Inside MvC's Mashable Moves: From the ROM Chip to the Extra Hit

Gold War Machine's Proton Cannon responds to directions. Ryu's Shinkuu Hadoken responds to buttons. Hulk's Gamma Quake keeps dropping the same eleven rocks. Those differences are written into the moves themselves.

The question was simple: how does an attack become mashable, who has one, and could the behavior be added to another move? Answering it took following the input through the 68000 instructions, into a timer or repeat counter, and finally into the attack that reaches the opponent.

A button press never means “add one hit.” The move decides what another press buys. Sometimes it buys time. Sometimes it buys another animation cycle. Sometimes it fires another projectile that misses entirely.

Mashable supers at a glance

Find your character and super here first. Mashable does not mean every press becomes a hit. The table separates extra connected hits from extra attack production, tested paths with no extension, and moves whose behavior remains unresolved.

Baseline, best result, and what to press

“Normal total” below means no extra input after the command, in the printed setup. It is not a fixed number for every opponent and distance. The guide gives the strongest published comparison for each finding and the input sequence that produced it. Exact replay timing is expandable; the plain-language instructions explain the same input rule for a player.

A true maximum is not established for these moves. Spending all twelve Ryu credits or all sixty Chun-Li credits reaches the extension allowance, not a proved ceiling on collisions. Gold's 60-hit report is not a universal limit: the revised ordinary-health trials exceeded it before KO.

Three measured examples

MAME + instrumented FBNeo / observed

What did the extra input buy?

Three moves, different allowances. Each number belongs to the setup printed with it.

Earlier capture · move identity only

Ryu

Shinkuu Hadoken

16→20

first-move connected hits

No extra input → Rotate buttons

Spend a finite allowance12 extra active updates, not 12 hits

Distance matters too. This reproduces a four-hit gain, not the reported 22→26 setup.

Setup & limits

Airborne Hulk; separation 325; first-contact Y 466, ground Y 472.

observed, not maximum. Each core restores its own fixture. Three neutral replays match exactly.

Inspect the evidence ↗

Earlier capture · move identity only

Chun-Li

Senretsukyaku

15→30

first-move connected hits

No extra input → Rotate buttons

Use all sixty credits60 spends + 60 ordinary countdowns

All 60 credits spent. Thirty is demonstrated here; no universal contact maximum is claimed.

Setup & limits

Standing Spider-Man; separation 60.

observed, not maximum. Each core restores its own fixture. Three neutral replays match exactly.

Inspect the evidence ↗

Earlier capture · move identity only

Gold War Machine

Proton Cannon

25→116

FBNeo hit-processing events through KO

No extra input → Rotate directions

Reimburse the firing timer96 + 627 − 723 = 0

MAME: 114 through the lethal event; FBNeo: 116. The separate full-window count of 181 includes post-KO events. Both allocated 180 missiles; allocations are not hits.

Setup & limits

Standing Spider-Man; separation 70; initial HP 144.

observed, not maximum. Each core restores its own fixture. Three neutral replays match exactly.

Inspect the evidence ↗

New MAME 0.287 and custom instrumented FBNeo comparisons, stock Euro 980123 program. Health, meter and positions prepared before saving; input replays make no gameplay-state writes. Images identify moves from earlier captures, not the new numerical outcomes.

These comparisons restore a fixed setup within each emulator. They are observations, not maximum combos or a ranking of the moves. Gold's card uses 116 events through the lethal hit in FBNeo; MAME recorded 114. Its separate full-window total of 181 includes processing after KO. The sections below connect each result to its code; the evidence package contains the player guide, input harnesses, raw observations, runtime logs, ROM identities and selected instruction excerpts.

Follow the code for your move

Start with the named findings below. Each verdict carries its input rule, code locations and limitations. The complete native-table inventory further below exposes the remaining naming and reachability questions as well.

Audit the native routes

The named guide above answers the findings we can explain directly. The code census below makes the wider coverage inspectable: normal contexts, special/super selectors, throws and character styles. A row can identify a real table and handler while still having an unresolved human command or conditional branch. Those gaps remain visible.

Challenge the numbers

Ryu is mashable, and distance matters. Those statements can both be true. With the same standing Spider-Man setup, fresh buttons changed 9 to 10; with the same airborne-Hulk setup, they changed 16 to 20 in MAME and instrumented FBNeo. Moving the opponent independently changed the baseline. We did not reproduce the exact reported 22→26 macro.

