FINAL_BUY_ROUND.md

music/exemplars/FINAL_BUY_ROUND.md

FINAL BUY ROUND — the last music acquisition, capped and closed

Ruled 2026-08-07 (docs/spine/DECISIONS_PENDING_JOSH.md, commit cd2619f7): one final capped buy

round, and the acquisition treadmill ends with it. Anything unbuyable is studied from what exists,

never re-queued. Generation starts now on the N = 28 findings regardless of what lands here.

Cap: ~$40–60. This list: $57.05. Every price and every URL was fetched live on 2026-08-07 —

the 2026-08-05 sheet's estimates drifted, in both directions, and four of them are corrected below.

Legal posture unchanged and binding: analysis for understanding only, the bytes are never model

input, nothing is redistributed.

Predecessors: BUY_SHEET.md (the 111-album survey, 2026-08-05) · PATTERN_FINDINGS_V2.md

(what the pool still needs, §11) · CORPUS.json defect_register (the flagged rows).

Files land at D:/audio/exemplars/.

PHONE-FIRST PUBLISH (the standing law: a board is not done until it is phone-viewable):

https://claude.ai/code/artifact/0067d421-1a66-4ece-8f76-23c872840239

Same rows as this file, with tap-target cards and a running total that counts $57.05 down as each is

bought. Its source is build/audio/exemplars/FINAL_BUY_ROUND.html — edit the markdown first and the

HTML second, so the two never drift, and republish the same path to keep that URL.

STATUS, refreshed 2026-08-07 after commit 9f4f7a94: PRIORITY 2 IS DONE AND COST $0.00. Both

zero-cost rows are landed and CARDED, both albums are now in the study pool with their full sibling

sets, and the corpus reads 152 MISSING · 18 ACQUIRED · 12 CARDED of 182. Nothing on the $57.05

capped list has been bought yet, so every price row below stands exactly as verified. Two claims this

sheet made about the Tekken 2 and Rayman Legends card counts were WRONG and are corrected in place —

see PRIORITY 2; the correction makes the round cheaper in compute, never in money.

---

THE TOTAL

$57.05 — five checkouts, four storefronts, every cent verified live today. Still unbought.

Zero estimates in the total. Two further rows on this sheet cost $0.00 because the audio was already

on disk and already paid for — and those two are now LANDED, CARDED and in the study pool.

---

PRIORITY 1 — the MGS recovery, and the truth about the score albums

The ask was the two MGS score albums. Here is what is actually true, verified today.

• The 1998 Metal Gear Solid Original Game Soundtrack (KICA-7895) is NOT sold digitally anywhere.

Absent from the US iTunes Store and the JP iTunes Store (both queried directly against Apple's own

catalogue API — every "Metal Gear Solid" album that returns a price is a cover, a piano tribute, a

pachislot score, an MGS V album, or Gigi Meroni's third-party re-recording). Absent from Steam —

Konami sells no soundtrack DLC at all; the only carrier is the $59.99 MASTER COLLECTION Vol.1 BONUS

CONTENT pack, which is a game bundle. Absent from Qobuz.

• The 2004 Metal Gear Solid 3: Snake Eater Original Soundtrack (KOLA-089/090) is worse. No digital

route in any territory, and the one in-print physical seller reads "Sold Out"

(CDJapan KOLA-89, $18.95 when it existed). The only

Snake Eater album a store will sell today is the 2025 Δ remake ($23.99 Apple / $35.09 Qobuz) — a

different recording of a different production, which measures as a different card.

• So the two score albums cannot both be bought, and the second cannot be bought at all.

What recovers the two dropped hits instead — and it is better, cheaper and instant.

EX_082 and EX_092 are not dropped because their album is unowned. They are dropped because **only two

of its thirteen tracks were bought** — the $1.29 singles. `D:/audio/exemplars/iTunes_Sweep/Compilations/

Metal Gear Solid Vocal Tracks/` holds exactly two files. Zero siblings, so both rows are silently

excluded from every within-album statistic in PATTERN_FINDINGS_V2 (which is why its headline was

computed on 26 of 28; at today's N = 30 the same exclusion makes it 28 of 30, and these two are

still the only zero-sibling rows in the pool). **Completing the album they are already carded against

fixes both at once, and it is the only remaining unbought row that recovers a hit already paid for.**

Tracklist verified today: 8 distinct vocals across MGS1/2/3/Peace Walker/Portable Ops/MGS V, plus 5

Donna Burke re-arrangements (Jazz Lounge, Trip-Hop, Tango, Tender, DB English). Stated honestly:

five of the eleven siblings are re-arrangements of four of the other eight, so this sibling set is

narrower than a normal album's filler — but it is a real set of eleven, it is the set the cards already

name, and it converts the usable denominator from 26 to 28 before anything else is bought. It also

hands the rig a free instrumentation-delta control: one melody in five arrangement states.

