One human, a long bibliography, some AI sub-agents

The Team

A small newsroom-lab, with a single byline.

~4 min read · 1015 words

The PS1 branch reads like the notebook of a small newsroom-lab: several distinct voices, all aligned on the same standard. The voice on a March agent prompt — sceptical, refusing to count black frames as wins — is recognizably the same project as the voice in the April foreground-pilot worklog that decides the captured pixels are the answer. The shared values aren’t framed as a manifesto but they show up in every commit message: evidence before confidence, reproducibility before storytelling, human review before claimed validation, memory and runtime reality over elegant theory, bespoke pragmatism when generic frameworks stop paying off, preserve the paper trail even when the direction changes.

The team that produced that work is not traditional. There is one human author. There are several prior open-source projects that this port stands on. There are a handful of AI assistants that helped draft, debug, and pattern-match through the long reverse-engineering stretches. That’s it. The rest of this page names what each contributor actually did, because the line between “wrote it,” “authored the work it depends on,” and “helped the author think” matters and is worth being honest about.

On this page

The author

Hunter Davis wrote the PS1 port. That includes every line in graphics_ps1.c, cdrom_ps1.c, sound_ps1.c, events_ps1.c, spi.c, pause_menu.c, foreground_pilot.c, the offline pipeline scripts under scripts/, the documentation, and the website you’re reading this on. Hunter also did the human signoff for FISHING 1 and FISHING 2 — sat in front of DuckStation, watched every variant, listened for missing or misaligned SFX, and decided when each was good enough to promote. The git history is sole-author. So is the credits screen on the disc.

The team-perspective archaeology document describes Hunter playing several roles across the project’s six-month arc: operator-editor who rejects convenient wins, embedded porter who respects the exact build scripts, instrumentarian who builds debug overlays, validator-skeptic who refuses to count title frames as scenes, archaeology engineer who maintains the historical binary library, bespoke scene engineer who authored the foreground playback pivot. Those roles are different jobs in the same skull. The unifying read: different voices handled different phases, but they stayed aligned on one ethic — make the machine legible, distrust convenient narratives, keep pushing until the screen tells the truth.

The prior ports

Three projects did the reverse-engineering this port stands on. None of them are part of this project’s team in the day-to-day sense; all of them are part of the bibliography in the load-bearing sense.

jno6809/jc_reborn is the SDL2 port that decoded the original engine’s bytecode — ADS, TTM, RES interpreters, the dirty-rect machinery, the resource format. Without that work the PS1 host build doesn’t exist. This project’s host capture is a fork of it.

nivs1978/JCOS is an alternate decoding effort, useful for cross-validating frames where the bytecode could be read more than one way. When this project’s capture disagreed with what a scene should look like, having a second reference implementation was the fast way to ask which interpretation was correct.

xesf/Castaway is more notes than running code, but the notes caught engine quirks that running interpreters elide. Several rare opcodes and edge cases got resolved by comparing against xesf’s reading.

The Sierra Chest archive kept the original JOHNNY.EXE and resource files in a verifiable form. Without that there’s no way to tell whether a decoding mismatch is the engine being weird or the input being corrupt.

The toolchain authors

The PS1 build is impossible without PSn00bSDK (Lameguy64 et al.), the open-source PS1 SDK that replaces Sony’s proprietary toolchain. The SPI pad driver in src/platform/ps1/spi.c is derived from spicyjpeg’s MPL-licensed pad-poll example — with the DuckStation-specific tx_len=5 fix on top. The emulator that every commit gets tested against is DuckStation (Connor McLaughlin et al.). The CD image is packed by mkpsxiso. The Easter algorithm that drives the movable-feast holidays is Meeus / Jones / Butcher. Each one of those names solved a problem that would otherwise have eaten a month.

The AI assistants

This project used AI assistants for drafting, debugging, and pattern-matching through the reverse-engineering stretches. That’s worth naming because pretending otherwise would be dishonest and because the kind of help matters. The assistants drafted documentation, summarized long worklogs, helped read SIO0 datasheet fragments, and pattern-matched against PS1-era reference code when something on DuckStation behaved differently from what the spicyjpeg example expected. They did not author scenes. They did not decide what counted as validated. They did not press signoff. The validator-skeptic role — the one that refused to count black frames as wins — was always Hunter’s, and the AI assistants sometimes had to be reminded that running and correct are not the same thing. The drafts on this site went through human editing before they shipped.

The reader

This is a small distinction but worth making. The project’s “team” also implicitly includes the reader of the GitHub issues, the DuckStation forum thread, the friend who said the early build’s sky color looked off. None of those names appear in the contributors graph. They’re part of the loop anyway.

The team is small. The bibliography is long. The work is sole-author in the sense the credits page means it: one human accountable for every shipped commit. Everything else is generously donated upstream work this port could not have existed without. The site doesn’t put anyone’s photo on it. Hunter’s name is on the credits screen. The prior ports are linked from the Credits page. That’s the whole accounting.

  • Credits — the full acknowledgements page; toolchain authors, prior ports, AI sub-agents, fonts, ocean-ambience source.
  • About: Voice — the editorial standard the prose-drafted-by-agents had to be edited up to.
  • Docs: AI sub-agents — honest accounting of where the agents helped on this project and where they didn’t.
  • Lab: the LLM pass — methodology essay on running parallel agents under human review.
  • Lab: voice-anchor problem — long-form retrospective on keeping a single voice across human + agent prose.
  • Lab: dunking-bird — the auto-poker that keeps the agents productive between human review passes.