Thirty Senretsukyaku hits are demonstrated. Sixty is not a universal Gold Proton ceiling. Maximum Spider's alleged 25 is unsupported by these tests. Each claim below keeps the controls and scoring boundary alongside the number.

The input recipe rotates individual buttons at emulator callbacks; the trace separately counts actual native input samples and move updates. That distinction matters during hitstop. Holding buttons, using the wrong mask, and introducing new presses before or after the active window are controls. The outside-window receipts contain actual sampled edges, zero credit spends or timer reimbursements, and unchanged first-move hit counts for Ryu, Chun-Li and Gold in both cores. Late Chun-Li presses caused later normal attacks; those hits are excluded.

A maximum needs both a code-derived bound and an attainable example. This study establishes finite update or repeat budgets where the code supports them, and reports the contacts observed under each setup. It does not promote a best search result into a universal contact limit. The community reports are paraphrased hypotheses, not independent trace evidence. Protocol and disproof controls.

Gold War Machine: directions buy more firing time

Gold shares War Machine's top-level character code, but he does not share every attack implementation. Quarter-circle forward plus two punches selects +$E6=$02. At $02A5A4, character ID $28 branches to Gold's routine at $02A7E2.

Its firing timer starts at 96 at $02A8E2. The helper call at $02A9D0 accepts new directions, and $02A9D8 adds one to the timer. The ordinary update subtracts one at $02A944. One accepted direction reimburses that decrement.

A separate counter schedules a missile every fourth firing update. Keep the firing state alive longer, and the emitter gets more opportunities:

// Simplified firing-phase logic; surrounding gates omitted.
if (new_direction()) firing_timer += 1;
if (++emission_counter % 4 == 0) launch_missile();
if (--firing_timer == 0) begin_recovery();

To extend it: finish QCF + two punches, then continuously change directions while the cannon fires. The strongest recorded script cycles Down → Left → Up → Right, one cardinal per emulator callback with no neutral gaps. A held direction stops producing fresh edges; attack buttons do not reach this timer-extension branch. The timer must still be in its active firing state. Human stick rotation follows that rule, but the exact scripted count is not a promised hand-speed result.

The earlier ordinary-health fixture produced 25 hits without extra input and 40 with direction pulses. The firing state also lasted longer. Holding all buttons produced 23 hits and tapping them 24 in that earlier investigation: missile placement uses randomness, and contact totals can vary without an extension. The claim rests on the timer code and the longer firing state together. Those earlier controls use a different fixture from the revised dense-cardinal comparison. Gold and War Machine receipt.

The revised dense-cardinal test went further. With ordinary initial health 144, both engines counted 181 hit-processing events during the first move. MAME had 114 through the lethal event, then 67 after KO; FBNeo had 116 through the lethal event, then 65 after KO. The displayed counter saturated at 99. A successful missile allocation is a different event: both recorded 180 allocations.

The timer arithmetic closes: 96 initial + 627 reimbursements − 723 decrements = 0, followed by recovery. There is no separate finite mash-credit counter in this inspected loop, but the run ended. The launch-phase byte wrapped twice in MAME; that cycling byte is neither the hit counter nor an infinite-time proof. No tested reimbursement wrapped the timer from 255 to zero.

The visibly separate 1024-HP diagnostic produced 166 MAME / 149 FBNeo connections, despite matching 165 allocations and 565 reimbursements. Its initial RNG seeds differ between cores; both log native RNG changes without freezing them. It cannot stand in for an ordinary-health result. Counts before the first move ends exclude later buffered attacks; this controlled-window attribution does not claim a general per-projectile collision-owner trace. Revised Gold events, RNG and limits.

Ordinary War Machine's Proton Cannon follows another branch. Its timer starts at 120, decrements normally, and lacks Gold's reimbursement call. Its four input conditions all gave 27 fixture hits. Compare conditions within each saved character fixture; Gold and ordinary War Machine were captured on different stages.

Ryu: the style and the super both matter

Shinkuu Hadoken initializes twelve human mash credits and a 64-update active timer. The relevant instructions are in the shared beam routine at $0478D2.

At $047AB6, an accepted attack-button press spends one credit at $047ABE, then branches past the timer decrement at $047AC4. The ordinary animation update still runs. After the credits reach zero, every eligible update advances the timer toward the finish.

