Tachyon: The Fringe (demo)
NovaLogic, 2000 (copyright screen says 1999). Space combat, first five
single-player missions. Candidate tachyon-demo, app tachyon_demo, local
only.
Source and layout
- archive.org
TachyonTheFringe,TachyonDemo.zip(61,461,980 bytes, sha1 6a51db79d5d29c0e6f036b9dd29c2ddfd0a42e70). An InstallShield 5 package;tools/is5-extract.jsunpacksDATA1.CAB(14 files, 92 MB) toinstalled/tachyon(guestc:\tachyon). LICENSE.TXT(NovaLogic): personal use of one copy, no reproduction or transfer without written permission. SolocalOnly, never deployed.Front.exeis the setup front end (Play / Video Setup / Control Setup / Update);Front.cfgsays it runstachyon.exeand writestachyon.cfg. The registry launchesTachyon.exedirectly.- Data:
Space.pff(68 MB),Menubgs.sbf,Exile.sbf. Renderers:Hw_d3d.dll(Direct3D, loads at 0x6e8000 relocated from 0x10000000) andHw_3dfx.dll(Glide 2, wantsglide2x.dll).
Renderer configuration
tachyon.cfg [VIDEO], read by the game itself with _lopen/ReadFile (not
the profile API, so startupIni cannot change it):
USE_HARDWARE=0(shipped): software renderer, no hw DLL.USE_HARDWARE=1,HW_TYPE=0:hw_d3d.dll(IDirectDraw4 + IDirect3D3).USE_HARDWARE=1,HW_TYPE=1:hw_3dfx.dll->glide2x.dll.RES=1(shipped): under D3D the game sets 640x480x16, then 800x600x32 on a second DirectDraw interface, but its menu still draws into 640x480 buffers. Our input then scales the page's 640x480 canvas to 800x600 (a click at 320,226 arrives as WM_MOUSEMOVE 400,282) and every menu click misses (d3d-res1-800x600-click-miss-b1899.png). Not fixed; open question whether the canvas should follow the second mode.RES=0: stays 640x480 and clicks land.
The registry mounts Tachyon.cfg with iniSet USE_HARDWARE=1, HW_TYPE=0,
RES=0. That needed dbc778344: a registry file now replaces the generated
local manifest's file at the same guest path.
Fixes it needed
- dd75c48d3:
IDirectDrawSurface::GetDDInterfacereturned the v1 IDirectDraw. hw_d3d.dll creates its chain on IDirectDraw4, asks the back buffer for its DirectDraw (RVA 0x3a7e area) and calls slot 27 (GetDeviceIdentifier, vendor compared with 0x12d2); on the v1 table it hit something else and the game posted WM_CLOSE at batch 6.
Input
- Menus are WM_MOUSE* driven (DirectInput is created only to enumerate joysticks, then released).
- The pilot-name field reads WM_KEYDOWN together with WM_CHAR: CLI
keydownalone orkeypressalone types nothing;keydown+keypresskeyupper letter works (only one of three fast letters stuck; a name of one letter enables Play). A real keyboard sends all three.
- Flight keys (readme keychart): arrows pitch/yaw, Space fire, Tab afterburner, Backspace full throttle, 0 zero throttle, +/- throttle, comma/period roll, 1/2 weapons, ? in-game key chart, Esc exit.
Route (headless, --batch-size=100000, D3D registry config)
Guest 640x480 coordinates.
| batch | input | result |
|---|---|---|
| ~150 | - | main menu (Single Player / Multiplayer / Exit) |
| 200-210 | click 320,226 | pilot slots |
| 320-330 | click 320,63 | slot 1 selected, caret from ~500 |
| 470-494 | keydown+keypress+keyup A, C, E | name "C" |
| 520 | Enter | name accepted |
| 580-590 | click Play 235,365 | intro cutscene (~700) |
| ~2000 | - | starbase interior |
| 2100-2130 | click the "01" door 233,240 | Ship Bay (Pegasus; Configure / Launch / Lobby) |
| 2300-2310 | click Launch 320,440 | launch |
| ~3300 | - | cockpit, training objectives (throttle, roll, ?) |
| ~3700 | - | jump gate |
| ~4800 | - | "Destination: Venus - Training Sector, head to Luna" |
The right-hand console (470,300) opens the news terminal.
