Compatibility Notes¶

This document tracks what the emulator intends to support, what it deliberately does not support yet, and the game/BIOS quirks inherited from FreeDOCore/4DO. New compatibility work uses References/opera-libretro-master/libopera as the primary behavioral reference; see ReferenceStrategy.md.

Current position (2026-10-29, twenty-first pass): all five golden-corpus titles boot, load their own OS, render real frames, and produce non-silent DSP audio. Seven additional titles were validated on 2026-10-26 (Gex, Need for Speed, Super Street Fighter II Turbo, Total Eclipse, Killing Time, Theme Park, and Return Fire): each boots, renders, and save-state roundtrips deterministically; six produce non-silent 44.1 kHz stereo audio, while Theme Park's captured menu is silent. The 2026-10-28 pass fixed live OpenAL crackle, added the PCM tap and FSR 1.0; compatibility status is unchanged. Road Rash's 3D race renders correctly and the attract race continues past 60,000 frames with no allocator error; Wolfenstein 3D starts a fresh game (FLOOR 1-1), walks/turns/fires, and its save states roundtrip; Alone in the Dark's in-game character renders upright at the staircase, and its Begin/Resume menu and intro are reachable; Lemmings reaches its title menu and passcode screen, its 640x480 high-resolution title matches the reference, and its Fun 4 passcode now starts gameplay with a save/load roundtrip after being typed entirely with the D-pad; Space Ace plays its animated demo and starts its interactive scene (X menu, A play) with a save/load roundtrip. The fifteenth pass also added the --gl-backend selector and validated Windows ANGLE/GLES 3.0 with pixel-identical CPU/GPU output. The tenth pass added CRT bloom and committed the D-pad passcode script. The eighth pass fixed the GPU palette shader so real frames use independent R/G/B CLUT indices and added the opt-in xBR-lv2 filter. The 2026-10-29 twenty-first pass added the optional Vulkan xBR/FSR GPU filters and HQx CPU path and removed the CPU palette-upload workaround from the OpenGL xBR path; compatibility status is unchanged.

Full-collection matrix (2026-10-30): the curated evidence below is being expanded into GameCompatibilityList.md, which contains the 241-title 4DO reference list, local filename matches from J:\Panasonic 3DO and G:\MAME\MAME Software List CHDs\3do, and empty 5DO status columns. The sweep procedure and status definitions are in CompatibilitySweep.md. The high-resolution research plan is in HighResolutionCelResearch.md.

The ninth pass captured per-title save-state, PNG, and WAV evidence (see TestCorpus.md 1.5a) and fixed the old Road Rash ~42,000/~92,000 stops. They were a static game-side wait (only READ_ID polls, all input ignored), not wall/rider pins. A CD media-access pulse now advances those screens to BUSTED, race select, and a new race; the autopilot then cycles race -> wreck -> menu -> new race to 100k. The Wolfenstein fresh-game walk, Lemmings passcode, Space Ace interactive scene, and Alone movement loop are verified. The open Road Rash work is finishing a full race without wrecking, not the old CD wait.

The CD wait fix is a media-access pulse: two consecutive READ_ID commands with no intervening command and at least 50 frames between them set poll bit 0x40. Portfolio uses that event to finish the post-race/post-load transition. The fix advances Road Rash's ~42k/~92k/100k screens to BUSTED, race select, and a new race. The five 3,000-frame corpus hashes were unchanged by that fix at the time; the exact DSP audio clock later re-baselined them (current keys in TestCorpus.md 1.1).

The 2026-10-24 sessions found four root causes:

• DSP immediate operands. The loader shifted un-justified immediates left by two instead of zero. The AudioFolio tail program tests Sema4Status & 8; the extra shift turned the mask into 0x20, so the DSP always skipped the ARM command and SAudio AllocInstrument timed out. The fix uses IMMEDIATE << (JUSTIFY & 3). The DSP now runs instrument code and acknowledges the ARM.
• CCB field order. The CCB parser read PRE0/PRE1 before the position, size, perspective, and pixel-processor fields, but FreeDO reads the geometry first and the preamble last. Every 3D CEL was decoded from the wrong offsets. Road Rash's race produced a bogus 1987x583 literal CEL whose arbitrary quads took minutes to rasterize; with the correct order the race renders at normal speed. The fix also corrects the CCB fields for all 3D titles.
• NVRAM initialization. The CLI never loaded an NVRAM image, so games that check the NVRAM volume for save space saw a zero-filled device. Alone in the Dark interpreted that as zero free slots, showed its "not enough space to save" warning, and shut down. The CLI now loads a fresh 32 KB volume by default and accepts --nvram <path> to persist it. Alone reaches its in-game room/inventory scene.
• Arbitrary-CEL X/Y transposition. The new rasterizer helpers take (x, y), but the arbitrary-quad fill passed (i, j) (row, column), following FreeDO's readPIX(src, i, j) parameter order. Every asymmetric quad was written transposed. Road Rash's race CELs wrote texels into the top of VRAM where the OS keeps its VRAM free list, producing AllocMem error, bad node address and a frozen race; Alone in the Dark drew its character horizontally. Passing (j, i) fixes both. The same pass fixed the VDLP bitmap copy to use the modulo-derived line width instead of a fixed 320 pixels, so 384/512/640-wide modes are no longer cropped, and made save-state loads restore the VDL pixel buffers immediately.

The 2026-10-25 eighth pass fixed a renderer-only issue: the GPU palette shader used the low five bits of each pixel as one CLUT index for all three channels, so real content appeared grayscale even though the CPU oracle was correct. The shader now uses the red, green, and blue indices independently. Lemmings gameplay was captured through sharp-bilinear and xBR with correct colors. The 2026-10-29 pass removed the former CPU-expanded xBR source; xBR now samples the GPU palette texture directly and the 320x240 WGL/ANGLE captures remain top-down and visually correct.

Space Ace's former frame-681 runaway was a downstream consequence of the DSP failure: the ReadySoft player's audio task copied a decoded buffer to a null [audioState+0x9C] pointer, wiping the ARM exception vectors. With the handshake working the copy no longer starts at address 0, and the title runs past 10,000 frames.

See Status.md and ImplementationPlan.md.

1. BIOS ROMs¶

The 3DO boot ROM is 1 MB. 4DO also accepts an optional second 1 MB ROM (BIOS2); if absent, the second bank is zero-filled.

The BIOS manager scans a folder, reports file name, size, MD5, region, video standard, Kanji, Anvil, and patch state, and stores the selected primary/secondary paths. Auto-selection uses the game database region, known title requirements, and per-game MD5 overrides; the configured BIOS remains the fallback. A missing/invalid BIOS prevents starting a session with a clear error. See UserGuide.md and Configuration.md.

2. Disc images¶

Sector count is read from the 3DO volume header at sector 0 (big-endian blockCount at byte offset 80). If the header is invalid, fall back to fileSize / sectorSize for ISO/BIN or the CHD data track's frame count.

CHD specifics:

• CHDSharp's ReadSector(lba) returns the first 2352 bytes of the 2448-byte CD frame. ChdDiscImage extracts 2048 user-data bytes from each track using its native layout:
• cooked Mode1/2048 -> offset 0
• raw Mode1/2352 -> offset 16
• raw Mode2/Form1 -> offset 24
• audio/2352 -> raw sector bytes (byte-swapped from CHD's Red Book order)
• Standalone and parent/child CHDs are supported; an optional parent path can be configured.
• Flat CHDs with 2048-byte units are also supported and exposed as one data track. This covers cooked ISOs converted with chdman createdvd (or createhd with 2048-byte units), which have no CD track table. CD audio is unavailable for these data-only images.
• DVD/HDD/raw CHDs with other unit sizes and audio-only CHDs are rejected; 3DO software is CD-ROM based.
• Read-ahead is enabled by default (4 hunks) for sequential streaming.

Game ID is the MD5 of the first two 2048-byte sectors (4096 bytes), matching 4DO. This ID keys save states, screenshots, and optional per-game settings. It must never change, or existing save data will be orphaned.

3. NVRAM¶

• 32 KB, persisted as NVRAM_SaveData.ram (v1 global file) in the saves directory.
• Writes are debounced (250 ms, ported from 4DO) and written atomically (temp file + replace).
• The NVRAM address mapping has a FreeDO quirk: one byte per 32-bit word address. Do not "fix" it.
• A fresh NVRAM image is initialized with the 3DO volume header (record type 1, ZZZZZ sync, label nvram, 32768 blocks, root directory copy at 0x84), matching 4DO.
• Per-game NVRAM is a possible later enhancement; v1 uses the 4DO-compatible global file.