Twelve credits can therefore buy twelve extra active updates, not twelve guaranteed hits. The original standing-Spider-Man fixture went from 9 to 10, independently reproduced in the revised MAME and FBNeo tests. Akuma-mode Messatsu Gou Hadou uses the same extension and also went from 9 to 10. The harness reached the styles through the game's ordinary transformation commands and checked the resulting style values. Ryu and styles receipt.

The new geometry grid separates distance from input. Against standing Spider-Man, initial separations 62, 160 and 325 produced 9→10, 20→21 and 15→17 in MAME. More distance did not monotonically add hits. Against airborne Hulk at realized separation 325, both cores produced 16→20, with first-contact Y=466 versus ground Y=472. A case nickname such as “air16” does not prove the opponent was airborne on impact; first-contact state and coordinates are in each receipt.

Standing Hulk at separation 325 produced the largest published matched Ryu result, 20→24 in MAME. At separation 160 he produced 22→22 even though the input mechanism exists. That counterexample matters: spending the beam's allowance need not change connected hits. After QCF + two punches, rotate fresh attack buttons during the beam; the measured rotation is MP → HP → LK → MK → HK → LP. All twelve credits can be spent, but holding the buttons or pressing after the active window buys nothing. The exact community 22→26 setup remains unreproduced. All Ryu distance and jump fixtures.

Ken-mode Shinryuken spends the press differentlyAnimation advancement, measured against the same saved opponentCONFIRMED IN GAME

native combo peak	12 → 15
input read	$048428: fighter + $D4
accepted mask	$0770 — attack buttons
effect	$048432: animation delay = 1
next step	$048438: ordinary animation update

The rows describe the paired experiment and code path; the still identifies the move. The peak result is not the counter at this particular capture.

Stock Euro 980123 • MAME 0.287 • ryu/ken_shinryuken.csv and .log in evidence archive

Ken-mode Shinryuken uses an inline input test. A fresh attack-button press sets the animation-delay word to 1, causing the next animation step immediately when the normal updater runs. Upward movement continues separately. Faster animation creates more attack opportunities during that ascent; the measured combo rose from 12 to 15.

Base-Ryu Shin Shoryuken is a different move and did not gain hits here: 3 across all five conditions. Shinkuu Tatsumaki Senpukyaku stayed at 18. Its loop consumes a fixed count at an animation boundary without the tested beam's press reimbursement. Similar names are not evidence of shared mechanics.

Hulk: follow the rocks, not just the hit counter

Hulk's three supers showed no attack-button extension of hits, active time or emitted attacks in the examined stock paths. This is backed by fifteen paired trials and by following the child objects—not simply by failing to find a shared-helper call.

Gamma Quake initializes a controller timer to 48 at $0A96DA. Each update decrements it. When the nonzero timer is divisible by four, the controller emits a falling rock. The qualifying values are 44, 40, 36, down to 4: eleven rocks.

All five input conditions allocated eleven rocks and 22 debris objects. Only seven rocks' attacks connected in this fixture. Eleven scheduled objects and seven connected hits answer different questions. The rock positions also use randomness. Hulk receipt and allocation addresses.

Gamma Wave follows a travelling controller with a fixed child-emission cadence. The allocation trace counted eight wave children in every condition; its fixture combo stayed at eleven. Those children can remain active after Hulk recovers, so the result includes their continuation.

Gamma Crush reads a held, facing-relative direction once to choose the descent trajectory. It does not count repeated presses. Its four-hit fixture result stayed unchanged under button mashing. Changing the trajectory can change which attacks connect; that is aiming, not a bonus-hit counter.

Hulk does have a separately commanded Gamma Charge follow-up, and a bounded light-kick repeat path was found in his normal-attack code. Those are distinct input mechanics. They do not establish a mash extension for his supers.

Chun-Li exposes the two simplest designs

The first experiments in this investigation used Chun-Li and five input patterns: release, hold, light-punch taps, all-button taps, and directions. Kikosho went from 9 to 13 hits; Senretsukyaku went from 15 to 25. Holding and directions did not produce either increase.

Both supers call the same button helper. What happens after it succeeds is the revealing part.

Kikosho buys child-object repeats. At $04E2A4, the player gets eight credits at +$398. Each accepted press spends one. Every second credit follows the pointer at +$392 to the energy field and increases that object's repeat counter. Eight accepted presses can buy four additional repeats. The child routine at $09879A..$0987CE rearms the attack and consumes the repeat counter.