Evidence: scratch/runs/20261011T2025Z-tachyon-demo-claude1863.
Local flight controls reviewed 2026-11-01
One local original CLI invocation reached the training cockpit and demonstrated
ordinary throttle, direction and fire input. Evidence is in the ignored directory
scratch/runs/20261101-tachyon-flight-gated/; review.json identifies the
visually inspected pairs and commands.ndjson / responses.json preserve the
actual input sequence. This qualifies these observed controls only, not mission
completion, audio, frame rate or the rest of the demo.
The prior offline review was reused after verifying all 15 original/manifest hashes (92,143,006 bytes), the note and registry. It found no later accepted paired controls evidence; the Oct11 capture directory is locally absent. The existing candidate licence record authorizes personal local use; the owner task authorized this invocation. No new EULA was accepted. The original licence text and installer archive are absent locally, so licence authority remains the recorded candidate provenance, not a newly inspected licence document.
The run waited for Redline's fresh final/result and 06:31:31 UTC explicit
runtime/source/build release. Gate-open source baseline was main e06fe9784.
Full comparison against the accepted 149286afd source found the subsequent
capture-chooser repair and its regression test, so a new locked build was run
and passed. The actual loaded module was
330b120d64cfd4aebdd945c99a95022dbd97db7ee6962d7ec534c24da466011b;
source, module, mirror and original identities were checked again after exit.
Before invocation, synthetic sync/async checks exercised actual emitted
dx_trace callbacks, verified the whole shared 512 MB memory and guest clock
remained unchanged by observation, and checked bounded retention after 10,001
late events. No source, test or renderer changes were needed for Tachyon.
The registered tachyon_demo route and its established iniSet mapping above
were used without new overrides. Shipped configuration and originals remained
byte-identical. Each mouse/key press was released and allowed to dispatch.
Same-batch click only highlighted Single Player; holding the mouse down across
batches opened it. Pilot A used keydown + WM_CHAR + keyup, then Enter and Play.
Reviewed progress was main menu (150), pilot list (540), accepted name (680),
starbase interior (2230), Ship Bay/Pegasus (2410), training briefing (3220), and
cockpit (3370). These are observed batch numbers with batch size 100,000, not
timing guarantees.
| Control | Reviewed before / after PNG | Observed response |
|---|---|---|
Zero throttle, 0 down at 3450, up at 3453 |
baseline-3450.png / throttle-zero-3488.png |
HUD thrust 300 → 0; speed 410 → 2. |
| Full throttle, Backspace down at 3488, up at 3491 | throttle-zero-3488.png / throttle-full-3526.png |
HUD thrust 0 → 300; speed 2 → 288. |
| Right arrow down at 3526, up at 3541 | throttle-full-3526.png / right-held-3541.png |
Star/asteroid view changes and cockpit banks; release captured at 3546. The two held/release captures are identical, so they do not independently establish when turning ceased. |
| Space down at 3546, up at 3561 | right-released-3546.png / fire-held-3561.png / fire-released-3581.png |
Blue laser bolts appear around the reticle while held and are absent after release. |
The last controlled capture's objective text reads “Destination: Venus - Training
Sector”, “Jumps: 2”, “Head to Luna”. The separate exit PNG omits that text;
the paired observations above use the controlled captures. Input trace records
the ordinary down/up messages; the final
Space keyup was dispatched and the scheduled/pending/renderer queues were empty.
One rejected read-only input-state request included unsupported key arguments;
the corrected request succeeded. The run quit through the CLI at batch 3581,
outer exit 0 after 206.987 seconds, within the recorded 850/900-second inner/outer
limits. No second original invocation, remote box, browser or GPU was used.
Training onward attempt, 2026-11-01
scratch/runs/20261101-tachyon-training-onward/ records an attempt to continue
with ordinary Insert slide input and the next displayed objective toward Luna.
