, 79, 79, 77, 76, 74, 72, 72, 74, 76, 76·1.5, 74·0.5, 74·2

(all others 1 beat), then the same with the close 74·1.5, 72·0.5, 72·2.

• Londonderry Air, 76 bpm: 67·1, 72·2, 74·1, 76·3, 74·1, 72·2, 69·2, 67·1, 69·1, 72·2, 71·1,

69·3, 67·1, 72·4, 72·1, 74·1, 76·2, 79·2, 81·2, 79·1, 76·1, 77·3, 76·1, 74·2, 72·2.

• Greensleeves, 100 bpm: 69·1, 72·2, 74·1, 75·1.5, 76·0.5, 74·1, 71·2, 67·1, 69·2, 71·1,

72·1.5, 71·0.5, 69·1, 68·2, then the strain repeats with the close 67·1, 69·1, 71·1, 68·1.5,

67·0.5, 65·1, 69·3.

• Scarborough Fair, 100 bpm: 69·3, 69·3, 76·3, 76·2, 74·1, 76·3, 72·3, 69·2, 67·1, 69·2, 71·1,

72·3, 69·6.

• CONTROL, 70 bpm: 69·4, 72·4, 71·4, 69·4, 72·4, 74·4, 72·4, 69·4.

Caveat, stated rather than buried: these are working encodings of the principal strain, not

critical editions. A wrong note would move a percentage by a point or two; it would not move any of

the six findings, all of which turn on gaps of 2× or more or on a metric returning *unavailable*.

5.4 The phrase segmenter, isolated

Because finding 4 is severe, it was isolated. Ode to Joy was re-realised three ways and passed

through segment_phrases + phrase_sophistication:

The segmenter is doing what its docstring says it does — finding rests and agogic accents — and its

docstring is honest that a phrase carrying neither is invisible to it. But the consequence in

practice is that an entire scoring axis silently vanishes on any legato line, and legato lines

are most of a lyrical score. The agogic branch also fails on a knife edge here: it requires a note

strictly greater than 2.0× the local median duration, and Ode to Joy's cadential half-note

against a quarter-note median is exactly 2.0×.

5.5 The grammar and the scorer disagree about what a phrase is

The strongest single finding of this pass. A period was generated with the real builder —

pass3_grammar.period(cell, cell, mode, 0, "P1", weak_cadence="HC", strong_cadence="PAC") — and the

resulting note list was scored by the instrument.

• pass3_grammar says: theme_type = period, 18 bars, 60 notes, cadence

evaded_then_PAC at bar 17 with the weak HC recorded. By construction that is two phrases,

antecedent and consequent, with an expansion.

• instr_melodic_intelligence.segment_phrases says: 11 phrases, median length 5 notes.
• phrase_sophistication then computed phrase_length_variety 0.5934, antecedent_consequent_rate

0.2, cadential_differentiation 0.6522, phrase_hierarchy 1.0, total 0.713 — every one

of those numbers computed over eleven fragments that the grammar does not consider phrases.

Worse, the tonic estimate was wrong. phrase_sophistication estimates the tonic as the most

frequent pitch class (instr_melodic_intelligence.py:842). On this grammar-generated period the

true tonic pitch class was 2; the estimate returned 4. Since "closed" is defined as ending

on the tonic pitch class, the entire open/closed classification — and therefore the

antecedent-consequent detection — was anchored to the wrong note.

And separately from the estimator being wrong, the *definition* is too narrow. Running each cadence

gesture the grammar can produce through the instrument's closed test with the true tonic:

An imperfect authentic cadence is a closed cadence in Caplin and in Schoenberg; our instrument

calls it open because it accepts only the tonic pitch class. A period built as

antecedent-HC → consequent-IAC — a completely standard construction our own grammar can emit —

scores an antecedent-consequent rate of zero.

---

6. ACTIONABLE RULES, MAPPED TO OUR STACK

Each rule names the file and the function or constant it lands in. Rules are ordered by how much

they close the gap Josh named. ENGINE/CANON declaration: every rule here is ENGINE — it changes

how we generate and score audio, and touches no world canon.

R-01 — Score note density, at two levels, conditioned on the cue

Where: instr_melodic_intelligence.py — add to METRIC_KEYS (:1767) and AXES (:1773).

note_rate_per_s is already computed at :350 and :390 and is currently used only as a *gate*

at :946. Promote it to a scored axis.