The original Senretsukyaku 25-hit observation left eighteen credits unused. It establishes an input-sensitive extension, not a 25-hit ceiling.

Senretsukyaku buys time. The human path starts with sixty credits. A press spends one at $04E44E and returns before the active-phase countdown at $04E454. The animation has already advanced earlier in that update, so the existing kicking machinery gets more time before the finish. The credit limit keeps this particular extension finite. Chun-Li code and paired-trial receipt.

In the revised Senretsukyaku fixture, alternating on/off LP or all buttons reached 26 and spent 45 credits. Rotating individual buttons reached 30, spending all 60, in both cores. Holding and direction controls remained at 15. The trace records sixty ordinary countdowns in each condition: qualifying presses postponed those countdowns while the attack animation continued. Thirty is now attainable; this is still not a proof that every target and variant has a 30-contact ceiling. New Chun-Li scored trials.

‹›

Release · 15 hitsTap LP · 25 hits

The trace is the time-series evidence. Callback counts measure the observation schedule, not native frame data; hitstop and game speed intervene between those two clocks.

Who else has a supported mash mechanic?

The roster explorer separates live-measured behavior, static code findings, and tested controls. “More attacks” includes extra objects and throw slams, even when the HUD does not count them as more combo hits. Each entry links to its research receipt.

Wolverine's Berserker Barrage accepts fresh punches into a continuation latch at $0352D8..$0352E4. The later cycle checks the latch, within a strength-dependent finite repeat limit. Its light-punch fixture rose from 2 to 4 hits. Tornado Claw accepts current or buffered punches in specific animation windows and enables further stages; it went from 3 to 5.

Morrigan's Soul Eraser has two distinct responses. Buttons reimburse up to ten timer ticks and increased the fixture combo from 18 to 20. Directions set a flag that doubles a satellite's orbit step on that update. They accelerate satellite motion; they do not add button credits. Shadow Lady's Big Bang Laser uses sixteen button credits to delay the active countdown, measuring 21 to 25. Additional roster receipt.

Hyper Mega Man and Hyper Roll shorten a projectile scheduler by four additional ticks when a button press qualifies. Successful projectile allocations rose 14 to 23 for Mega Man and 12 to 20 for Roll. Their connected combo totals remained 17 and 15 in these fixtures. An earlier version of this article had only a static trace of that scheduler; the follow-up reached and instrumented both moves.

Red Venom's Venom Rush collects fresh kicks in a local buffer, then checks them at an animation boundary. A qualifying kick reimburses the loop decrement, permitting another attack cycle. It measured 2 to 7 hits with kick taps; punches and directions stayed at two. Holding all kicks bought one continuation because MK/HK were newly introduced after the LK command, then stopped helping. The ordinary Venom control remained at two across all five conditions. Red Venom receipt.

Venom's Web Throw illustrates why the combo byte alone is insufficient evidence. Twenty-two qualifying new directions or punches unlock one extra two-slam animation loop: two slams become four, while the HUD remains at one combo hit. The extra-slam state changes the damage attributes too; the no-input and button-tap trials ended with equal total damage. Red Venom has the corresponding throw code, but his Web Throw was not separately tested live. Throw receipt.

Strider's Ouroboros has a code-confirmed punch-triggered projectile path in the orbiting objects. A qualifying owner punch fires a shot, then applies a cooldown. It does not refill the 240-update effect lifetime. This is a static finding here, not a claimed live hit-count comparison.

Maximum Spider and the two Beat Planes

Maximum Spider was tested as Maximum Spider: QCF + PP, selector $08, its own dispatch and attack passes. The examined MAME and FBNeo conditions produced five first-move hits, including fresh buttons and directions. This supplies no support for a 25-hit mash claim. Spider-Man's separate Web Throw rotation routine cannot establish that result, and five observed hits is not a proof of a global five-hit maximum. Move-specific handlers and MAME controls, independent FBNeo events.

Mega Man and Roll's Beat Plane use QCF + KK, selector $08. In the tested MAME paths, fresh buttons initiated four additional firing objects at allocation store $0A05E4; neutral and held conditions initiated none. The vehicle lifetime still consumed 180 decrements. This is gated firing during the existing vehicle state, rather than extending its duration; steering remains a separate response.