Possible discount, marked as an estimate: Apple's Complete My Album may credit the $2.58 already

spent on the two singles, bringing this row to about $12.41. Not verifiable from outside the store —

budget the full $14.99 and take the credit if it renders at checkout.

---

PRIORITY 2 — the two already-owned rows: RESOLVED. CARDED. $0.00 spent.

This was the single biggest finding of the round, and it is now executed. Commit 9f4f7a94 landed

both rows; commit below wired their siblings into the study. Neither needed a purchase, and neither

needed the $8.99 the 2026-08-05 sheet quoted, which had already been spent in the starter-ten sweep.

EX_014 (Tekken 2) — RESOLVED BY INSPECTION

`D:/audio/exemplars/iTunes_Sweep/TEKKEN Project & Bandai Namco Game Music/Tekken 2 (Original

Soundtrack)/` holds all 32 tracks. The defect register's own instruction — *"the underlying question

… is answered by buying the album and inspecting it, not by this tool"* — was discharged by looking:

• The album carries no track named "Opening Theme." Track 01 is **"Tekken 2 Attract Movie: Sound

Track"**, its only attract/opening cue. Track 16 is "Ending: Sound Track", track 18 "Are You Ready?".

• The false Pokémon X & Y attach that had been sitting on the row was DEMOTED with its reason —

logged to DEMOTION_LOG.jsonl, prior acq kept under demotion.prior_acq, card and record moved to

repudiated/ — before --id EX_014 attached track 01 at game agreement 1.000, exact. It reads

CARDED on a human's evidence. Josh's title is kept in t_as_listed. No override_open: with exactly

one attract cue on the release, the inspection is determinate.

• **THE HONEST CAVEAT, AND IT MUST TRAVEL WITH EVERY NUMBER STATED FROM THIS ROW: the cue is 21.5

seconds.** It is an attract STING, not a developed theme, and it is the shortest measured thing in the

whole pool. TITLE_CHARACTER's fourth row is a twenty-one-second card, so any per-class

MASTERPIECE_STANDARD band computed from TITLE_CHARACTER must say so on its face or it is quietly

reporting a sting as a theme. PATTERN_FINDINGS_V1 already priced this exact hazard — most of its

ranking turned out to be duration wearing a costume — so a 21.5 s positive is the one row most able to

move a length-correlated axis for the wrong reason. The album's longer alternates are on the record if

it ever reads wrong: t.18 "Are You Ready?" 102.1 s, t.16 "Ending: Sound Track" 83.4 s, t.02 "Figthter

Select" 38.7 s (Bandai Namco's own typo, kept).

EX_094 (Rayman Legends) — RE-KEYED, override_open SET

Rayman Legends (Original Game Soundtrack) is on disk, 46 of 47 tracks. It carries **no track named

"Escape from Vector"** — verified against the on-disk filenames and against Apple's 47-track listing.

That is the third data defect of the EX_076 / EX_116 class: the row named a cue the official release

does not have. Re-keyed onto "Dark Creature Pursuit" (t.35) — the album's escape cues are its three

timed "Pursuit" tracks, this is the one that names the pursuer, and at 115.4 s it is the longest, so

the card measures a developed cue rather than a sting. override_open is set with four runners-up named

on the row ("Infernal Pursuit" 85.5 s, "Enchanted Forest Pursuit" 111.4 s, "Missile Airlines" 90.1 s,

"Hell's Gate" 62.3 s), so Josh's call still wins.

THE CORRECTION THIS SHEET OWES: both "one card" claims above were FALSE

This sheet said Tekken 2 had one formula card and Rayman Legends one, and concluded that 76

tracks needed measuring. Both albums were already fully carded — 32 Tekken cards and 46 Rayman cards,

all measured 2026-08-07 at 24/24 positive control in commit 9270acab, the wholesale sweep that also

found the six poisoned rows. Not one minute of measurement was owed.

The wrong number came from counting an album's cards by **grepping the album name against card

FILENAMES**. Card filenames are built from the TRACK title, not the album, so that search can only ever

match a track whose own title happens to contain the game name — and each of these albums has exactly

one: UNMATCHED__01_Tekken_2_Attract_Movie__Sound_Tra and

UNMATCHED__45_Rayman_Theme_8_Bit_Version_. The search returned 1 and 1 because **1 and 1 are the only

answers it could return**, not because the bank was empty. Counted the only way that is sound — by the

source_path recorded inside each card — the answers are 32 and 46. This is the repo's own

a-search-that-cannot-match-reports-zero law, and the positive control that catches it costs one line:

run the count both ways and require them to agree, or explain why they do not.