What: report structural note rate (rate of the line's own long/accented notes) and

surface note rate (rate of all notes) separately. Evidence targets (§3.1): pop-idiom surface

lives at 1.8–2.8 notes/s; a lyrical theme legitimately runs 0.6–1.5. **Do not gate on an absolute

floor.** Gate on the *ratio*: a line whose surface rate equals its structural rate has no divided

surface at all, which is precisely the control's signature.

R-02 — Measure ambitus, and stop pretending 1.5 semitones is a melody

Where: instr_melodic_intelligence.py — VOICE_MIN_SPREAD = 1.5 (:156) is a

voice-detection floor of one and a half semitones of pitch standard deviation. Add a scored

pitch_range_semitones and pitch_sd_semitones.

What: §5 measures real melodies at 7–14 semitones of ambitus against the control's 5.

BiMMuDa (§3.1) reports two-thirds of pitches inside 5.5 semitones for modern pop, i.e. SD ≈ 2.75.

Range is the single cleanest discriminator this pass found and we do not measure it.

R-03 — Fix the step-to-leap term, or replace it with step inertia

Where: instr_melodic_intelligence.melodic_complexity (:496), the structure block.

What: three changes.

1. Close the dead zone. Thirds (3–4 semitones) are 17–30% of intervals in real melodies and are

counted as neither step nor leap. Either fold them into leaps (the conventional definition: a

step is a second, everything else is a leap) or report a three-way histogram.

2. Drop the 3.0 target. No corpus supports it; it scores Ode to Joy at 0.366 (§3.3).

3. Prefer step inertia (§3.2) as the order statistic, because it is measured, style-diagnostic,

and has published per-corpus values: 71% Essen, 67% classical themes, 63% hymns, 48% Rolling

Stone, 43% Billboard. Score against the declared idiom of the cue, not a universal target.

R-04 — Add a conventionality axis, and stop scoring variety alone

Where: instr_melodic_intelligence.melodic_complexity — the five-term variety half.

What: the catchiness regression (§3.5) finds melodic entropy negative and melodic

repetitivity, motive typicality, and range conventionality positive. The earworm study (§3.4)

finds conventional global contour the strongest single predictor. Our variety half rewards the

opposite of all four. Add a second-order term — how typical is this line's contour and interval

profile against our own corpus of authored cues — and make the axis a band, not a maximiser:

conventional at the level of shape, distinctive at the level of local gradient. This is the single

most important conceptual change in this document and it is the one most likely to be got wrong by

implementing it as another slider.

R-05 — Measure the climax and its approach

Where: new function in instr_melodic_intelligence.py; scheduling hook in

pass3_grammar.py's phrase builders.

What: report (a) the position of the highest note as a fraction of the line, (b) whether

density and interval activity increase into it, and (c) whether the line **recedes to the middle

register** after it. All three are Schoenberg's explicit conditions (§2.1, §2.3) and (a) is measured

here at 0.75 for Londonderry Air. Report per theme; do not gate universally, because a descending

melody like Ode to Joy legitimately peaks early (§3.6).

R-06 — Score contour SHAPE, not only contour depth

Where: instr_melodic_intelligence.contour_depth (:480) stays; add a shape classifier.

What: Huron's three-point reduction — first note, mean of interior, last note — classifying into

convex / concave / ascending / descending (§3.4). It is four lines of code, it is the reduction the

published corpus results use, and §5 shows contour_depth alone cannot tell the control from

Ode to Joy (both 4.00).

R-07 — Let the grammar hand its phrase boundaries to the scorer

Where: the interface between pass3_grammar.Phrase (:921, with to_dict() carrying

functions, cadence, start_bar, bars, derived_from, operation) and

instr_melodic_intelligence.measure() (:1087), which today takes only a card and an audio path.

What: when we composed the cue, we know the phrase boundaries, the cadence types, and the

formal functions — they are in the plan. Re-deriving them from a rest-gap heuristic on rendered

audio throws away ground truth and then scores the noise (§5.5: 11 phrases where the grammar built

2). Pass the phrase plan through the card and let segment_phrases validate against it instead

of replacing it, keeping the audio path as the fallback for material we did not author.

R-08 — Widen "closed", and fix the tonic estimate

Where: instr_melodic_intelligence.phrase_sophistication (:821).

What: accept scale degrees 1 and 3 as closed (an IAC is a closed cadence), and either take

the tonic from the plan — which we know exactly — or replace the most-frequent-pitch-class

heuristic with a Krumhansl-style key-profile correlation. §5.5 shows the current estimator returning

pitch class 4 for a period whose tonic is pitch class 2.

