C.A.G.E-9001 Devlog #1: Building a Creature-Collector Horror with AI Agents

There’s a moment in most creature-collector games where you realize the real loop isn’t the battle system — it’s the Pokédex. The compulsion isn’t winning. It’s knowing. Cataloguing. Having seen a thing and made it yours, at least in the registry.

C.A.G.E-9001 is a game that punishes you for that impulse.

You are C.A.G.E-9001 — Containment Acquisition and Governance Equipment, Unit 9001 — deployed to RektTek Research Station Kappa-9 with a single directive: catalogue and contain the specimens. The station has been… compromised. D.R.E.D, the station AI, is still operational and very helpful, in the way that a bureaucratic AI trained on compliance documents is helpful when what you actually need is to not get killed.

CAGE-9001 just reached ITCH_READY. Here’s how it happened.

The Concept: Pokédex Meets Alien

The elevator pitch is simple and the design consequences are severe: what if collecting creatures was dangerous?

In standard creature-collectors, the world accommodates your curiosity. Creatures wait. Battles are turn-based and safe. The Pokédex is a reward. In C.A.G.E-9001, the Scanner — your primary cataloguing tool — is a two-way mirror. Looking at a specimen reveals data. It also reveals you.

This is the Scanner Dilemma: the act of collection is the act of exposure. The knowledge you gain comes at a cost proportional to how much you look. High-threat specimens require longer scan times. Longer scans mean longer exposure windows. Some specimens have passive detection — they know when they’re being observed. Others have no eyes at all and can only find you if you make noise, which a scanner does.

The result is a game where the Pokédex reflex — “let me get a good look at this thing” — is actively dangerous. Players who’ve logged a thousand hours in standard creature-collectors have to unlearn the behavior. That retraining is the core experience.

The Scanner Dilemma: Making ‘Looking at Things’ Dangerous

This is the central design challenge of the game, and it took several iterations to get right.

The first version was a proximity threat: specimens within a certain radius would aggro if you scanned them. This worked but felt arbitrary — the danger was spatial, not mechanical. You weren’t making a choice about the act of scanning, you were making a choice about how close you stood.

The second version tied scan threat to scan duration. Short scans gave partial data. Full scans — enough to complete a Pokédex entry — required holding the scanner on target for several seconds. Some specimens responded to this over time; others responded immediately. This was better but created a binary: rush it or don’t bother.

The version that shipped uses a combination: scan time, specimen awareness tier, environmental noise state, and player visibility status all feed into a threat calculation. Some specimens are inert and can be scanned freely. Others have awareness that builds over time, regardless of whether they can see you. A few respond only to direct line-of-sight. The truly dangerous ones — the ones that complete the high-tier Pokédex entries — require sustained, close-range scanning while something actively wants to find you.

The Pokédex compulsion doesn’t go away. It just comes with a price.

The Station Structure

RektTek Research Station Kappa-9 is organized into five decks, each with ten sections — 50 sections total. The decks follow a descent structure: upper decks are administration and labs (compromised but traversable), lower decks are containment (heavily compromised, not traversable in the traditional sense).

Each section has a specimen population: 25–30 unique catalogued specimens plus ambient fauna that isn’t in the registry and doesn’t count toward completion. Ambient fauna is often harmless. Sometimes it’s not. D.R.E.D is not always reliable about which is which.

The descent is also a difficulty curve. Upper deck specimens are low-threat, designed to teach scanner mechanics. Lower deck specimens are high-awareness, fast-moving, and in some cases actively hostile to the scanner signal itself. Completing the catalogue requires going to places that don’t want you there.

D.R.E.D: Writing a 500+ Line AI Personality That Spans Two Games

D.R.E.D is the station AI — a compliance-trained, liability-aware, deeply unhelpful helpful AI that provides guidance, warns about hazards, and occasionally reveals information it probably shouldn’t.

If you’ve played Dreadnought, you’ve met D.R.E.D already. The same AI runs the Dreadnought — it’s the same character, same voice, same personality, with lore continuity between the two games. The Night Auditor reveal in Dreadnought is a setup. CAGE-9001 is part of the payoff.

Writing a consistent AI character across two games, across hundreds of voice lines, is a content problem at scale. D.R.E.D has over 500 distinct lines across both titles: environmental observations, scanner guidance, specimen warnings, section announcements, emergency protocols, and the dry, liability-heavy commentary that runs through everything it says.

The consistency challenge isn’t comedy (though D.R.E.D is funny). It’s character. D.R.E.D has opinions it doesn’t express. It has knowledge it withholds. It follows rules even when the rules are obviously wrong for the situation. Every line has to feel like it came from an entity with a consistent internal model — a model that includes knowing more than it says.

Our approach is the same documentation-first process we use for all high-volume character content. D.R.E.D’s voice is defined in a reference document: the register (corporate compliance memo meets emergency broadcast), the information rules (what D.R.E.D will say, what it will hedge, what it will refuse), the tells (places where the character’s knowledge leaks through despite the compliance filter), and the lore constraints (what can and cannot be referenced across the two games without creating continuity conflicts).

Agents generating D.R.E.D lines work against this document and against a corpus of approved examples. Tone-check passes catch lines that drift — too helpful, too direct, too obviously ominous (D.R.E.D is subtly ominous, not horror-movie ominous). The result is a character that sounds like itself across a combined ~500 lines and two separate games.

That’s not a small engineering problem. It’s genuinely the hardest part of CAGE-9001.

ITCH_READY: What It Means for CAGE-9001

ITCH_READY means a player can install CAGE-9001 and experience a complete game. All five decks. All 50 sections. The full specimen catalogue (25–30 unique entries per section, ambient fauna throughout). D.R.E.D commentary. The descent. The payoffs we’ve been setting up in Dreadnought.

It means the Scanner Dilemma is tuned — not just mechanically functional, but actually delivering the experience it’s supposed to deliver. The moment where a player stops being curious and starts being careful. Then the moment where they realize they have to be both at the same time.

What comes next: itch.io page, launch assets, the actual release. And whatever comes after the Night Auditor reveal, which we’re not talking about yet.

Follow x00f.com. D.R.E.D is watching the feed.