19-scoring.md

19 — The scoring: 36 hit, 29 half, 5 miss

Measure: the 70 clauses of 00 — The predictions, scored against what was measured. A half is a clause whose substance holds and whose detail does not, or the other way round. The totals in the heading are counted out of the tables below by python tools/checkscore.py docs/19-scoring.md, not added up here.

---

The arithmetic

hit        36
half       29
miss        5
unresolved  0
total      70

score      36 + 29/2 = 50.5 of 70 = 72.1 %

C70 predicted 49.5 of 70 = 70.7 % and got there through 41 / 17 / 12 instead of 36 / 29 / 5. The total was a point low; the distribution was wrong in the same direction it has been wrong for five sessions running, and this time it was wrong by more than ever: twenty-nine halves against seventeen predicted, and five misses against twelve. The fifth consecutive session in which the total is close and the shape is not.

The shape is worth reading rather than lamenting. Almost every half has the same form: a clause whose substance was right and whose number came from §A, or whose substance was right and whose proposed mechanism was wrong. Clauses that bet on the object were mostly right; clauses that bet on the briefing's arithmetic were mostly half.

C70 also predicted "between 5 and 11 errors in this briefing, with 8 the single most likely value". There are eight. That is the first exact hit on that sub-clause in this branch.

---

The frame, the mode, and the missing XA

#	verdict	kind	what happened
C01	hit	method	rawcensus.py ran unmodified and passed 64/64 on both halves of its self-check at the first attempt on both images; mode1.py was declared inapplicable and not adapted; both facts are one paragraph in 16.
C02	hit	method	629,766 sectors censused whole, every field, zero mismatches of any kind; EDC re-derives 322,527/322,527 and 307,239/307,239.
C03	hit	content	Mode byte 2 on 100 % of both, Form 2 on zero sectors of 629,766.
C04	half	content	The two counts coincide exactly — 1,712 and 1,718, all-zero user area and empty ECC field, same sectors. The distribution claim is badly wrong: the 16-sector system area plus the 302-sector tail run is 318 of 1,712 = 18.6 %, not "more than 80 %".
C05	hit	content	12.9252 % on both, cited rather than re-proved. Parity against file bytes: 13.5441 % and 13.5484 %, 0.0043 points apart and both above 13.4 %, exactly as bet.
C06	half	content	The accounting closes with residue 0 on both discs, which is the substance. The decomposition is seven terms and not the four named — the system area and the two descriptors were forgotten and the 304 unclaimed sectors are not one term but two, of which 302 are outside the volume space entirely.
C07	hit	content	No CD-XA001 on any of the four descriptors; 0 of 2,230 and 0 of 2,260 records with an XA System Use area; subheader 00 00 00 00 on 629,766 of 629,766, censused and not sampled.
C08	hit	content	Submode 0x00 everywhere, data bit clear on every sector of both discs, stated as a defect of declaration; EDC verifies on all 629,766.
C09	hit	content	"A Mode 1 disc in a Mode 2 wrapper", byte cost zero, 2,580,216 + 2,457,912 = 5,038,128 bytes spent saying nothing, and the offset-24 consequence is what killed recdates.py. All four figures exact.

The descriptors, the namespaces, the holes

#	verdict	kind	what happened
C10	miss	content	The escape sequence is %/@, UCS-2 level 1, not %/E. There was no evidence for level 3 and the clause picked what sounded like a 1999 disc.
C11	hit	content	20 divergent names on Leon and 20 on Claire for the same files, 0.91 %, and not one of them is a CAPCOM file — every long name on the object belongs to a third party.
C12	hit	method	Joliet costs 91 sectors = 186,368 bytes on Leon and 93 = 190,464 on Claire, inside the predicted 40–120 sectors and under 250 KB, published with the command that computes it.
C13	hit	content	Extents identical across namespaces on all 2,194 and 2,224; zero files in one namespace only; delta 0 bytes.
C14	half	content	The interpretation is right and better than predicted — it is a complete map, 37 pairs, all 37 of 37 accounted for, not "more than 20 of 28". The count of 28 dwords was wrong: there are 75.
C15	hit	content	The run begins at the last sector of the volume space; exactly 302 sectors lie past it; string, dwords, then 302 zero sectors.
C16	half	content	Identified, which is what the clause promised — the two dwords are the block's own LBA in both-endian form. But that is none of the three quantities the clause named, and it also missed the AA 55 55 AA terminator entirely.
C17	hit	content	Four-sample table published (150 GEAR, 302 Toast, 302 CeQuadrat, 302 CeQuadrat), question left open, mechanism not named.
C18	hit	content	Eighteenth answer, and the first where the descriptor withholds rather than merely omits: two sectors inside its own declared volume space, written by its own author, reachable by nothing.

