Saga

One player is the Harrowed: a single siege predator. Up to eight play Architects: defenders bonded to pre-placed strongholds called Bastions. They fight over map control, survival, and a shared resource called Divine Light, on a map called The Wound. The mode is Hunted Grounds.

It is deliberately slow. Third-person, grounded, weighty: there is no free jumping, and that is a rule rather than a missing feature. What it is not, stated as plainly in the design canon as what it is: not a twitch arena fighter, not a hero brawler, not a hero shooter, not a chaotic sandbox, not a base-building RTS, not survival crafting. Stronghold defence under pressure, a controlled economy, and an endgame that tightens.

The Last Light starts somewhere unglamorous: a StarCraft II custom map called Catch 'n Run, played to death with friends as a teenager. One side hunts, one side builds and survives: the whole evening turning on whether the defenders could hold. Nobody involved was a designer. Everybody had opinions about what was wrong with it.

That is the actual origin. Not a market gap, not a genre analysis: a game we loved that we spent years talking about fixing. The conversations were always the same shape: the hunter is unstoppable or useless with nothing in between, the defenders' best move is boring, the match either ends in four minutes or drags forever. This project is the attempt to build the version those conversations kept describing.

A note on sourcing: that map is not documented in any public archive we can find, so nothing above is a citable fact about StarCraft II. It is a recollection, and it is recorded here as one.

The visual and moral contrast is the design's spine: sacred blue light against red corruption, one map with a strong identity rather than a rotation of arenas. The Wound is a place that has been damaged and is being fought over: the Architects bond to Bastions and try to hold ground that means something; the Harrowed is the pressure that ground is held against.

Both sides climb the same ladder from opposite ends. Architects win by killing the Harrowed, or by completing the Ascension Engine: a long, expensive channel they have to protect while it runs. The Harrowed wins by wiping the Architects when they have no Rite of Return left, or by completing Bound Ascendant, its final evolution. Divine Light pays for all of it, and both sides earn and spend it in different ways. Escalation is meant to be readable: you should be able to feel the match tipping before it tips.

The ambition is narrow and hard: make one-against-many feel great from both chairs. Asymmetric games usually fail in one specific way: the rare role is a power fantasy and the common role is chores, or the reverse. The target here is that the Harrowed player and the eighth Architect both leave a match feeling like the outcome was theirs.

That is a balance problem before it is an art problem, which is why the work so far has gone into economy, pacing, and behaviour rather than polish. It is also why the design fixes the tone first, slow, heavy, no jumping, because a pacing rule you can break is a pacing rule you will break the first time a fight feels sluggish.

The game runs. It is built in s&box: earlier Unity work is legacy, and the design canon says so explicitly so old documents cannot quietly become current instructions. The validation target is 1v1 first, with systems structured to scale to the full one-against-eight rather than hardcoded for duels.

The most useful thing so far came from the first live voice playtest, which produced thirty-four findings in one sitting. The headline was not a bug: matches were ending in four to seven minutes against a thirty-minute target. Objective thresholds were roughly quadrupled and event progress halved in response, and the balance work since has run through several waves of bot behaviour, including per-round hunter temperaments so the AI stops playing the same match every time. That gap between the intended match and the actual match is the honest state of this project.

This is the part worth being precise about, because “released” means two different things and only one of them is true here. The game is published on s&box’s own platform, where anyone with s&box can download and play it. It is not on Steam as a standalone title, it has never been sold, and it has no commercial launch behind it. What it has is a public listing, a version history, and a meter running on it.

Platform figures as reported on the s&box listing, 1 August 2026. Reviews stood at 75% from two ratings; the package was last updated in mid-June. Average session length is not the same measurement as average match length, a session can hold several matches, or a menu and an abandonment, so treat it as a proxy rather than a match clock.

Ten minutes. The playtest above found matches ending in four to seven against a thirty-minute target, and the response was to roughly quadruple objective thresholds and halve event progress. Months later the platform’s own meter puts the average session at ten minutes: better, and still a third of the intended match. The clock problem is not fixed. It is publicly not fixed, which is a more useful place to be than privately optimistic.

Tagged 1v8, asymmetrical, siege, builder, strategy, multiplayer, pvp, horror, monster, and enrolled in the Play Fund, s&box’s payout pool, which pays creators on how much their game is actually played rather than on sales. Thirteen hours of playtime is not income. It is a signal, measured by someone other than us.

The claim that it runs on Source 2 is worth stating carefully, because it is unusual enough to sound like marketing. It is accurate: Facepunch’s own repository describes s&box as an engine built on Valve’s Source 2 with a modern .NET layer, and the Valve Developer Community documents it the same way. Source 2 is the engine behind Half-Life: Alyx. This is the first time Valve has licensed it to a third party: s&box is the only place outside Valve where it is available to build on. Facepunch put s&box’s own source under the MIT licence in November 2025; the Source 2 code underneath stays proprietary.

Day to day the part that matters is less glamorous: gameplay is written in C# rather than C++, and a hotload system applies code changes without a rebuild. For one person iterating on a balance problem, the difference between a fifteen-second reload and a two-minute compile is the difference between testing an idea and deciding not to bother.

Because it is the same shape as the thesis. An industrial-grade production tool, one that until now sat inside a single studio, became available to a solo builder, along with a route to distribution and a payout mechanism, with no capital raise and no studio around it. Moduloa’s argument about manufacturing is that capacity stops being a fortress when the tooling and the certification travel. Games got there first. It is worth watching what happened next in a field that has already made the transition, rather than only theorising about the one that has not.

Facepunch has a licence from Valve that lets projects built in s&box be exported and published on Steam as standalone games, royalty free. The plan for The Last Light is to use it early rather than late: put a rough build on Steam, free, so friends can install it and play without owning s&box first.

A Steamworks account exists for this, registered in the founder’s own name, Sondre Hegerland Kristiansen, rather than under a company. That is the honest scale of the thing. It is not a publishing arrangement and there is no launch behind it.

Partly because the game began as a custom map played with friends, and handing the first Steam build back to the same people closes a loop that started fifteen years ago. Mostly because of the design goal. “Fair from both chairs” cannot be validated against bots, and it cannot be validated at 1v1: it needs a full lobby of humans who will say the unflattering thing out loud. Eighteen players on a platform listing will not produce that. Eight friends in a voice call already did it once, and produced thirty-four findings in a sitting.

There is no announced date, no store page, and no price: free is a decision, not a launch plan. When there is a Steam page, it will be linked here.

Both readings point the same way. The licence bill for a text-driven game rounds to nothing, and the failures teams report are about how much was promised rather than how it was built. A branching story has no systemic multiplier: every scene a player might reach is written, edited, localised or cut by a person, so the work scales with content and not with the engine. That last step is our inference, not a finding: the dataset codes problem types, not genres, and no public dataset we found separates narrative titles from the rest.

Three tests. One: a scope statement written before production in countable units, scenes, words, minutes, divided by an observed production rate, because “unrealistic scope” is a number nobody wrote down. Two: a slice that survives being public, graded against the store’s own visibility surfaces, not internal enthusiasm. Three: a revenue expectation set at the median of the field rather than its head, planning against the roughly 300 releases a year that clear $1 million is planning against a rounding error. None of the three is written down here. Until they are, this page stays as thin as the record deserves.