4. Controllers¶

The 3DO pad has: A, B, C, P, X, L, R, plus D-pad directions. P is the pause button. There is no separate Start/Select in the standard pad protocol used by 4DO.

PBUS encoding is documented in docs/Architecture.md. The legacy packet is ported from 4DO's JohnnyInputPlugin; the device records follow libopera/opera_pbus.c.

Portfolio OS fixed-point math SWIs (0x50000-0x50012) can be handled on the host with --swi-hle or Settings > Emulation > Portfolio OS SWI HLE. This is an opt-in performance mode: the host routines match opera_fixedpoint_math.c, but HLE skips the guest handler's cycles and register clobbers. Titles that poll cycle-sensitive hardware around matrix calls can therefore change behavior; AutoBahn Tokio and Ballz boot and run with HLE but produce different frame hashes. Leave it off for compatibility, or use a per-game profile when a specific title benefits. See docs/Performance.md §8.2.

5. Game-specific fixes¶

FreeDOCore carries a fixmode bitmask. 4DO maps known game IDs to flags. The C# port keeps the same names and behavior.

The 4DO game database also maps:

• F3AF1B13, 217344B0 -> FIX_BIT_TIMING_1
• DBB419FA, 07C32F10, 5282889F, 7340307E -> FIX_BIT_TIMING_2
• CB8EE795 -> FIX_BIT_TIMING_3
• 1A370EBA, B35C911D -> FIX_BIT_TIMING_5
• 0511D3D2, 9F7D72BC, E843C635 -> FIX_BIT_TIMING_6
• 2AABA5B9, BF61BB32 -> FIX_BIT_GRAPHICS_STEP_Y

The game database is ported with attribution. Unknown game IDs get the first 8 hex characters of the MD5, with collision avoidance, matching 4DO. GameCompatibility applies the mapping automatically when the CLI or desktop shell opens a disc; --fix-mode/--no-game-fixes and the per-game override in Settings provide explicit control. The 3,000-frame Alone boot hash is unchanged with FIX_BIT_TIMING_6 applied.

The database grew from 518 to 653 records in SYS-17. Record-level fixMode, bios, bios2, and inputDevices fields now feed GameCompatibility, BiosLibrary, and GameControlDefaults; revision, discNumber/discTotal, nvram, and cheats are carried for profile seeding and diagnostics. On the local 1,366-image collection, gamedb stats --folder "J:\Panasonic 3DO" reports 1,366 readable images and 100% metadata coverage of those images. The three exceptions are audio-only [CD-Audio] bonus discs with no data track, not game images. See GameDatabase.md for the schema and the gamedb contribution workflow.

6. Video¶

6.1 High-resolution mode¶

EmulatorTuning.HighResolution / --high-res is an opt-in 640x480 renderer. CEL pixels address the 640-wide X coordinate and the 240-line Y coordinate; the even/odd X column is stored in plane 0/1 (and mirrored to planes 2/3 for the second interlace field). The VDLP interleaves planes 0/1 and 2/3 into two 640-wide scanlines and the frame source becomes 640x480. The extra planes follow the active VRAM capacity (4 * VramSize) and are not CPU-visible. SPORT fill/copy operations mirror the source page into the extra planes. Save states, frame hashes, screenshots, the CPU assembler, and the OpenGL uploader all carry the larger frame.

Important limitation: this is FreeDO-style line doubling, not true internal 640x480 rendering. Each 240-line source line is written to both interlace fields, so the vertical detail is still 240 lines. The mode improves horizontal CEL sampling and matches 640-wide titles such as Lemmings, but it does not add true vertical geometry. See HighResolutionCelResearch.md for the M14 research options (true 640x480, 2x supersampling, and GPU CEL compositing). M14 is post-1.0 and does not block v1.0.

Evidence: from a fresh Lemmings boot, the 640x480 title screen shows all four buttons and the full logo and matches the reference screenshot; the five 3,000-frame low-res corpus hashes were unchanged by high-res mode (the current keys are in TestCorpus.md 1.1). The mode is off by default because 320x240 titles must not be upscaled. The Lemmings passcode screen is a 320-wide UI scaled 2x; its five visible keyboard columns are the game's own window, not a framebuffer crop, so it does not need a separate high-resolution input path.