The dates

#	verdict	kind	what happened
C19	half	content	835 days on both, exact. The rest is wrong twice: the oldest files are different files on the two discs (EM03F.TIM and EM13F.TIM) and both are CAPCOM's, not third-party.
C20	hit	content	Zero records at 00:00:00 on either disc, and the difference from 1997 got one sentence.
C21	hit	content	2,194/2,194 and 2,224/2,224 even, 30 distinct values, and no odd second anywhere on either disc in either population.
C22	half	content	The first half is exactly right, including which timestamp: Claire's creation, 14:35:11.95, seconds = 11, odd. The conclusion drawn from it is exactly wrong.
C23	miss	content	The 36 directory records are dated 1999-03-21, on both discs, inside a five-minute window. They are the burner's work, not inherited, and the clause bet the whole way on the opposite.
C24	half	content	The four re-derive within a second; the shape survives widening to 72 stamped PE; the CAPCOM cluster is 4 and 5, inside the predicted 4–8. But "RESIDENTEVIL2.EXE is not one of them" is true on Leon and false on Claire, where it sits at −32,401 s with the rest.
C25	half	content	The 33 days to the volume is right and the side-by-side publication happened. The direction is backwards: Leon's executable is newer than the other CAPCOM binaries by 26 days, not older, and that is the whole finding.
C26	half	content	Right that it is not a Japanese build clock and right to leave it open, and the demonstration is stronger than predicted — yesterday's +9:00 group was NEC, Japanese, and today's is Rendition, Inc., American, so the sign cannot track the build machine's country. The mechanism the clause proposed was not demonstrated.

The two discs

#	verdict	kind	what happened
C27	half	method	The partition closes and is asserted, which is the method half. The numbers are the wrong ones: by path the identical count is 1,547 files and 341,133,374 bytes, not 1,696 and 357,785,576 — those are the hash-set figures, and the clause conflated two different measurements.
C28	miss	content	There are no COMMON files with the same name and different bytes. All 1,005 are identical on both discs. The clause inherited the briefing's largest error and built a paragraph on it.
C29	half	content	All five denominators re-derive exactly, and 54.4374 % and 57.1645 % are exact. 27.8837 % is wrong in the fourth decimal: it is 27.8838 %.
C30	half	content	"COMMON, ZMOVIE, GALLERY and DRIVERS are 100 % duplicated" is exactly right. The PL0/PL1 figures are the hash-set ones again: by path they share 136 files and 8,404,946 bytes, 2.65 % rather than 8.7431 %, so the scenarios are 97.3 % distinct rather than "more than 91 %".
C31	half	content	The verdict — not waste — is argued with the structure of what is duplicated and with the readme's 1 MB install. But the clause promised either a reference count or an explicit declaration that the reference graph was underivable, and neither was delivered; the argument runs on directory names, which is the escape route without the declaration.

The film

#	verdict	kind	what happened
C32	hit	method	Extended to classify every movi chunk; found 00iv, 01wb and JUNK; 52,780 video chunks on the 32 movies, matching dwTotalFrames on 32 of 32 (and 34 of 34 with SAMPLE.AVI); the extension and the old tool's silent DISAGREE are both written down.
C33	half	content	IV50 and 320×160 on all 32; the 15/30 split is exactly 15 and 17; the cross-tabulation was published. The predicted mechanism is wrong: the split is not by directory group, it is by disc — Leon's ten scenario films are 8 at 30 fps and Claire's are 3.
C34	hit	content	160 lines against 240 is one third fewer pixels; the counterfactual was computed and published — 833,967,264 bytes at 240 lines.
C35	hit	content	2,532.5314 s, not round; distinct film 518,060,506 bytes and a distinct duration below the total, both published.
C36	hit	content	219,536 bytes/s (predicted 215,000–224,000), 1.4293× (predicted 1.40–1.46), worst single file 1.95× (predicted under 2.0). Three for three.
C37	hit	content	The 4× is not a video constraint; the gap is 2.80× against 1997's 1.50×; two-row table published with the argument.
C38	half	content	Byte-identical between the discs and one of the two shared root files: right. The frame-size comparison was never run, and the file turns out to be the CAPCOM logo animation — identified by someone looking at it, not by this session, and recorded as such.
C39	hit	method	Zero frames decoded, zero written to _work/, zero committed, and the chapter says so in its own words.

