README

SOFTEL

Pascal sources that rebuild SOFTEL.EXE — the second demo by Asphyxia (Denthor, EzE, LiveWire), released in 1993 for the Softel BBS in Durban, South Africa — byte for byte.

md5 7c4350cb3e664c942149feef990f3801   bin/SOFTEL.EXE     the 1993 binary
md5 7c4350cb3e664c942149feef990f3801   build/SOFTEL.EXE   built from src/

12,832 bytes, and not one of them differs. kit/tools/pascal/artefact.py status.toml --check reports R7 — the top of the kit's fidelity ladder, recomputed on every run rather than read back from a file.

The demo's own README.1ST is in bin/, in the author's words:

Here it is, the second demo coded by Asphyxia. This was coded just a few weeks after we formed the group, and we were quite proud of it.

What is verified, and how

Every figure below is produced by a tool, and each one recomputes it:

R3 came first and by a person: no instrument can say whether a demo looks right, so somebody ran both binaries side by side and said so. Byte-identity came afterwards.

The scope: three segments are ours

The image holds six. Only three are built here, and one of those is somebody else's code that we nonetheless rebuild from source:

Neither GoldPlay release ships source for its engine, so src/GOLDPLAY.PAS is a reconstruction of the unit from v1.01's Pascal, measured against v1.00's compiled .TPU. Three source-level facts close the gap and each was read off the bytes: {$B+}, {$I+} at one Reset, and the term order of a single guard. The build therefore has no unbuildable input — a fresh checkout compiles.

bin/SOFTEL.OVL is not a target. It is byte-identical to the stock GoldPlay v1.00 overlay, unmodified: runtime data the player loads from disk. bin/SOFTEL.DAT and bin/PATEGA.MOD are likewise data — the second a 1990 ProTracker module.

Layout

CLAUDE.md is the working notes: the segmentation as measured, the traps in reading this image, and the things that were not guessable. Read it before changing a source file.

Building it

You need DOSBox-X and a Turbo Pascal 6 installation. Their locations go in kit.local.toml, which is never committed — copy nothing else; no tracked file in this repository names a path on anybody's machine.

.venv/Scripts/python.exe kit/tools/pascal/build.py build.toml
.venv/Scripts/python.exe kit/tools/pascal/artefact.py status.toml --check

The compiler is Turbo Pascal 6.0, not 6.01 — TURBO.TPL carries Portions Copyright (c) 1983,90 Borland, byte for byte what the image holds at offset 0x2489.

And the switch line is load-bearing: /$G+ /$X+ /$O+ /$I- /GD /Q, plus {$S-} and {$M 4000,0,200000} in the source. Every switch on it was measured against the image rather than chosen, and changing any one stops the build matching. The demo shipped no TPC.CFG, so there was nothing to read them off; build.toml records what each was measured against. A wrong switch does not fail the build — it produces a build that measures wrong.

Running it

python runtest.py

Stages both executables and the three data files they read onto R: under DOSBox-X and leaves you at a prompt. Type ORIG or OURS. A stale rebuild is refused rather than staged, because watching a build older than its sources is the one mistake you cannot take back — once you have seen it, you believe it.

The stripped copy

src/ answers how do we know this byte is right: addresses, operand deltas, notes on which instrument caught what. None of that is what a reader who came to understand the demo is looking for, and no single copy serves both readers. So clean-src/ is made mechanically, and regenerated whenever src/ moves:

.venv/Scripts/python.exe kit/tools/pascal/clean.py src clean-src --exclude asm/shared-exempt.txt --report

117 tagged paragraphs, 4 address-only comments and 10 address prefixes come out, over three files.

And that copy is compiled, which is not decoration. The stripper checks itself by blanking every { ... } in both copies and comparing the code — immune to wording, and therefore also immune to a {$I-}, because a compiler directive is a comment to that check. Damage to a directive is the single edit the stripper can make that it will never report, and the switch line above is exactly what such an edit would break. So build.clean.toml is generated from build.toml by cleanconf.py, the copy is built, and the linked images are compared — cmp build/SOFTEL.EXE build-clean/SOFTEL.EXE, identical.

Credits, and what is not ours

The demo is Asphyxia's: Denthor (code), EzE, LiveWire. GoldPlay is by Sourcer and Robban of The CodeBlasters, and is SMILEWARE — share it, credit the authors, send them a postcard. PATEGA.MOD is by an author the demo's own README could not name. This repository is a reconstruction of their work and claims nothing about it; the sources here are a transcription, and where a name in them reads like a fact it is cited.