It did not invoke the original: the first launch blocker was free disk below
the task's required 2,300,000,000 bytes. At the first pre-launch recheck there
were 2,299,285,504 bytes available (714,496 bytes short), then 2,299,256,832
(743,168 short). The entire owned scratch directory was only about 269 KB;
removing it could not recover admission. Foreign artifacts were preserved.
No new captures, Insert consumption, slide response or Luna progress are claimed.
The worker waited scratch-only for Redline worker
01a195da-2b40-7ff0-ad0b-c710dc6a5349's fresh final/result and its 06:54:25 UTC
explicit runtime/source/build release. The gate opened at 06:55:12 UTC, with
2,300,731,392 bytes free. Gate-open main was 574a35dc415c9788208798d900e57b8af345d89d.
All 4,390 source inputs, 122 prior actual loaded inputs, the module and region
mirror matched the accepted Tachyon locked build, so no rebuild was needed.
The canonical module SHA-256 remained
330b120d64cfd4aebdd945c99a95022dbd97db7ee6962d7ec534c24da466011b;
it was pinned directly, avoiding a duplicate module on the nearly full disk.
Synthetic sync/async validation passed actual emitted dx_trace callbacks,
whole shared 512 MB memory and guest-clock read-only checks, and bounded
retention after 10,001 late events. These were observer checks, not game runs.
All 15 original/manifest files were pinned and unchanged, including the actual
installed/tachyon/Tachyon.exe; no compilation-container identity was used.
The existing recorded local-use licence authority and established registry
configuration were retained without a new EULA acceptance or config override.
The full note, prior driver, keychart and prior controlled captures were read.
The prepared scratch route deliberately stops before repeating the prior
throttle/turn/fire checks. The next prerequisite is sufficient launch disk
headroom; ordinary Insert input and the current displayed instruction remain
unverified by this attempt.
Native Insert and Luna approach, 2026-11-01
One original invocation in scratch/runs/20261101-tachyon-training-margin/
advanced the observed route beyond the earlier controls checkpoint. The reviewed
baseline at batch 3450 displayed “HOLD DOWN THE INSERT KEY TO SLIDE”. Ordinary
Insert (VK 45) was held at 3450–3480 and 3490–3590, with releases after each hold.
Actual DispatchMessageA traces record down/up at 3451/3482 and 3490/3591;
before/held/released input-state snapshots show empty scheduled, pending,
cross-thread and renderer queues. The 3590 held and 3610 released captures display
“Destination: Venus - Training Sector”, “Jumps: 2”, “Head to Luna”. This proves
native input consumption and the displayed transition during the hold; it does
not isolate slide physics or prove that Insert caused the timed objective change.
The native ? keychart was reviewed, then targeting and arrow inputs followed
the Luna objective. Luna is visibly selected by batch 3738. A gate-opening
animation appears at 3761/3781, but the cockpit at 3981 still says “Jumps: 2 / Head
to Luna”, so that animation is not evidence of a completed jump. Later reviewed
approach captures show Luna range decreasing from about 27.9 at 4080 to 22.1 at
4257. The final controlled capture at 4277 shows range 22.51 and the same objective.
Luna arrival and mission completion remain unverified. review.json names the
reviewed pairs; raw commands, responses, dispatch traces and PNGs are retained.
The Launch cursor had remained at (320,440). Neutral intervals continued to change heading; moving the native cursor to (320,240) at 4147 produced more stable paired views. This is a navigation confound, not a proven emulator defect. Three intermediate input-state snapshots still had renderer messages pending; later steps consumed them. Every requested keydown has a matching actual dispatched keyup, and the final observed queues are empty. Physical Insert/arrow key-state fields and the guest post queue were not independently sampled.
The initial gate stayed scratch-only until Redline worker
01a195ed-24ef-7f61-a7ba-05ddac7b9401 had fresh final/result files and its
07:14:16 UTC explicit runtime/source/build release. Gate-open main was
0f9825de5abb2ce9c06f78380662da975be6b072. All 4,390 source inputs, 122 prior
loaded inputs, the module and region mirror matched the accepted locked build;
no rebuild was required. Actual loaded module SHA-256 was
330b120d64cfd4aebdd945c99a95022dbd97db7ee6962d7ec534c24da466011b.