Both vehicles reached 3→23 first-move hits in the MAME and FBNeo comparisons. Roll's raw FBNeo aggregate was 24 because it included a later hit after the first move ended. Scoring that boundary consistently gives 23. These findings are specific to Beat Plane, separate from the Hyper Mega Man / Hyper Roll schedulers above. They neither reproduce nor disprove a 36-hit community example under unknown conditions. Vehicle code and allocations.

Normals: another attack, another classification

The normal-attack census separates restarting or continuing a normal from extending one super. Independent instruction review found local repeat paths in War Machine's standing LP; Hulk's standing and crouching LK; selected Wolverine normals; Morrigan's standing LP; Captain America's standing LP/MK; Venom's standing LP/LK; and Spider-Man and Jin's standing/crouching LP. The ledger records their exact masks, animation gates, shared-call frontiers and whether an effect has been measured in play.

Hulk's two LK initializers write one to +$3A1 ($0306D0 / $030772). A qualifying new or buffered LK spends that allowance and restarts an animation ($030AAC / $030BA6). That is a one-use local continuation, not proof of a mashable Gamma Quake. Captain America's standing MK advances from local state zero to two after its continuation, bypassing that same input branch on the next update.

Wolverine supplies an especially useful counterexample. His air-MP code reads the same input word in both $BA branches, but +$BA=0 returns before the local attack reset and animation restart. The nonzero branch can reach that restart. An input reader alone is not enough to label both paths mashable. Shared global gates and human reachability remain qualified in the independent normal review.

Findings that remain separate

Jin and Zangief have direction-sensitive grab routines affecting animation or damage parameters. Their exact grab identities and live damage differences were not established in this pass. Spider-Man's Web Throw collects directional rotations for additional spin continuation and throw parameters; it is not yet evidence of extra damaging combo hits.

Lilith has copied helper code, but the tested Brilliant Shower did not execute it. Its changed contacts under directional input do not prove that this unreachable-looking helper is the cause. Ordinary War Machine's War Destroyer also produced an unresolved contact difference without a longer launch phase in the traced code. Neither observation belongs in a positive mash-extension list until the causal path is established. Roster caveats, War Destroyer comparison.

This is a supported set of named findings, not an exhaustive negative verdict on every unlisted move, airborne version, team combination or CPU path.

Follow the rule into the board

The target is the unmodified Euro 980123 arcade parent set, commonly named mvsc.zip. Every one of its 21 members matches the stock identities recorded by MAME. The member names and hashes matter more than the archive's filename: another revision or a modified set can have the same ZIP name. ROM provenance.

The 3D location map uses readable, simplified packages positioned from the board photograph, with program sockets, audio and graphics ROMs, connectors and the processor identified. Selecting a program window highlights its package and the corresponding address range. Top and focused views make the labels easier to inspect; the photograph and schematic remain available beside it.

The board photograph is a physical reference, not proof that its installed program chips contain this Euro revision. Its visible MVC 03D and 04D labels differ from the tested mvce.03a and mvce.04a. The map distinguishes a socket on the photographed board, a ROM member, and an address the CPU executes. Those are related locations, not interchangeable numbers.

The main 68000 is integrated into the B-board's DL-1525 custom chip. The detailed board notes in MAME and a physical decapsulation investigation identify that package. This matters when tracing the path from a program EPROM to an executing processor; the plastic shell alone does not identify the silicon. MAME PCB notes, DL-1525 decapsulation.

The shared input helper and the player-side mash routines shown here lie below $080000, in the first 512 KiB program window supplied by 03. Child attack logic such as Kikosho's $0986FA routine lies in 04, the next window. CPS-2 instruction fetches need the decrypted opcode view; ordinary absolute ROM data reads need the data view. A displayed CPU address is therefore an execution location, not a ready-to-paste patch offset or the encrypted bytes in a chip dump. Selected instruction bytes.

For a concrete example, the button-mask instruction's immediate word at CPU $01A3DE resides in mvce.03a. Its aligned raw dump bytes are 65 A0; correcting byte order gives the ordinary ROM word $A065; decoding the instruction gives $0770. The board explorer shows these as separate views. Its bit grid models address space—it is not a microscope image of the ROM die or a claim that ciphertext bits correspond directly to buttons. MAME opcode map, CPS-2 cipher implementation.

