World Circuit DOS demo: keyboard consumption and driving boundary
Run: scratch/runs/20261107-world-circuit-driving/; owner Telegram177.
Tested source: 90e01588dec6ee9eece5b0744dbe5b6a69fe466d, after the
startup repair. No production code changed.
Ordinary driving remains unverified. The original renders the start sequence and later advances the scene without held input. The observed original loop drains BIOS keystrokes; it does not consume hardware held-key state. This is evidence for a demo limitation in this execution path, with no demonstrated CPU/DOS/keyboard mismatch to repair. Automatic movement is not control proof.
Original input path
Static inspection used retained startup memory before allocating a fresh box,
then checked the fresh session's memory. At 1a66:323b, the original calls
INT 16h/AH=01, tests availability, calls AH=00 when a key exists, and returns
carry clear after consumption or carry set when empty. 1a66:324a repeatedly
calls that routine while carry is clear, then returns far. It drains the
keyboard queue. The main segment calls this drain at 0110:e7f5 and
0110:e951; the following instructions do not dispatch on the key value.
The run retains disassemblies segment-1a66.asm, main-f8b0.asm and
main-fa10.asm, plus the underlying captured memory.
The runner's existing Machine.keyDown queues one BIOS key and a hardware make;
keyUp independently queues the break and removes the held entry. BIOS reads
consume machine.keys and update its BDA mirror. IRQ1 delivery requires an
installed INT 9 vector; the original retained f000:0109, the default vector.
The fresh run observed 269 INT 16h availability polls and four reads,
zero IRQ1 deliveries, and zero reads of ports 60h, 61h, 64h or 201h.
All four supplied BIOS keys were consumed, while their eight hardware
make/break bytes remained queued. Thus the ordinary keys reached the guest;
there is no evidence that a missing runner release prevented driving.
Documented bounded inputs and visual result
The MicroProse DOS technical booklet lists A for acceleration, Z for braking, angle-bracket keys for steering, and A+Space for an upshift. The manual describes K for keyboard/joystick mode. Their OCR was retrieved through Archive metadata/download and retained. These 1993 retail references are candidates, not proof that this 1992 demo supports interactive driving.
After reviewing the grid, the single session used:
| Input | Bounded work while held | Release |
|---|---|---|
| A | 15 million dispatches | Separate UP, then 2 million |
| K, documented native mode request | 5 million | Separate UP, then 1 million |
| A, then Space | A alone 1 million, chord 5 million | Space UP, A held another 20 million; separate A UP |
K was consumed, but no visible mode selection was established. A and the chord left the reviewed cockpit in neutral at 0 MPH, with no demonstrated gear or acceleration response. No Escape/function-key survey, edited guest settings, bindings, memory, clocks, forced control state or guest patch was used. The drain evidence ended input probing; steering/braking probes would not establish a held-control path here and were not sent.
The bounded passive continuation ended at 990,055,972 dispatches. Captures show the start lights progressing from dark to red, then the tower leaving view, cars advancing and changed road/track framing. The final display reads neutral, MPH 004, lap 1, position 24. This scene advance occurred after all releases, without further input. No steering consequence or user-controlled gear/speed/track displacement was demonstrated. No decode refusal, missing file, trap, guest exit, stuck-detector stop or UNKNOWN command occurred.
Fixture, validation and recovery
The same hash-pinned three-file native ToyVM fixture was used: original
test/binaries/dos-games/world-circuit-demo/F1GP/{F1GPDEMO.EXE,ADLIB.CAT,XSOUND.BIN}.
The registered EXE SHA-256 is
5c88d286945690fc7340ce98aaab0ca51912e7b0cc1ed1f89754fa0e82721511;
media-refs.json records all three paths, sizes and hashes. Static corpus checks
passed before allocation; 139 runtime/media/helper pins passed before and
after execution. The actual module SHA-256 was
b6bacd0b52e72aea90a1ddad6016d07b59f4102de3dba720c66c5d3d855c9550,
matching startup. Node v24.22.1/V8 13.6.233.17-node.58 ran the existing
createMachineRunner, DosSession.step and production PNG capture functions,
with default CPU/devices/clocks and empty guest arguments. Run-specific wrappers
only observed keyboard methods. Established runner/capture/lifecycle checks
were reused; driver syntax, ordered ACK/release validation, complete retrieval
hash/size validation and documentation whitespace checks passed.
Sixteen commands, including terminal quit, had exact ordered acknowledgments;
four physical DOWNs had four independent UPs. A local retrieval attempt first
encountered an already retained immutable file; the corrected retrieval verified
existing bytes and fetched missing files without replaying guest commands.
All 70 remote artifact paths, 61 unique contents, 2,159,921 bytes were
recovered and SHA-256/size verified before requesting box shutdown. Intermediate
memory captures are retained separately. Git SHA, relevant binary diffs, helper
patches, media references and the actual module provide recovery; no source-tree
or media archives were made. The corpus has no candidateId, so this investigation
uses run-summary.json rather than an unrelated dashboard candidate.
One persisted UUID allocated bx_3cvmp865 at 22:23:02Z, initial expiry
22:48:02.297Z, with more than 5 GB free on /home. Observable detached
25-minute watchdog and normal-stop delegates were armed before allocation.
The task explicitly delegated shutdown. After recovery, the sole driver PID
5727 was absent, held keys were empty, and this CLI route had opened no ports
or browser profiles. Normal stop stop_4c7857b59be1d06a completed
22:28:07.827Z, forced:false; final box state ARCHIVED, expiry null.
Both delegates exited zero; owned watchers were closed. Custody returns to the
coordinator; the box was not deleted or resumed.
Next: preserve the demo limitation. An interactive-driving claim needs an authentic demo entry/control path or separately authorized interactive media; these retail controls alone do not justify an emulator change. No public deployment, FPS, audio or parity claim is made.