7. Audio¶

8. Save states¶

• Native C# format, versioned, explicit little-endian, CRC-protected.
• Includes CPU, the full 22 MB RAM backing (largest DRAM/VRAM layout plus four VRAM planes), ROM, NVRAM, MADAM, CLIO, DSP, SPORT, VDLP, XBUS, CD, Quartz, compatibility flags, and the parsed VDL frame. Format version 2 compresses large sections with Brotli; a typical state is 0.6-2.0 MB on disk (the uncompressed payload is ~28 MB). Version-1 states still load.
• Does not include disc contents or input state.
• Includes the parsed VDL frame lines, so a zero-frame load is pixel-identical to the saved frame even in high-resolution mode.
• Stores BIOS and disc hashes. A BIOS or emulator-version mismatch produces a warning but the load continues; a differing expected disc hash adds a warning but still loads.
• 4DO .4dosav import/export is not a v1 goal and is not implemented. The native format is explicit and versioned; a best-effort importer can be added later if there is demand. See SaveStates.md.

9. Known limitations / out of scope for v1¶

• No netplay.
• Rewind, fast-forward, cheats, TAS movies, OSD/quick menu, multi-track/CD-DA sectors, CD-DA volume/mute, and multi-disc playlists are implemented; CD-DA read-ahead/subcode, guest ODE prompts, and known-title disc audio evidence remain.
• No PAL.
• High-resolution CEL rendering is opt-in (--high-res) and validated for the Lemmings title; it is FreeDO-style line doubling, not true internal 640x480. Output upscaling via OpenGL is supported separately. See HighResolutionCelResearch.md.
• No GPU CEL compositing in v1; CEL rendering is software, matching FreeDO. The OpenGL renderer accelerates presentation, palette expansion, scaling, and filters.
• Mouse, light gun, arcade light gun, flightstick, and Orbatak trackball devices are implemented on the eight PBUS ports with an active-device cap of 0-8 and a 256-byte chain buffer; per-device calibration and scripted gameplay validation remain.
• No 3DO modem/network emulation.
• No cycle-exact ARM bus timing yet. The interpreter preserves FreeDOCore's Ncycle = 4; Opera uses Ncycle = 2 with datasheet references. The difference is under evaluation and must not be changed by default until a differential test proves the impact.
• No guarantee of 4DO save-state compatibility.
• The primary renderer requires desktop OpenGL 3.3 or OpenGL ES 3.0/3.1 (see docs/Performance.md); systems without either use the CPU fallback.
• CHD is supported for CD/GD-ROM images and for flat 2048-byte-unit CHDs made from cooked ISOs; DVD/HDD/raw CHDs with other unit sizes are not 3DO discs.
• CHD decompression is pure managed code and may be slower than ISO on low-end arm64 devices; read-ahead/cache mitigates this.

10. Curated compatibility matrix¶

This is the validated showcase set. The full 241-title 4DO reference list and the local collection matrix are in GameCompatibilityList.md. Use this table when testing real discs. Record the disc hash/ID, BIOS, build, and result.

Pass means a real game frame was rendered and verified visually. The Opera reference column is the bundled Windows OperaCore with the same BIOS and an ISO extracted from the same CHD; it verifies game progress, not audio output. The DSP/CCB/NVRAM and CEL transposition fixes are described above; input results come from --input-script runs.