What was actually blocking the two rows was never the cards. It was the pool's ability to see them,

and that is repaired in this refresh — see WHAT THIS ROUND BUYS THE STUDY below.

---

PRIORITY 3 — boss and tension fills, cheapest first

PATTERN_FINDINGS_V2 §11 names the constraint precisely: *"Five of seven purpose classes are still

below five rows … BOSS and TENSION are the binding constraint."* No per-class MASTERPIECE_STANDARD band

can be stated until each class carries five-plus rows. These four purchases clear both, and are ranked

by rows-per-dollar on the binding classes.

Priority-3 subtotal: $42.06. Full list total: $57.05.

If the cap has to bite

Stop after row 4 and the total is $45.06: BOSS reaches 4 instead of 5 and the EX_076 defect stays

open. Every other gain above survives. Nothing in this list depends on anything below it.

Declared alternate, considered and not taken

ELDEN RING (OST), iTunes $17.99 VERIFIED, 67 tracks — carries EX_062 and EX_063, two BOSS rows in

one purchase, and would reach BOSS 5 on its own for $17.99 against the $31.97 of rows 2, 4 and 5. It is

not the recommendation because those three land six rows across four classes rather than two in one,

they close a data defect, two of them are FLAC rather than 256 kbps AAC, and Elden Ring may already be

free in the Steam library (app 1896300 ships with Digital Deluxe and exports all 67 tracks) — a check

that cannot be made here because the Steam client is not installed on this box.

Two checks worth ninety seconds, both possible savings

• UNDERTALE may already be in the Steam library as app 391570 (320 kbps MP3, no base game needed).
• Hollow Knight may already be in the Steam library as app 598190 (declares a lossless tier).

Both are still worth the $9.99 each: the Bandcamp masters are FLAC and the Steam copies are not, and

the rig's fidelity control exists precisely to price that difference. Buy them regardless unless the

cap bites.

---

WHAT THIS ROUND BUYS THE STUDY

Purpose classes — the binding constraint, cleared

The two free rows are in the "today" column now rather than the "after" column: TENSION moved 2 → 3 and

TITLE_CHARACTER 3 → 4 at zero cost. BOSS did not move and is now the sole binding constraint — it is

the only class the capped buy must clear, and rows 2, 4 and 5 are what clear it. Five of seven classes

reach the five-row floor the MASTERPIECE_STANDARD's per-class bands need; the two that do not

(CREDITS_TRIUMPH, TITLE_CHARACTER at four) are not the constraint. **TITLE_CHARACTER's fourth row is the

21.5 s attract sting**, so that class is the one to read most sceptically, not the one to celebrate.

The pool

The "today" column is measured, not projected — derive_patterns.py --perm 400 reports

n_hits = 30, n_ctrl = 927, n_other_albums = 223 over n_cards_measured = 1182, with

Tekken 2 (Original Soundtrack) joining the album list. EX_014 draws 31 within-album siblings and

EX_094 draws 45 — the two numbers this sheet promised, delivered without spending a cent or a minute

of measurement, because the 76 cards already existed.

Two code defects had to be repaired before the pool could see them, and both were real:

• A stale quarantine. derive_patterns.py hardcoded KNOWN_FALSE = {"EX_014"} from the era when

that row measured a Pokémon file, and dropped it from the pool in both directions. After the repair

the row was still being withheld, and with it the Tekken 2 album's entire 31-track sibling set, because

an album with no hit contributes no controls. The set is now empty, on the corpus's own

defect_register wording, and the control that guarded it no longer names a row: it asserts that **no

attached corpus row is quarantined at all**, which cannot go stale as rows are repaired.

• A self-comparison leak, and it was live. A wholesale sweep cards every file on an album; a later

--id attach then gives one of those same files a second card under an EX_ id. Both new rows had

such a twin — byte-identical sha256, different track_id — and the control that claimed "no hit

leaked into controls" compared track_id SETS, so it reported clean exactly when the leak was present.

EX_094 was being scored against a copy of itself in the run before this one. Controls are now keyed on

the audio's hash, not its name, and a third control asserts the twins really are on disk so the

check cannot pass vacuously. All four repairs are armed by mutation: reintroduce any one defect and the

named control goes red. 25/25 controls green at HEAD.