Two boards, two different RAM roles

The cabinet inputs enter the A-board through the JAMMA and kick connections, reaching its UD6 / DL-1123 I/O interface. The main bus connects that path to the B-board processor and program ROMs. The fighter's input words and mash credits occupy the decoded $FFxxxx work-RAM region, associated with the A-board UR13 / UR14 HM658128 devices. The B-board UA3 / UA4 HM6264 pair serves the separate sprite-object path. Highlighting that pair as the fighter's mash-credit storage would be wrong.

This logical map adapts Loïc “WydD” Petit's CPS-2 schematics, pinned to commit cbd37b3. Inspect the input sheet, work-RAM sheet, and B-board SRAM sheet. The original schematics are licensed CC BY 4.0 unless otherwise noted; this page simplifies their grouping and adds interactive highlighting. The author's uncertainty notes remain part of the evidence.

The lights mean logical execution or memory access. Explanatory steps identify where the described code and data live; recorded events must include a real program counter or access receipt. They are not measurements of voltage, power or exact electrical timing. MAME explicitly notes that RAM wait states are not emulated.

The bit that makes a press new

At $0136BC..$0136C4, the game computes the difference between the current controls and the previous sample:

newly_pressed = held_now & ~held_previous;

The result lands in the fighter's +$D4 word. A held button produces its initial edge and then stops being new. Release it, press again, and another edge appears. Use the input lab to see the individual bits and the helper's answer.

THE SIGNAL BENCH / KIKOSHO RULE MODEL

Press. Trace. Find the extra attack.

Follow one input from its changing bits to the code that grants another opportunity.

Try the ruleRecorded events 4

EXPLANATION / code-derived rule model

01 / MAKE A NEW PRESS

Eight credits. Four extra repeats.

Reset

↑←+→↓

LPMPHPLKMKHK

Six attacksLP onlyDirection only

Sample active update →Release & sample

0Waiting for a samplePress LP, then sample twice.

New-input word · +$D4$0000

HK0MK0LK0ST0HP0MP0LP0↑0↓0←0→0

new = held & ~previous
eligible = (new & $0770) ≠ 0
spend only while credits remain

PLAYER +$3988 / 8 credits

CHILD REPEAT OPPORTUNITIES0 extra

One additional repeat per two spent credits

Assumes human-path gates and Kikosho’s active window pass. This samples the input/credit rule; it does not emulate collision, hitstop or the whole game.

CPS-2 / location → address → instruction

Where the rule lives.

Physical reference

3D location mapPhotographSignal map

4E · DL-1525
68000 core345678910

Selected packageROM 03 left program column · row 6

mvce.03a · 512 KiB · CPU $000000–$07FFFF

Amber · inspected chipCyan · explanation

A · input UD6A · work RAMB · DL-1525B · ROM 3

The photograph locates B-board components. The signal map separates A-board input and work RAM from the B-board.

Jump to a verified code locationStore the fresh-input word in fighter RAMNew direction helperNew attack-button helperANDI instruction testing attack-button maskLogical attack-button mask immediate, 0x0770Gold War Machine direction-helper callRyu Shinkuu Hadoken button-helper callChun-Li Kikosho button-helper callChun-Li Senretsukyaku button-helper callMega Man Hyper Mega Man button-helper callRed Venom Venom Rush repeat-credit incrementKikosho child repeat count initializationKikosho child repeat processingRyu Shinkuu Hadoken: spend one mash creditRyu Shinkuu Hadoken: decrement the active timerStore the connected-hit counter, capped at 99

0304050607080910

PROGRAM ROM 03mvce.03a

$000000$07FFFF

$0136C0 · decoded instructionmove.w d0,$d4(a6)

Address bits, raw bytes & chip receipt

Capacity512 KiB / 256K words

CRC32824E4A90

Photo stickerMVC 03D

Physical positionleft program column, row 6

CPU byte address$0136C0

Chip word address$09B60

Chip word index · 18 address bits

001001101101100000

01 / raw dump bytes8E F4 B2 11

02 / CPU word orderF4 8E 11 B2

03 / decrypted opcode view3D 40 00 D4

Raw CPU-order word

1111010010001110

Decrypted opcode-view word

0011110101000000

CPU-byte offset $0136C0; its first byte is at raw member offset $0136C1 after accounting for word swap. Decryption is a separate transformation.

01Input→02Helper→03Credit→04Repeat→05Contact→

← PreviousNext →

Following explanation

01

Detect the edge

$0136BC–$0136C4

Current controls are compared with the previous sample. Only newly pressed bits survive. This word lives in A-board work RAM, at the fighter’s +$D4.