10.1 Input validation (2026-10-24)¶

• Alone in the Dark: from the 16,000-frame in-game state, Up/Down and Left/Right/A/B/C produce different hashes and PCs; the D-pad/A sequence reached the Begin/Resume main menu. The control run walks the character through the room, so the in-game scene is live without input. From the intro, A every 300 frames leaves the static wait and reaches the room/mansion/title scenes (29,200 frames, PC 0x00020640, hash 0x62304BCF). From alone-intro-A2.state (39,200), Up, Left, A, Down, Right moves the character in the room; 39,800 is PC 0x0008D0EC, hash 0x44C7576B, and the state roundtrips (0x44C7576B -> two 120-frame continuations at PC 0x0008BF00, 0xE490C8C7).
• Lemmings: the title screen shows Play/New Game, but A/Left/Right/P/X do not leave the attract loop. Down + A opens the passcode screen ("Enter the access code"). The keyboard is a 9-column grid of which five columns are visible; A types the selected letter, B confirms early, and ten letters auto-confirm. A 15-frame D-pad hold moves the selection one key. A wrong code shows "Incorrect" and returns to the title. The 10-letter entry is permuted across RAM 0x96574-0x9657B + 0x9657E-0x9657F (display order 77,76,75,74,7B,7A,79,78, 7F,7E; 0x96576 is the preview slot refreshed from the selection). From the N-selected state, poking the other nine bytes of HNLHCIOECW (Fun 4) and pressing B is accepted (Code for 1 / Rating = Fun), then A shows the briefing and A starts gameplay (21,640 frames, PC 0x83FEC, hash 0x18F38B4D; zero-frame load roundtrips, two 120-frame continuations both 0x8D628FBD). A sequential poke at 0x96574 alone renders scrambled and reports Incorrect.
• Road Rash: the attract race renders correctly from ~17,000 frames. The 3DO controls are B accelerate, A brake, D-pad left/right steer, C attack, and P pause. Holding B starts a player race; a hard left around frame 31,000 clears the first oncoming taxi. --autopilot road v9 steered to the road centre measured low (rows 190-214) with lookahead obstacle steering, closing-side crossover, sharp-bend braking, wall avoidance, emergency braking, urban throttle pacing, a coast floor, and stuck recovery. The previously reported terminal wall pin at ~34,000 frames is the race-over BUSTED screen: from rr-auto-v8.state, no-input plus a B proves the plainclothes-cop screen, and no input resumes a lost race. v10 adds a close-traffic coast/steer rule (the 33,820 racer filled the centre while every old rule required a 0.6-wide escape), urgent preemption only for close traffic (applying every obstacle bias immediately oversteered into a taxi at ~32,000), an emergency-brake steer hold (the old path chased a lineX jump into the 33,890 tree), and a stopped/wreck restart that holds B with straight bars for 180 frames based on a left-scenery + speed-readout signature. The restart recovers the rider-running phase (34,500/80,000 -> 52 mph) and wall stops, but a cyclic right/left search was tried and drove the 83,000 run into BUSTED (0x9F44FAA2), so throttle-only is the proven policy. Best evidence: 91,000 live frames (PC 0x00085F48, hash 0x02ABE78E, cov=0.93), zero-frame 0x02ABE78E, two 120-frame continuations both PC 0x000973EC / 0x3F27B3D6. The ninth pass fixed the old ~92,000/~42,000 stop: the CD device raises a media-access event (0x40) after two consecutive READ_ID polls with an idle gap. The game advances to BUSTED, the race-select menu, and a new race (PC 0x0007DCBC, hash 0x277CCED7 at 113,000). The five 3,000-frame keys were later re-baselined by the exact DSP clock; see TestCorpus.md 1.1. The autopilot pulses A/releases throttle during long-static, low-coverage, CD-load, and 180-frame restart-stuck states (autopilotMenuAdvance). From the 60k state it starts a new race at 80,000 (PC 0x00092FF8, hash 0x1074EF3F) and cycles race/wreck/menu to 100,000 (PC 0x0008214C, hash 0x0D46CF56). The remaining work is finishing a full race without wrecking, not the old CD wait or another side search.
• Wolfenstein 3D: from the 12,000-frame demo state, A opens the Select Previews menu; D-pad/A navigate to Select a Level and start a fresh game at FLOOR 1-1 with SCORE 0000000. From a live 60-frame state, two independent 120-frame continuations both end at hash 0x370158AC, and a zero-frame load reproduces the saved hash 0xC9989182. The long fresh-game load (w3d-loadlong.state) stays static with no input, but A every 300 frames reaches character select and FLOOR 1-1 at 64,000 (PC 0x00069F3C, hash 0x2F5B9864).
• Space Ace: from a 12,000-frame state, X at 12,000 returns to the title/menu (PC 0x0000E12C, hash 0x596B6A81); A at 12,200 starts the interactive scene. The run reaches 15,000 frames (PC 0x000765F0, hash 0x7E780D93), the state roundtrips, and two independent 120-frame continuations both end at PC 0x000284EC, hash 0x70B16422. P also changes state; A/B/C and the D-pad do not change the demo by themselves.

