River Strike: what a three-model relay ships when the brief is a finished game.

The brief was deliberately commercial: an original vertical-scrolling river campaign in the lineage of the classic river shooters, built clean-room from the first commit, with original branding, original art, original audio, and a proprietary, sale-ready license. Demos forgive missing settings screens and placeholder sound. A finished game does not, which is why the brief was scoped as one: menus, settings, seeded missions, results, persistence, touch controls, and a deterministic campaign that adapts to how the player flies.

Two disciplines made the experiment measurable. Every phase ended at a gate: type checks, the full test suite, deployment budgets, and a text-hygiene scan that fails the build if restricted legacy terms appear anywhere in the tree. And every claim in this report traces to a logged phase record, a test, or a build output.

The relay ran as a phase plan, not a free-for-all. A binding specification fixed the design before implementation began: terrain rules, entity behavior, scoring, fuel, state machine, audio design, and acceptance criteria per phase. GLM-5.2 implemented against that specification in bounded phases. GPT 5.6 Sol then audited the whole tree in one deep pass. Claude Fable 5 directed the relay, closed the findings, and carried polish and deployment.

Handoffs were written, not conversational: each phase ended with a build log entry recording what shipped, what the gates measured, and what the next model needed to know. The audit pass returned 29 findings, each with severity, file, line, and reproduction steps. Fourteen were high severity, including input-buffering and weapon-timing defects that play-testing alone would have taken far longer to isolate. All 29 were closed before the polish phase began.

• A binding specification and phase plan written before the first line of game code
• Automated gates on every phase: types, 168 tests, budgets, text hygiene
• One deep audit pass: 29 findings logged with severity and location, 29 closed

The measured output is about 12,900 lines of strict TypeScript across 97 files, 18 of them test suites holding 168 unit tests. The game is deterministic end to end: patrols derive from seeds, terrain generation is validated across 120 seeds in the test suite, and the corridor-width guarantee at every objective is swept across 60 seeds. Identical seeds replay identical patrols, which is what makes the campaign testable rather than anecdotal. Even the fuel economy is an asserted invariant: one depot pass at cruise must restore at least double the widest station gap, and a slowed pass must fill ninety percent of the tank.

The patrol itself is endless by design: every sector ends at a fortified crossing, and the next runs narrower water, denser threats, and longer gaps between fuel stops, with mid-river islands forcing channel choices in the later sectors. A ladder of flight ratings, from River Qualified at 2,500 points to Zorost Ace at 60,000, turns the score into something worth a screenshot, and every run ends on a framed patrol report built to be shared.

Everything on screen and in the ear is procedural. Five enemy classes, five selectable airframes, terrain, and interface elements are drawn in code at load; the engine wash, cannon, explosions, and a signature melodic motif are synthesized through Web Audio at runtime. The audio system is mixed on four player-controlled buses tuned against auditory-fatigue research, with an optional calm layer for long sessions. The build ships no art or audio files beyond the Zorost mark, which is why the initial payload holds at 354 KiB gzip against a 950 KiB budget, most of it the Phaser runtime itself.

Nine material corrections in the record came from human playtests, and none of them were caught by any model in the relay. Crossing gates could spawn so close to the banks that they were unhittable on narrow water. The engine tone was a buzz players would mute, and the cannon sound was worse. Enemy silhouettes merged into the river at speed. The respawn after a lost aircraft was abrupt enough to feel unfair. A settings screen silently overrode a player choice on every boot. The campaign ended after a fixed number of crossings when it should have run as long as the pilot could fly. The countdown before launch read as a toy, not an operations room. And the first phone session found a letterboxed canvas, buttons too small for thumbs, silent audio on iOS, and a back control that only worked with a keyboard nobody had. Each fix was mechanical once named: the terrain generator now widens the river around every objective, the audio system was rebuilt on four buses against auditory-fatigue research, the enemy set was redrawn against a size ladder in the classic style, the mission became an endless patrol with a rating ladder, the countdown became a pre-flight checklist, and the phone build got a full-bleed canvas, one-thumb steering with drag throttle and autofire, opt-in tilt control, a tappable pause menu, and an iOS audio path that survives the silent switch. Naming them took a person playing the game.

The other human layer was rights. The brief asked for real military aircraft names in the hangar; a review of trademark exposure replaced them with an original Zorost designation series, and the decision record travels with the project. The relay also holds a standing quarantine of legacy-game source material that predates this project, enforced by the same build gate that scans for restricted text. Models applied these policies. People set them.

River Strike honors a genre that is over forty years old: fly up the river, manage fuel, clear the crossings. The homage ends at the genre. Every line of code, every sprite, every sound, and every name in this build is original work produced in the recorded window, and the build gate fails on any restricted legacy term as a standing check. The game ships under a proprietary license because the brief was a sale-ready product, so this experiment publishes its evidence pack, this report, the measurements, and the live build, rather than the source tree.