Observer validation passed actual sync/async emitted callbacks, whole shared
512 MB memory and guest-clock read-only checks, and bounded late-10,001 retention
before runtime diagnostics. Source, actual loaded inputs and all 15 original/
manifest identities were rechecked after exit. Existing recorded licence
authority and registry configuration were retained; no new EULA or file/config
changes were made. The original exited 0 after 415.898 seconds at batch 4277,
within the 850/900-second inner/outer limits. No second invocation, runtime
repair, remote box, browser or GPU was used. No audio, FPS or broader demo claims.
Luna arrival attempt blocked before launch, 2026-11-02
scratch/runs/20261102-tachyon-luna-arrival-next/ preserves this attempt's
preparation and measured launch blocker. No original session or boat was
started, and no new arrival, completed jump or objective progress is claimed.
The accepted Insert transition and range 22.1/22.51 with “Jumps: 2 / Head to
Luna” remain the previous checkpoint; gate animation alone remains insufficient.
The worker stayed scratch-only until fresh GL query worker35014 and Nox item
worker76354 each had actual terminal exit 0, turn.completed, result/final,
exact AUTHOR RELEASE and scoped helper cleanup. Their releases were
13:31:13 and 14:03:02 UTC; named Nox helper PIDs and both local helper scopes
were independently checked absent. After both gates cleared, the first launch
check at 14:03:58 UTC measured 2,376,925,184 bytes free on /home, below
the task's requirement of strictly more than 2,400,000,000 bytes by
23,074,817 bytes. The owned scratch directory used only 962,560 allocated
bytes, so removing it could not restore the margin. The worker stopped at
this first concrete causal blocker, preserving foreign evidence and the
owner's 2 GB floor. admission.json and gate.json record these checks.
All 15 original/manifest identities matched the accepted pins (92,143,006 bytes). The shipped readme and both prior native keychart pages were reviewed: Enter selects a target, T cycles targets, arrows turn, Backspace sets full throttle, 0 sets zero throttle and Tab activates afterburners. No separate jump key was shown. Prepared, unexecuted controls center the native cursor at (320,240) after Launch and retain before/held/released/settled captures plus matched neutral intervals and input-state observations. They are preparation, not input-consumption or flight evidence.
A preliminary comparison matched all 7,228 inputs of the newly accepted GL
build, whose retrieved module SHA-256 is
d6e20ba99816b54b75d2423983ab7c80a0bf9bb1776b570a46d244f881b3291d.
It was not loaded or reused here; the older Nov01 module was not assumed
current. No build, runtime tests, guest/configuration changes, downloads or
subworkers ran. A future authorized attempt needs restored launch headroom,
fresh dispatch and module/input checks, then one finite-TTL boat recorded
before use. The gameplay requirement remains a visibly completed Luna
arrival/jump or changed objective under ordinary controls.
Recovered Luna approach, 2026-11-02
One original session in scratch/runs/20261102-tachyon-luna-arrival-recovered/
reached a new reviewed Luna range 4.58, but did not establish arrival,
a completed jump or a new objective. The last controlled capture at batch 6092
still reads “Destination: Venus - Training Sector”, “Jumps: 2”, “Head to Luna”.
The requested arrival goal remains incomplete. This attempt ended at the
controller's 1,260-second outer deadline; the original then exited 0 after quit,
at 1,261.187 seconds. No concrete game/runtime defect was established, and the
timeout must not be presented as a game blocker or justification for a repair.
There was no second invocation or replay after exit.
Both prescribed gates cleared before shared edits, tests, boat or runtime:
GL pixel-map worker6843 and Die by the Sword worker39566 each had launcher
exit 0, turn.completed, result/final files, exact AUTHOR RELEASE and
independently absent named/scoped helpers. The latter had stopped before its
original launch because its own mutable transfer input changed. Tachyon's fresh
boat was bx_wur5nmgf, finite TTL 3,600 seconds, with original expiry
2026-11-02T15:31:27.538Z recorded before use. Local launch free space was
2,500,419,584 bytes; the guard's recorded minimum through retrieval was
2,418,479,104 bytes, above the owner's 2 GB floor.