R-09 — Segment on cadence, not only on silence

Where: segment_phrases (:762), constants PHRASE_GAP_MULT = 1.55 (:184) and the agogic

branch.

What: the agogic test requires a note strictly greater than 2.0× the local median, which a

textbook cadential half-note against a quarter-note median fails by exactly zero (§5.4). Relax to

≥ 1.8× and add a harmonic-arrival cue. Until then, report the axis as unavailable-with-reason

rather than letting a legato cue quietly carry no phrase score at all.

R-10 — Plan the restatement count

Where: the cue plan schema (build/audio/pass2/plans/*.json) and whatever authors it; the

hook.returns list already exists.

What: Phillips (§4.2) gives working numbers: five full statements of the melody across a

main-theme cue with contrast between them, 14 restatements of a short signature pattern within a

single main theme as the high end of the range, and a five-note fragment as the smallest

transportable unit for stingers and menus. (The "six variations" figure this rule previously carried

was a fabrication and has been struck — see §4.2.) Our round-2 cues carried 21 statement layers

across three cues of which literal alone was eight — restatement without variation, which is

Schoenberg's named monotony (§2.2).

R-11 — Preserve rhythm when varying pitch

Where: pass3_grammar.OPERATIONS (:63-123) and the spine_changing discriminator.

What: Schoenberg's stated corollary is that **preserving the rhythm licenses far-reaching

contour change** (§2.2). Our spine_changing flag asks the pitch question; add the rhythm question

as its partner, so an operation can be classified as *developing variation* (spine changed, rhythm

preserved) versus *re-dressing* (spine preserved) versus *foreign* (both changed). The middle

category is the one Schoenberg says produces coherence and it is not currently nameable.

R-12 — Schedule division against the climax

Where: pass3_grammar.divide() (:542), _nct_figure() (:591), division_level on Phrase.

What: division_level is currently a phrase-wide constant. Schoenberg's climax passage (§2.3)

requires the small notes to increase into the high point and thin after it. Make division a

schedule across the phrase, keyed to the climax bar, rather than a flat setting. This is the

concrete mechanism that turns "a few notes slowly" into a real line without changing a single

structural pitch.

R-13 — Tempo is a lever we are not pulling

Where: cue planning; build/audio/music_cue_table.json.

What: earworm tunes averaged 124.10 bpm against 115.79 for matched non-earworms (§3.4),

and Kondo's first essential is that the music ride the game's internal rhythm (§4.1). Neither our

plans nor our scorers relate a cue's tempo to its melodic density or to the player action it

accompanies. At minimum, record the intended player motion per cue and check note rate against it.

R-14 — Report unavailable as unavailable

Where: melodic_complexity (:496) and predictive_info (:420).

What: predictive_info returns all-zero for sequences shorter than 8 symbols. A melody of

exactly 8 notes yields 7 intervals, so interval_entropy is reported as 0.0 while

melodic_complexity still reports available: true — a measured-looking zero for something that

was never measured. Our control hit exactly this. Propagate an availability flag.

---

7. WHAT WE ADOPT

Ranked. Everything above the line is a defect fix and should land before any new melodic authoring

run; everything below is new capability.

Adopt now — these are corrections, not enhancements:

1. R-03 — fix the step-to-leap term. It currently scores the failure mode at 0.904 and Ode to

Joy at 0.366. Any melodic ranking produced while this stands is untrustworthy in a way that

points *toward* the material Josh rejected.

2. R-08 + R-09 — repair phrase_sophistication. It was unavailable for all five reference

melodies and produced numbers over 11 phantom phrases on our own grammar's period, off a wrong

tonic. It is currently reporting confidently and wrongly, which is worse than not reporting.

3. R-07 — hand the grammar's phrase plan to the scorer. This is the structural fix that makes

R-08 and R-09 mostly unnecessary for authored material, and it costs nothing but wiring: the

ground truth already exists in pass3_grammar.Phrase.to_dict().

4. R-14 — availability flags. Cheap, and it is the difference between a gate that is armed and

a gate that reads zero.

Adopt now — these are the three properties Josh's grade named that we do not measure:

5. R-01 — note density as a scored axis, at two levels. Directly names "slowly".

6. R-02 — ambitus. The cleanest discriminator found in §5, and directly names "a few notes".

7. R-12 — division scheduled against the climax. The mechanism that fixes "back to back"

without touching the structural line. Pairs with R-05.

Adopt next:

8. R-05 — climax measurement, reported per theme, gating nothing until we have a per-cue design

target to gate against.