The projected "after" figures are lower than this sheet's first estimate (~1,440 cards, ~259 new)

because 76 of those cards were never new. The purchases add roughly 183 cards, not 259.

Where that lands on the power curve

PATTERN_FINDINGS_V2 §11's re-fitted model (K_eff = 103, calibration error 0.007):

The two free rows put the pool exactly on the number V1's power model named. PATTERN_FINDINGS_V1

priced the next-data question at 30 carded rows to make a true 0.75 effect visible at 80% power, and

the corpus now sits at 30 — at the bar its own null result set, rather than at a third of it. Read that

as *worth re-running*, not as *answered*: 30 buys a 0.75 effect, and the melodic-craft axes that failed

in V1 would have to live nearer 0.70, which still needs the buy.

The pool crosses the line V2 named. Today only a true 0.75 effect is reliably visible, which is why

exactly six axes appeared and all six are measures of structural scale. At 38 the bar falls to about

0.667 and a true 0.70 effect becomes detectable at roughly four times in five — which is the range

the melodic-craft axes would have to live in if they live anywhere. That is the specific thing this

round buys: not more of the same finding, but the first honest chance of a different one.

It also fixes a defect in the finding itself. Two records — Chrono Trigger and FINAL FANTASY VI —

carry 11 of the 28 hits today, 39% of the positives on two albums. After this round the six new

albums spread the positives across 22 records, and the leave-one-album-out robustness check

V2 §12 names as "the obvious next check, not run here" becomes runnable.

---

PRIORITY 4 — the Kondo lineage: naming the truth

There is no legitimate digital route to any Nintendo first-party game music. Nintendo sells nothing.

This is not a gap in the research; it is the state of the market, and it was verified in seven live

lanes on 2026-08-05 and spot-re-checked today. Concretely:

• Nintendo operates no digital music storefront in any territory. Not the eShop, not iTunes, not

Steam, not Bandcamp, not Qobuz. Zero of the corpus's 46 Nintendo first-party rows — every Kondo

cue, every Zelda cue, every N64 cue, Super Metroid, EarthBound, Donkey Kong Country, Kirby, Star Fox,

F-Zero, Super Mario RPG — has a Class-A route at any price.

• Nintendo Music (Switch Online, $19.99/yr) is not a route. It streams, its in-app downloads are

encrypted, and it has no export path. It buys the ability to HEAR a cue, never a file, and the rig

reads files. It is genuinely useful for one thing — auditioning a pressing before spending money on a

used import — and worth exactly nothing to the measurement. It does not even carry the corpus's

hardest holes: no Majora's Mask, no Twilight Princess, no EarthBound, no Kirby, no Super Mario RPG,

no Pokémon Gold & Silver.

• This is why PATTERN_FINDINGS_V2 §9 says the lineage is still only half covered — the Square/Enix

16-bit and Sega 16-bit lineages opened, and the Nintendo lineage did not. It is the last untested era.

It stays untested. That is the honest end state, and per the 2026-08-07 ruling it is not re-queued.

The one-or-two cheapest legitimate physical options — OPTIONAL, and outside the total

Listed because they are real, priced live today, and never blocking. **Do not buy these to unblock

anything** — nothing downstream waits on them. They are class B: a CD arrives, the LG BD-RE at F:

rips it into the same drop folder, and the rig cards it like any other file.

Everything else in the Nintendo lineage — Ocarina of Time and Majora's Mask on Pony Canyon, the Sound &

Drama disc, Super Metroid, DK Jamz, Mother 2, the Kirby compilations, Star Fox 64, F-Zero X, Mario

Kart 64, Perfect Dark — is a used-import hunt with no live-verifiable price, catalogued in

BUY_SHEET.md § "Used market and imports". It stays there. It is not part of this round.

---

DECLARED PERMANENTLY UNBUYABLE

The treadmill ends here, so these are said plainly and not re-queued. Each is studied from what exists

or declared ABSENT on its card — never measured as a cover and called the source.

One genuine footnote, because it is true and cheap. The 1998 MGS1 CD is not vapour: CDJapan lists

KICA-7895 at **US$13.49, in stock, ships 2-4 days,

verified today, 21 tracks — and its tracklist carries "ENCOUNTER" at track 6 (EX_093, TENSION)** and

"END TITLE / THE BEST IS YET TO COME" at track 18 (EX_082's cue in its original score context). It

is a real class-B route to twenty genuine score siblings plus a sixth TENSION row. It is not in the

capped total because it is physical: it ships from Japan, it adds shipping, it needs ripping, and

generation starts now rather than in two weeks. If Josh wants one optional disc, this is the better of

the two on offer — a whole score album and a corpus row for $13.49.