02

Ask the helper

$01A3D8–$01A3E6

The 68000 executes the ROM 03 helper. Mask $0770 accepts attack buttons. One eligible edge sets carry; six simultaneous edges still produce one yes/no result.

03

Spend one allowance

$04E2FA / player +$398

Inside Kikosho’s active window, an accepted press spends one of eight credits. A press is an opportunity granted by this move’s code, not a command to increment the hit counter.

04

Reach the child

$09879A / child +$50

Every second credit buys an additional energy-field repeat through the player’s +$392 pointer. The child’s repeat-processing instructions live in ROM 04. Eight credits can buy four additional repeats.

05

Collision decides

player +$120

The rearmed attack must still connect. Collision, opponent state and spacing decide the result. This rule model stops at attack opportunities; the separate recorded tests measured nine versus thirteen hits.

The shared attack-button helper at $01A3CC masks that word with $0770: three punches and three kicks. The direction helper at $01A3A8 uses $000F. Both return a yes/no result, through the CPU's carry flag. Six simultaneous button edges are one accepted event, not six credits.

; Human attack-button path, stock Euro 980123
$01A3D8  move.w  $d4(a6),d0
$01A3DC  andi.w  #$0770,d0
$01A3E0  beq     $01A3E8       ; no qualifying new press
$01A3E2  ori.b   #$01,ccr      ; success: set carry
$01A3E6  rts

These shared helpers also check the human-control flag and the +$2B9 gate. CPU fighters take a separate random/table path. The results here concern ordinary human-controlled attacks; a CPU demonstration is not a substitute for that test. Several moves perform their own inline checks, so searching only for calls to these helpers misses real mash mechanics. Input and helper receipt.

Could another move be made mashable?

Yes, through a move-specific modification. The examples provide several designs: extend a timer, add a child repeat, accelerate an animation, shorten an emission schedule, or unlock another attack cycle. There is no universal “mashable” bit established by this investigation.

---
config:
  layout: elk
---
flowchart TD
  input["New qualifying input"] --> window{"Active window<br/>and allowance?"}
  window -->|yes| effect["Change this move's<br/>timer, cycle or child"]
  effect --> attack["Another attack opportunity"]
  attack --> collision{"Connects with<br/>eligible opponent?"}
  collision -->|yes| hit["Normal hit processing"]
  window -->|no| finish["Continue ordinary update"]
  collision -->|no| miss["No added hit"]

A sustained beam must coordinate its owner and child lifetime. A repeated strike must rearm an attack identity so a lingering hitbox is not assumed to damage the same opponent again. A cinematic grab needs authored additional phases and a reachable finish. Each design also needs fresh move-local state and cleanup on interruption.

A useful first patch would pick one move, one accepted mask, one active window and a small bonus budget, then test contact, block, whiff, airborne opponents, hitstop, interruption, tags and DHCs. Calling the existing input helper is only the beginning. No universal patch or new balance behavior is shipped with this article.

For the related display question, the 999-hit counter investigation follows what happens after attacks connect. Counting more visible hits and creating additional attacks are separate pieces of code.

Why the diagrams work this way

The visual design draws on recent research, while the MvC conclusions come from the ROM and experiments. July 2026's VizPilot combines guided explanation with synchronized highlighting and free exploration. The design inference here is to keep a chip's location, its address and the selected instruction connected as the reader moves between views.

Two further papers helped sharpen this revision. Debugging as Evidence-Driven Reasoning, uploaded June 29, 2026 and listed as submitted to IEEE VIS, identifies opportunities for connecting expected behavior, observed state and evidence across artifacts. WireYourWay, a CHI 2026 accepted-paper preprint uploaded March 5, 2026, explores contextual guidance around physical hardware. The design inference here is a shared cursor: selecting an event should locate its instruction, changed value, chip and gameplay consequence together. Those are arXiv upload dates and venue statuses, not interchangeable publication dates. These studies inform the presentation; they do not establish MvC mechanics.

A July 2026 comparison of static, scrollytelling and chatbot onboarding found engagement and guideline-adherence advantages in its pooled interactive conditions, but no significant comprehension difference. Animation is therefore not treated as proof of better understanding: the numbers, labels and written explanation remain available. Code Arcades, from September 2025, informed the overview-to-detail structure; it does not establish that 3D is universally better than a flat diagram.