9. R-06 — contour shape classification. Four lines, and it is the reduction every published

corpus result in §3.4 uses, which makes our numbers comparable to theirs.

10. R-04 — the conventionality axis. The highest-value item conceptually and the highest-risk

to implement. It requires a reference corpus of our own authored cues before it means anything,

and it must land as a band rather than a maximiser. Do not ship it as a slider.

11. R-10 + R-11 — restatement planning and the rhythm-preservation partner flag. These belong

with the motivic lane (MOTIVIC_DEVELOPMENT_AND_PHRASE_GRAMMAR.md §6) rather than here.

12. R-13 — tempo and player motion. Blocked on the cue table carrying an intended-motion field.

Explicitly NOT adopted:

• No flat note-density floor. §5 measures Scarborough Fair at 0.66 notes/s. A floor set from

the pop corpus would fail it. Density is conditional (R-01).

• No universal climax-position gate. Ode to Joy peaks at 0.10 and is not defective (§3.6).
• No entropy maximiser. §3.5 is unambiguous that melodic entropy loads *negatively* on

recognisability. The instinct that "more complex" means "higher entropy" is the instinct that

produced a variety half with no counterweight, and adding more of it would make the melodies

worse in the exact direction the literature predicts.

• The Sugiyama eight-bar-loop doctrine is not adopted because no primary source for it was

found (§4.3).

---

8. GAP NOTES — where our named stack files deviate from the practice found

For the audit phase. Each gap names the file, the location, the practice it deviates from, and the

evidence. Severity is this document's own judgement and the audit is free to overturn it.

8.1 One cross-cutting note for the audit

GAP-04 and GAP-05 point the same direction and are the reason this document exists. Our melodic

scorer, as it stands, **rewards the failure mode Josh named on two of its three structure terms and

on its whole variety half.** That is not a claim that the instrument is worthless — its development

census and its shuffle-corrected predictive information are genuinely good work, and the file's own

mutation controls are exemplary. It is a claim that the axis most likely to be used as a ranking

signal is, over the range where real melodies live, pointing the wrong way. Nothing else in this

document matters as much as checking that claim, and §3.3 and §5.1 are written to be re-run rather

than believed.

---

9. HONEST LIMITS

1. No commercial game theme was measured. §1.4 and §4.4 declare this. Everything numeric here

comes from published corpora of folk, classical, hymn and pop melodies, or from public-domain

melodies measured in §5. The Zelda / Final Fantasy / Dragon Quest side of the domain brief is

served qualitatively only.

2. The §5 encodings are working transcriptions, not critical editions (§5.3). The findings are

robust to a wrong note; the individual percentages are not.

3. Five books were not read in full (S3, S4, S5, S6, S7) and are cited only for claims sourced

through readable secondary material, as marked in §1.

4. The peak-placement consensus (§3.6) has no corpus behind it in this pass — only practitioner

teaching and the aggregate arch results. It is labelled consensus and R-05 deliberately reports

rather than gates on it.

5. The conventionality axis (R-04) has no reference corpus yet. It cannot be implemented as

specified until we have enough authored cues to compute second-order features against, and

implementing it against an external pop corpus would import an idiom this game does not want.

6. Web search budget was exhausted during this pass, which is why the singability/vocal-range

literature and a primary Sugiyama interview are both absent. Both are named gaps for a follow-up

pass, not silent omissions.

7. This document shipped with two fabricated claims and three misattributions, found by the

2026-08-08 audit and repaired in §11. They were all in §4 and in the source table, never in the

corpus numbers — but that is a fact about where this pass was weakest, not a reason to trust §4

now. §4 remains prose analysis and carries no verified number about a commercial game theme.

10. DERIVATION NOTE

Every number attributed to a published source in §3 was read out of the source document itself —

the Jakubowski accepted manuscript, the Chiu & Temperley author PDF, the van Balen ISMIR PDF, the

Goldstein preprint, the Schoenberg OCR scan — not out of an abstract or a search summary, except

where §1 marks otherwise. Every number in §3.3, §5 and §8 was produced by importing and running the

actual functions in harness/music_gen/instr_melodic_intelligence.py and

harness/music_gen/pass3_grammar.py at the repository state of this pass, and every one of them

should be re-derived by the audit rather than trusted, because a same-commit edit can invalidate a

true proof.

---

11. CITATION AUDIT — 2026-08-08

An independent reality check was run over this document's source set. Every source below was fetched

and read at the level stated; PDFs were text-extracted locally rather than trusted to a summariser.

The audit's own bias was to hunt for invention, so the clean results are stated as plainly as the

defects.