10.2 Renderer fixes and validation (2026-10-24)¶

• Arbitrary CEL transposition. TexelDrawArbitrary passed (i, j) to the (x, y) read/write helpers. FreeDO's readPIX(src, i, j) treats i as the row, so the port's renamed helpers inverted every texel. Asymmetric quads were written transposed. Road Rash's race CELs landed in the VRAM allocator's free list at 0x2FC000, producing AllocMem error, bad node address and a frozen frame; Alone in the Dark drew its character horizontally. The fix passes (j, i).
• VDLP line width. The bitmap copy always read 320 pixels, so 384/512/640-wide modes were cropped. It now uses the modulo parsed from the VDL control word, and the machine derives SourceWidth from the widest rendered line. The 320-wide corpus output is unchanged.
• Save-state pixels. Loading a state now restores the VDL pixel buffers instead of discarding them. A zero-frame load reproduces the saved frame hash exactly.
• Audio measurements. 600-frame WAV captures: Road Rash 44.1 kHz stereo, 10.03 s, peak 32768, RMS 9051/9146, 191 clipped samples; Wolfenstein 44.1 kHz stereo, 10.02 s, peak 30743, RMS 5620/5457, no clipping. Both run at ~737 samples/frame.

The 2026-10-23 CD fix is significant: CdromDevice.InitCd used FreeDO's double-add of the 150-frame lead-in, so the disc end reported to the CD DIPIR was 150 frames larger than the real disc. The DIPIR rejects a disc when the reported end exceeds the volume block count by more than 512 frames. Lemmings had enough padding to pass; Road Rash, Space Ace, and Wolfenstein did not, and were misidentified as non-3DO discs. Reporting the real end fixed all three.

Space Ace's former runaway is now explained. At frame ~681 the ReadySoft player's audio task copies a decoded audio chunk to the pointer at audioState+0x9C. That pointer was zero because the player's mixer/instrument setup failed (AllocInstrument returned d52bf113). The 27,538-byte memcpy started at address 0, wiped the ARM exception vectors, and a timer FIQ then vectored through the zeroed table. After the DSP handshake fix the pointer is valid, the copy no longer starts at 0, and Space Ace runs past 10,000 frames.

The DSP failure was an operand-decoding bug, not a resource-map or microcode issue. The AudioFolio tail program uses Sema4Status & 8 to detect an ARM command. The port's immediate loader shifted by two when the one-bit JUSTIFY field was zero, so the constant 8 became 0x20 and the AND always produced zero. The DSP therefore took the "no command" branch, skipped Sema4Data, and went back to sleep without writing the ACK. FreeDO and Opera both use IMMEDIATE << (JUSTIFY & 3). A second multiplier-rounding difference was fixed in the same pass ((MULT1 * (Y >> 15)) & ~1, not MULT1 * ((Y >> 15) & ~1)).

The second root cause was in the CEL control-block parser. FreeDO reads the CCB body in this order: next pointer, source pointer, PLUT pointer, position, size, perspective, pixel-processor control, then the DSPP preamble. The port read the preamble immediately after the pointers and then read the geometry from the following words. Every 3D CEL was therefore built from shifted fields. Road Rash's race scene decoded a 1987x583 literal CEL with a perspective delta of 560 pixels per row; its arbitrary quads were pathological and the run appeared to hang. Parsing the fields in the reference order fixed the race and also changed Space Ace's 3,000-frame hash. The rasterizer's cnt_cross > 2 x-point sort was also missing; it is now ported from the reference.

The third root cause was the CLI's NVRAM. A zero-filled NVRAM has no valid volume header, so Alone in the Dark's save-space check saw zero free slots and displayed its "There is only enough space on your 3DO system to save X games" warning. After a timeout it entered its shutdown routine and spun on VCNT. Loading a fresh 32 KB NVRAM volume (the same image the desktop frontend already uses) removes the warning and the game reaches its in-game scene. The CLI also gained --nvram to persist the image.

Opera plays the same titles past these points, and the C# port now does too. The next compatibility step is input-driven gameplay validation, not another DSP or CD rewrite.