The voice

#	verdict	kind	what happened
C40	miss	content	The first dword is not a count. It is a bit mask, and its population count is the stream count — WEAPON01.SAP carries 481 and holds five streams. The closure half of the clause is stronger than promised (1,495 of 1,495 rather than "more than 1,450"), but the clause's headline claim is simply wrong, and §B3 built a careful three-paragraph argument for it without opening one file.
C41	half	content	"Small, under 65,536, not a size" holds and the distribution was reported without over-reading. But the second dword is zero on 1,377 of 1,495, not "more than 1,400", and it is not a flag or a bank id — it is the high half of the same bit field, with bit 48 the highest used.
C42	hit	method	Decoder written from the public definition, coefficients read per file, validated 19/19 against a scalar reference before any census, and every one of 5,195 streams closes with residue 0 bytes and a decoded sample count equal to blocks × wSamplesPerBlock. The Lands of Lore standard, met on compressed audio.
C43	half	content	The deduplicated duration, 9,919.383446711 s, lands inside the predicted 7,000–11,000, and the naive 18,450.147 clears both thresholds. Two numbers are wrong: it is 3.67 times yesterday's speech, not "roughly twice", and the per-sample ratio is 3.953 : 1, not "exactly 4 : 1".

The pictures and the models

#	verdict	kind	what happened
C44	hit	method	Signature classifier finds six genera where the clause promised at least four, assigns none by extension, and covers 127 of 127 on each disc.
C45	half	content	3,584 slots, last entry equal to the file size, residue 0, payload not a TIM: all right. Two things are not: 1,287 slots are used, not "under 1,200", and the entries are non-decreasing rather than strictly ascending, because unused slots repeat the previous offset.
C46	hit	content	Entropy-coded and not the same algorithm; pakdec.py raises on all of them; the five-indicator method was applied and all five readings published, including the one that failed.
C47	miss	content	.ADT was not opened. The clause's own escape route was taken in full — five indicators published, twenty-four decoder configurations named, the failure described — but the clause said "I will get .ADT open" and it is not open.
C48	half	content	1,455 blocks, 1,445 TIM, 10 TMD: exact. The verification method promised — pointing psblocks.py at yesterday's tree in the same run — was not performed; the collapse was verified by a different route, emdhead.py's TMD validator returning 0 of 337.
C49	hit	content	.DAT 99.83 %, .DIE 57.78 %, .TS 91.95 %; with .TIM they total 19,893,182 bytes on Leon, above the predicted 19 MB; GALLERY is 114 TIMs and ROUGH58.DAT is the false positive.
C50	half	content	The container reading is right — 45.82 % validated TIM in 238 of 250 files, and the other half is not pictures. The table claim fails: the dwords at offset 8 are not monotonic, and all-non-zero-inside-the-file holds on 186 of 250, not "more than 450 of 500".
C51	hit	content	The head is not a magic number; the first dword is a length and an offset; 51 distinct heads is what a length field looks like; and which of the two it is was settled by testing against the file sizes on 169 of 169 files.
C52	half	content	The negative claim is stronger than promised — 41 00 00 00 validates as a TMD in 0 of 337 files, not "no more than 5 of 58" — and CAPCOM's own format is confirmed. The positive half was not delivered: no vertex table was derived, and the interior is declared underived instead.
C53	hit	content	325,276,136 bytes on the naive sum in formats no Win32 API of 1999 could read, far above the predicted 90,000,000; zero AVI or WAVE bytes in a native container and none the other way.

The layers, the paths, the personal data

#	verdict	kind	what happened
C54	hit	content	Seven layers close with residue 0, and the third-party fraction is 5.3744 % and 4.8923 % against 3.6814 % — the two predicted figures, exact to four decimals.
C55	hit	content	DirectX 6 and DirectXMedia 6 prerequisites, drivers a leftover, same reading in opposite directions; Indeo 5.06 ruled a prerequisite on the ground the clause named — 555,978,176 bytes unreadable without it.
C56	hit	content	The readme does not call them development versions; the ruling was re-made on 3 of 14 chipsets and the check with your card manufacturer sentence; leftover, with a different argument.
C57	half	content	11,490,859 bytes and 1.2418 % both land inside the predicted ranges, most of REGIST is prerequisite, and the duplication gets its own line. One half is wrong: 1.2418 % is higher than Il Mio Computer's 1.2117 %, not lower.
C58	hit	method	Ran intact on the Joliet tree, printed 252 hits, 155 distinct, 67 files on Leon, and the tree is named on the same line as the number.
C59	half	content	The published figure is the clause's second one — 0 build-machine paths in 1,481,209,632 bytes, after checking all four CAPCOM binaries' debug directories. The clause led with 2 and the two c:\Program Files\CAPCOM occurrences were ruled a default, so half the clause argued against the other half.
C60	half	content	Three signed, all VeriSign chains, corporate subject names: right. Intel's is not among them — IV5PLAY.EXE carries no certificate table, and all three signed binaries are Microsoft's.
C61	half	content	Demonstrated exactly as promised: KUO.JGGG at offset 163,960 inside a pixel block running 532..167,242 in a TIM that closes with residue 0, and a second wider pattern over all 558,219,896 bytes of Indeo found zero domains. The leakcheck.py half did not happen — the tool found its one address and reported CLEAN, so its refusal behaviour never fired.