The page uses Mermaid 12.0.0, released September 10, 2026. Its bundled ELK layout places the process diagram above. The migration preserves this site's existing theme and explicitly retains Dagre for older diagrams, so the new defaults do not silently restyle earlier articles. Version 12 also adds use-case diagrams and per-container ELK choices; those features were reviewed, but this explanation needs a process flow and a separate board explorer. Official Mermaid 12 release notes.

How to inspect and reproduce the evidence

Download the complete public evidence package, or start with the file manifest, selected instruction bytes, and reproduction guide.

The package retains the original Chun-Li, Ryu/style, Hulk and additional roster investigations, then adds the revised scored events, the native-table inventory, exact input harnesses and independent code review. The comparison ledger above derives its case and trial counts directly from the scored records. It omits ROMs, decrypted images, save states, NVRAM and the exploratory whole-ROM linear scan. Each included file has a SHA-256 entry, and verify.py checks the extracted package.

Every live comparison restores a move-specific save. Both players are human controlled. The original fixtures use standing Spider-Man; the revision also varies Hulk, position and native jumping inputs. Health, meter and position are seeded before saving. Some additional-roster probes also choose the pre-match roster ID in selection RAM; their receipts say which ones. The original Gold fixture uses his native secret selection code; the revised core fixtures seed the pre-match character selection and then let native initialization run. The original Ryu style tests use native transformation commands. The gameplay ROM stays stock, and the input trials do not rewrite gameplay state.

The common +$B0=$2A attack state includes specials as well as supers. It cannot identify a Hyper by itself; the harness also checks the actual command and +$E6 move selector. The original Chun-Li comparisons are supers, but the broader roster includes specials and throws. Hulk's published input-read logs are the corrected -traced versions, which exclude reads generated by the observer itself.

An independent event recount checked every core trial and the headline counterexamples without calling the publication scorer. A separate hardware audit checked member identities, raw byte order, instruction samples and the two RAM roles.

The tools were MAME 0.287, an instruction-boundary-checked 68000 disassembly, and a custom instrumented FBNeo schema12 executable. Its build hash and source-base identity are in the revision protocol; this is an independent emulator core, not a claim about an unmodified retail Fightcade client. MAME records pass-through memory writes; FBNeo's pre-instruction hooks compute the imminent value from the verified instruction. The evidence identifies that distinction. MAME's Lua machine reference documents save/load and screenshots; its memory reference documents reads and taps. Those tools make the observations repeatable. The code, write traces and paired inputs together support the gameplay conclusions.

Corrections and open findings · 14 September 2026

• Fixture values are observed, not maximums. The original Senretsukyaku 25 left 18 credits; the revised complete-credit fixture reaches 30. The Ryu 9→10 example used standing Spider-Man, not airborne Hulk.
• Gold's HUD stops at 99; the external event count does not. The revised ordinary-health result separately reports events through KO, later overkill, missile allocations and timer changes. None is an infinity proof.
• Maximum Spider needs its own evidence. Web Throw cannot establish its alleged 25-hit result. Its examined conditions stayed at five. Beat Plane now has separate code, firing-object and runtime evidence.
• Roll's apparent 24 was a scoring-boundary error. One event occurred after the vehicle move; the corrected first-move total is 23. Hulk's wave children demonstrate the opposite trap: a parent-only window can omit genuine delayed child attacks, so his older eleven-hit wave result explicitly includes their continuation.
• The full native-table census is not a fully named move list. Shared style routes, conditional normal branches, unidentified commands and unresolved global callees remain individually marked. No untested row becomes a negative verdict.
• Baseline, best demonstrated and maximum are separate. The player guide now puts inputs and measured totals together. Gold's main result excludes post-KO events; its full-window count remains in the evidence. The super table marks unresolved cases explicitly instead of turning missing evidence into “not mashable.”
• The board view now emphasizes chip locations. The generated scan viewer has been removed. Cited photographs and schematics ground the more detailed location map. Fighter work RAM belongs to the A-board map, distinct from B-board sprite-object SRAM. Recorded lights cite actual events; explanatory lights are labeled as such.

The remaining work is specific: reproduce the exact airborne-Hulk 22→26 setup; close unresolved normal/command reachability paths; establish general delayed-projectile hit ownership; and derive an attainable contact bound before calling any observed total a universal maximum. The article ships the supporting and contrary evidence together so those claims can be challenged.