All 15 original/manifest identities (92,143,006 bytes) matched before launch,
on the boat and after exit. The established registered tachyon_demo arguments
and configuration mapping were retained. The accepted GL module was reused as
SHA-256 bdbe80167ac6970208664c1bb3241b667eabb4f7e25c6e18fbcdaedb8255cb5c.
Of its 7,230 recorded source-file identities, 7,229 were byte-identical; the sole
difference was the predecessor's newly published Die by the Sword findings note.
Every compiler/runtime input matched. Transfer verified 4,873 fixed file hashes;
post-run checks matched those inputs and all 124 observed actual file reads.
The observer's initial synthetic check lacked the existing local calc.exe
fixture; copying and hash-checking that fixture corrected the preparation before
Tachyon started. Sync/async actual emitted-callback, whole shared 512 MB memory,
guest-clock read-only and bounded retention checks then passed on the boat.
No game, emulator, configuration or guest-memory patch was made.
The shipped readme and both accepted native keychart pages were reviewed. The
first Launch release and immediate cursor-centering shared a dispatch batch and
left the game in the ship bay. A separate ordinary Launch hold/release, followed
by settled release before centering, opened the training briefing; its next
neutral interval reached the cockpit. This establishes successful input pacing,
not a general mouse defect. The previous Insert checkpoint was traversed, not
counted as the result. Bracket ] visibly selected Luna, then Heliopause; [
returned to Luna. T, a native-character T sequence, X and Enter did not visibly
select Luna in their tested views. X changed the weapon display and is not
qualified here as a targeting control.
Centered-cursor neutral intervals at zero throttle showed a stable view and Luna range near 37.1. Ordinary arrow holds changed the view; before/neutral/held/ released/settled captures preserve the differing responses. Short fixed-batch holds did not give uniform fine corrections. In one consumption-paced left tap, the keydown was dispatched only after three stepped batches (5647), followed by keyup at 5648. These observations do not isolate a clock, input or physics bug. Small cursor offsets were tested with matched centered intervals, then restored to (320,240); they did not establish useful fine steering. The earlier cursor confound is therefore not promoted to a proven navigation cause.
Full throttle and short arrow corrections produced the following reviewed approach, with Luna selected and the objective unchanged throughout:
| Capture | Batch | Visible Luna range |
|---|---|---|
approach-right-trim-settled.png |
5634 | 34.08 |
consumed-left-settled.png |
5658 | 32.48 |
luna-approach-neutral200.png |
5898 | 17.34 |
near-luna-forward100.png |
6015 | 8.84 |
final-luna-forward60.png |
6092 | 4.58 |
The last control snapshot has empty scheduled, pending, cross-thread and renderer queues, and all sampled movement keys have physical/async down state false. Pressed-since-query latches remain set for some released keys; they are not held key evidence. Guest post-queue and SDL consumption remain unknown. All 27 actual key down/up pairs and six mouse down/up pairs balance in the dispatch trace. A later requested 40-batch step found the original stdin already absent, so it executed no further game input. The exit PNG omits the objective overlay and is not substituted for the final controlled objective capture. No gate animation is counted as arrival. No audio, frame-rate, mission-completion or broader demo claim is made.
All 139 PNGs passed CRC/deflate/dimension checks. The 25,628,677-byte archive
SHA-256 91f3ce07187cf8d7ef09b0f7782e441920d27bee3578bdda59a40c32598cc6e2
and all 310 evidence-file identities were verified locally before delegated
shutdown. The boat was confirmed STOPPED at 2026-11-02T14:58:36.936Z; remote
runtime/controller and scoped helpers were absent beforehand, and local helpers
are absent at closeout. The local guard's shell reported SIGTERM/143 without a
final-file write; its last sample follows game exit and retrieval, and the
signal source is unestablished. cleanup.json preserves this limitation.
result.json, final.txt and RECOVERY.md retain the exact outcome and recovery
instructions. A future authorized attempt should retain the successful Launch
release boundary, select Luna with brackets early and spend the bounded session
on its final approach; it must independently establish arrival/objective change.
Do not resume this worker, automatically replay its scripts or repair generic
code without separately establishing a cause and bounding that implementation.