10.2 Release soak (1.0.0-rc.1, 2026-10-26)¶

Fresh 36,000-frame runs (10 minutes of emulated time) with panafz1.bin, Release build, and no input; all exited 0 with the DSP running:

The seven additional titles' source discs are not present on this machine, so the release soak continued their archived 12,000-frame states to 36,000. All exited 0:

Save/load is deterministic at the 36,000-frame states: the zero-frame load reproduces the live hash, and two independent 120-frame continuations match. Showcase examples: Wolfenstein 0x0814F039 -> 0x2853ADC5 twice, Lemmings 0x9939D345 -> 0xB069CA91 twice, Road Rash 0xBDC2DBEA -> 0xA41DD4CD twice, Space Ace 0xFD0F5B72 -> 0x2A9967FA twice, and Alone 0xBA7F2C8C -> 0x666EEDB7 twice. The 600-frame WAV captures are 44.1 kHz stereo and non-silent (Road Rash clips 74 samples in 10 s; Wolfenstein and Lemmings do not clip).

Scripted gameplay evidence: Wolfenstein's fresh-boot path reaches FLOOR 1-1 (0xF568EBCA at 36,000 after the documented input scripts), Lemmings' archived 23,000-frame gameplay state runs to 36,000 (0xD3961D3D), and Road Rash's 60,000-frame race state runs to 80,000 with --autopilot road. Space Ace's 15,000-frame interactive state renders the scene; a long no-input continuation reaches a black wait screen, which is documented in KnownIssues.md.

10.3 Linux x64 validation (2026-11-02)¶

Validated on the local Hyper-V LinuxMint VM: Linux Mint 22.3, x86_64, kernel 7.0.0-34-generic, Mesa 25.2.8 llvmpipe, OpenGL 4.5 core / GLES 3.2, 1920x1080. The guest has no .NET SDK; the self-contained linux-x64 publish was copied into the VM and run from /home/vm/linux-smoke.

The 3,000-frame Wolfenstein capture shows the Interplay logo (non-black), confirming disc boot and rendering on Linux x64. The full test suite also passes. Linux x64 is shipped as a preview/best-effort build: the CLI, frontend launch, and audio path are validated, while GPU presentation should be re-tested on a hardware-accelerated Linux desktop before claiming full GL support. See KnownIssues.md.

11. Reference comparison (2026-10-23, updated 2026-10-24)¶

The data track of each CHD was extracted to ISO with 5do disc dump, then run with the bundled Windows OperaCore (References/CPP_SimpleGui) and the same panafz1.bin BIOS. Opera rendered Road Rash, Wolfenstein, Alone, and Space Ace for 60 s. After the 2026-10-24 DSP fixes the C# port also renders all five titles and produces non-silent audio. The comparison method and commands are in ReferenceStrategy.md; the master plan and next actions are in ImplementationPlan.md.

The 2026-10-24 pass added the diagnostics needed to make the comparison mechanical: --watch-lowmem, --dump-ram, --dump-dsp, --dsp-trace, the disasm command, frame-stamped CLIO traces, DSP state in the run JSON, and --input-script for repeatable controller taps. The DSP trace and memory dumps identified the exact Sema4Status & 8 test that was being mis-decoded.

Local corpus (user-provided, not distributed; see TestCorpus.md):

• BIOS-only boot and menu navigation: panafz1.bin.
• Alone in the Dark (USA).chd — exercises FIX_BIT_TIMING_6 and literal CELs.
• Lemmings (USA).chd — 2D-heavy CEL/packed-sprite workload.
• Road Rash (USA).chd — 3D/matrix and streaming workload.
• Space Ace (USA).chd — FMV/streaming workload.
• Wolfenstein 3D (USA).chd — first-person 3D/matrix workload.
• Alternate BIOS variants (goldstar.bin, panafz10.bin, Anvil) for region/fix paths.

Suggested additional coverage:

• A CD-streaming title with FMV/audio.
• A title with a known 4DO fix flag not represented above.
• A homebrew/test disc if available.

12. Reporting a compatibility issue¶

Include:

• Emulator version/commit.
• BIOS MD5 and region.
• Disc ID (first 8 hex of the 2-sector MD5) and format (ISO/BIN/CUE).
• Fix mode flags and settings.
• Steps to reproduce.
• Save state and a short input recording if possible.
• Log output with the Emulation and Cdrom categories enabled.