---

HOW EVERY PRICE ON THIS SHEET WAS VERIFIED

The 2026-08-05 sheet's estimates drifted, which is the whole reason this round re-verified rather than

inherited. Corrections found today, all against the live stores:

Method, per storefront:

• Apple — queried Apple's own catalogue API by collection id (itunes.apple.com/lookup), which

returns the live collectionPrice, track count and release date, on the US and JP storefronts. This

is the same field that returns -1 for the Silent Hill 1999 album, Apple's sentinel for *streams but

is not for sale* — which is how "not purchasable" is distinguished from "not found" rather than

guessed at.

• Steam — queried the store search and appdetails APIs directly (cc=US), which is how the

absence of any Konami soundtrack DLC was established rather than assumed.

• Bandcamp, Qobuz, CDJapan — product pages fetched and the price string read off the page. Where a

page says "Starting at", this sheet says "Starting at".

• eBay — returns HTTP 403 to every automated fetcher. Its one row is marked ESTIMATE and links

to a search with the exact query. No product URL on this sheet was constructed by guessing.

• Disk — the already-owned findings were read off D:/audio/exemplars/ directly, file by file,

and cross-referenced against CORPUS.json and formula_cards/.

---

AFTER THE BUY — the ingest, in order

1. Drop every purchase into D:/audio/exemplars/ exactly as each store dumps it. No renaming, no

re-foldering — the game-agreement floor reads the folder and tag evidence.

2. ~~Card the two already-owned albums first.~~ DONE, and it needed no measuring at all — both albums

were already fully carded by the 9270acab sweep. **Never count an album's cards by grepping its name

against card filenames**; filenames carry the TRACK title, so that search returns 1 for any album and

it is the reason this sheet first claimed 76 tracks were owed. Count by the source_path inside each

card, and run the count both ways as a positive control.

The --scan caution in the defect_register still stands and is now doubly earned: --scan keys its

skip on corpus rows only, so it re-measures every sibling card from scratch — hours of recomputation

for work already on disk — and it would walk into the EX_051 containment trap on the way. Prefer a

scoped worklist: match first, decode-only, then measure only what is genuinely new.

3. ~~Attach the two inspection rows by hand.~~ DONE in commit 9f4f7a94 — --id EX_014 onto Tekken 2

track 01 at game agreement 1.000, and --id EX_094 onto "Dark Creature Pursuit" with override_open

set and four runners-up named on the row.

4. Check the pool can actually SEE a newly-landed row before trusting any N. Landing a row and

admitting it to the study are two different things, and both failed silently here: a stale

KNOWN_FALSE quarantine withheld EX_014 and its 31 siblings, and a name-keyed leak check let EX_094 be

scored against its own byte-identical twin. derive_patterns.py --self-test must be green at HEAD, and

n_hits / n_ctrl must move by the amount the ingest predicts — if they do not, the pool is lying.

5. Re-run the derivation at the new N — and re-derive its own bar for its own N, never inherit V2's:

python harness/music_gen/derive_patterns.py --perm 400 --json build/audio/exemplars/PATTERN_FINDINGS_V3.json

6. The per-class bands become statable for BOSS and TENSION for the first time. **State

TITLE_CHARACTER's band with the 21.5 s sting named in the band itself**, per the caveat in PRIORITY 2.

Run the leave-one-album-out check V2 §12 asks for while the pool is fresh.

---

PROVENANCE

Assembled 2026-08-07 against docs/spine/DECISIONS_PENDING_JOSH.md (commit cd2619f7),

build/audio/exemplars/BUY_SHEET.md, build/audio/exemplars/PATTERN_FINDINGS_V2.md §§9–12,

build/audio/exemplars/CORPUS.json (tracks, counts, defect_register), and the actual contents of

D:/audio/exemplars/. Every price fetched live 2026-08-07. Every row marked VERIFIED or ESTIMATE.

No product URL was invented.

**REFRESHED 2026-08-07, after commits 9f4f7a94 (the two free re-keys) and this one (the sibling

wiring).** Priority 2 moved from a plan to a landed result; the pool table's "today" column is now read

off a live derive_patterns.py --perm 400 run rather than projected; two false card-count claims are

corrected in place with the counting defect that produced them named; the 21.5 s EX_014 caveat is

carried into the class table, the ingest steps and the band-stating instruction. **No price, URL, store,

verification date or total was touched** — the capped $57.05 list is exactly as verified on 2026-08-07

and nothing on it has been bought.

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