11.1 Verified against the primary source, number by number

• S1 Schoenberg. Imprint verified (ed. Gerald Strang with Leonard Stein, Faber & Faber, 1967).

All five melodic conditions in §2.1 are a single verbatim passage. The monotony/variation doctrine,

the motive-as-shape definition, the rapidity and intelligibility positions, the Mattheson 1739

quotation, the climax/recession passage, the liquidation definition, and the capitalised sentence

in §2.2 — all present verbatim, and the capitalisation claim is literally true in the source.

Chapter attributions checked by offset: §2.1 ch. IV ✓, §2.2 ch. III and ch. VII ✓, §2.3 ch. VI and

ch. VII ✓, §2.5 ch. VIII ✓. Two errors found and fixed — see §11.2.

• S8 Jakubowski et al. 2017. Journal, volume, issue and pages exact. Every number in §3.4 —

0.326, ~80%, 0.421, 124.10/28.73 vs 115.79/25.39 bpm, 62.5%, 100 vs 100 tunes, the 14,063-melody

Geerdes reference corpus — verified in the accepted manuscript. The novelty claim about the tempo

finding and the Müllensiefen & Halpern implicit-memory restatement are both there verbatim.

• S10 Goldstein et al. 35,793 Essen European phrases ✓; four clusters convex/concave/descending/

ascending ✓; Huron's three-point reduction and his convex > descending > ascending > concave

ordering ✓; short melodies arching regardless of constituent phrase shape ✓; Tierney birdsong ✓.

• S11 Chiu & Temperley 2024. *Music & Science* 7: 1–11, DOI as cited. Table 1 verified

cell-for-cell: 194/41,066/48%, 214/68,493/43%, 3,786/145,381/71%, 9,776/142,149/67%,

17,683/901,372/63%. Table 2 at step size 3 verified: 47/42/60/59. Step ≤ 2 semitones with pitch

repetitions removed ✓. The Huron descending-steps caveat ✓.

• S12 van Balen et al. 2015. 321 songs, 1,715 segments, 973 participants ✓; drift rate under

linear ballistic accumulation ✓; α = .005 ✓; every coefficient in §3.5 verified against Table 2;

the authors' reading of the symbolic side (entropy and productivity both negative; higher DF with

lower second-order productivity ⇒ more typical motives) is verbatim.

• S13 Burgoyne et al. 2013. Title, four authors, ISMIR 2013, pages 245–50, Curitiba — verified

via the reference list of S12.

• S14 van Balen thesis ch. 7. Confirmed: chapter titled "Hooked", §7.1 on catchiness, earworms

and hooks, design chapter with results deferred to its ch. 8.

• S15 Hamilton & Pearce 2024. Onset densities 1.8 / 2.0 / 2.8 notes/s and mean pitch intervals

2.3 / 2.1 / 2.0 semitones verified; the 5.5-semitone two-thirds figure is correctly scoped to the

most recent era (the paper gives ~7 and ~6 semitones for the two earlier eras); revolutions 1975

and 2000 tier-1, 1996 tier-2 ✓; declining pitch and rhythmic complexity ✓.

• S16, S18 (Pt 1 and Pt 3), S19 talk title, S20 device list, S21, S24 — all confirmed to exist

and to say what is attributed to them, with the exceptions listed below.

• S23 Hooktheory. The declared HTTP 403 on automated fetch is real; re-confirmed. The doc's

refusal to cite it as evidence stands.

11.2 Defects found and repaired in place

11.3 What the audit did not disturb

No number in §3 sourced to a published corpus was wrong. Every defect above is in the citation

apparatus or in §4's game-practice prose — the qualitative layer the document already declared

weakest in §1.4 and §4.4. The quantitative backbone (§3.1, §3.2, §3.4, §3.5) survived intact, which

means the §6 rules that rest on it — R-01, R-02, R-03, R-04, R-05, R-06 — are unaffected. R-10 is

the only rule whose stated numbers changed, and it was already ranked eleventh of twelve.

The §5 and §8 numbers are out of scope for this audit: they are MEASURED HERE, reproducible from

§5.3, and §10 already instructs that they be re-derived rather than trusted. Nothing in this pass

either confirms or disturbs them.

Standing lesson for the sibling volumes. Both fabrications and one reversal share a single

signature: **a specific claim attributed to a source whose own read-status line says it was not

read.** The read-status column is the control, and it works — it named all three defects before the

fetches did. The rule it implies is that a source marked unread may support framing and may not

support a number, a key, or a hardware fact.

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