13. Full-collection sweep (2026-10-30)¶

The local collection is:

• J:\Panasonic 3DO — 589 CHD images.
• G:\MAME\MAME Software List CHDs\3do — 5 CHD images.

The sweep is automated by tools/compat-sweep.ps1; the reference table is GameCompatibilityList.md, and the procedure/status definitions are in CompatibilitySweep.md.

Automated labels:

• Not Loadable — CLI error, timeout, invalid JSON, or a short frame count.
• Loadable/No video — completed the frame count but the final frame is black/blank.
• Loadable/Rendered — completed and the final frame is non-black with several colors.

Final labels (after visual/input review): No data, Not Loadable, Unplayable, Playable, Good.

This sweep is the current compatibility gate. It is deliberately broader than the 12-title matrix above and does not require full gameplay completion. The 4DO list is used to prioritize known problem titles, not to assume that a filename match is the same disc revision.

13.1 First sweep results¶

The first 594-disc pass completed on 2026-10-30 and found 576 rendered, 13 black/blank, and 5 not loadable. The 2026-11-01 extended-boot retry reran all 18 non-rendered rows; the table below reflects the retried counts:

The per-disc table is GameCompatibilityResults.md.

13.2 Triage results (2026-10-30, extended 2026-11-01)¶

The baseline used panafz1.bin. Reruns with the other local BIOS images show that several failures are BIOS-selection problems rather than emulator defects:

The Arcade CD-ROM rows need the patched BIOS; unpatched panafz10 reaches the 3DO storage-manager shell instead of the game. The default pad alone does not fix them, and the patched BIOS alone is not enough either: the headless PBUS fix is what lets their boot code proceed past the controller probe.

Slayer has the opposite regional split: the Europe disc renders its credits with panafz1 by 12,000 frames, while the USA disc stays black with panafz1 and needs panafz10 to reach the title menu at 60,000 frames.

Battle Chess (Europe) is the first confirmed disc that rejects the NTSC system software outright. It prints "This is not a PAL TV system!" and halts at PC 0x00088EDC with panafz1, panafz10, and the GoldStar/Sanyo BIOSes. The European Anvil BIOS (panafz10e-anvil.bin) satisfies the region check: the title screen appears at 6,000 frames and the board at 30,000 frames. 5DO still emits NTSC 320x240 timing, so this is a region/OS check, not proof of PAL 50 Hz emulation. The USA and Japan discs boot with panafz10.

The 2026-11-01 retry reran all 18 non-rendered rows. At 20,000 frames with panafz10, Ishida Yoshio Kudan no Igo Seiha, both Konpeki no Kantai discs, all three Myst discs, Olympic Summer Games, all three Slayer discs, the four Arcade CD-ROM titles, and Mirai Shounen Conan render. Powers Kingdom (Europe) renders its live-action intro at 65,000 frames with panafz10e-anvil. Only Crash 'n Burn (Japan) and WorldCup Super Stadium (Japan) remain on the Anvil boot screen at 65,000 frames. The final counts are 593 rendered and 1 known failure (Crash 'n Burn Japan); WorldCup Super Stadium renders with the FZ-10J BIOS and its Kanji font ROM, and no title is Not Loadable.

Consequences for the sweep:

• The four Arcade CD-ROM rows are not timeouts. They are loadable with the patched BIOS plus the default headless pad, and should be recorded with that BIOS in their notes.
• The Loadable/No video set is not a stable failure class. Some titles need 20,000 frames, and some need a different BIOS. The retry used -RetryFrames 20000 in tools/compat-sweep.ps1; do not convert a row to Unplayable without a BIOS-aware extended-boot rerun.
• The final 5DO status must record the BIOS used, because a result is only meaningful for the BIOS that was tested.

The SYS-04 BIOS selector encodes the actionable part of this triage: arcade discs get the patched FZ-10 image, European games prefer the FZ-10E Anvil image, Japanese games prefer the FZ-10JA/Kanji image, and Slayer Europe/USA plus Powers Kingdom Europe have explicit known requirements. bios select prints the reason for each choice, and run --json records it in biosSelection. The full BIOS-aware sweep is still the acceptance step; the selector does not claim a title is compatible merely because a BIOS was chosen.