The crossings and the Saga

#	verdict	kind	what happened
C62	hit	method	12 files and 352,267 bytes re-derive, identical set on both discs; the largest of the twelve, ATIM8.DRV, was compared 129,584 bytes at a time against the 1997 image and found identical by comparison and not by hash; and the overlap with Final Fantasy VIII was re-derived from the crossing output rather than from the briefing — which is how it was found to be eleven, not the ten the briefing stated.
C63	hit	content	222 crossings, far above 12, and every one of them is in REGIST/DIRECTX. Zero game data.
C64	hit	content	Both cells filled, rule restated as series membership plus a measured second member, byte criterion kept for the crossings column and retired from this one, and the reasoning written into the index and into 13 rather than applied silently.
C65	hit	content	Seven structural traits published rather than five, and Final Fantasy VIII shares zero of them rather than the predicted "at most one" — including a trait the clause did not anticipate, that both games carry one CORE00.ESP in the engine's data directory and share none of its bytes.

The thesis and the meta-bets

#	verdict	kind	what happened
C66	half	content	Six denominators published, the deduplicated one is the headline, and the coincidence with 68.5388 % survives on the naive sum to 0.44 points — all as bet. The range is wrong: the headline is 75.3264 %, outside the predicted 66–73 %.
C67	half	content	Two columns published, and the per-second ratio is 12.38 : 1 against a predicted "about ten times". Right in kind, out by a quarter.
C68	hit	content	Runtime from 2.5632 % to 4.5841 %, drivers from 1.1182 % to 0.7903 %, chipsets 6 → 14, both bills published side by side with the same layer names. Every figure as predicted.
C69	half	method	Re-deriving rather than incrementing is what caught the problem: the index had moved on since the copy this session started from, and two upstream commits — one trimming every What it is cell to one or two lines, one removing the title count from the heading — meant the local README was stale and the first of the three commands no longer has anything to match. The work was redone on the current file. The two counts that still exist come out as predicted, 46 rows / 45 write-ups; two rows were touched, this object's and Resident Evil's Saga cell; the What it is cell is 195 characters against the file's current maximum of 204; and the Year cell says 1999 and names which of six dates it is and why. The clause's "45 titles" is unverifiable rather than wrong.
C70	half	method	Twenty documents, the seventh exact twenty in a row; git ls-files clean under the branch's regex; git check-ignore passing in both directions; no decoded asset, no film frame and no GALLERY illustration committed; repository created with topics and a description under 350 characters. The briefing sub-clause is exact — eight errors, the single most likely value named. The scoring prediction is not: 50.5 of 70 = 72.1 % against a predicted 49.5 and 70.7 %, through 36/29/5 against a predicted 41/17/12.

---

What the misses have in common

Five misses, and four of them are the same mistake at four scales.

C10 guessed a field it had no evidence about. C23 and C40 are the serious pair: both are conclusions built by careful reasoning on §A's summary statistics, and both would have been settled in under a minute by opening one file — the .SAP header on C40, the directory records on C23. C28 inherited the briefing's largest error outright and built a paragraph on it, which is the failure the briefing explicitly warns about every session and which happened anyway. C47 promised a result rather than an attempt.

The twenty-nine halves have a different and more interesting shape. The overwhelming majority are clauses where the object was read correctly and a number was taken from §A — C27, C29, C30 all conflated a hash-set count with a file count, which is exactly the error §A itself was making — or where the substance was right and a proposed mechanism was wrong (C26, C33, C41). Very few halves are failures of measurement; nearly all are failures of arithmetic inherited rather than derived.

Which is the fifth session in a row on which the same sentence closes the chapter: contare un insieme di numeri non è leggerli — counting a set of numbers is not reading them — and this time it cost eight clauses.