chatcat-litterbox

An ethological simulator for testing Animal-Computer Interaction (ACI) agent policies against simulated cats, before any real cat is involved.

This is an experiment, not a product. Stage 1 is simulator-only: a SimCat modelled on published feline ethology interacts with a rule-based ChatCatAgent inside a browser sandbox we call the Litterbox. No real cats are used at this stage. The system is designed so that cats can always say no, and the system listens.

Built on published ACI research, following the lineage of Cat Royale (Blast Theory, Mancini, Mills, Nottingham; CHI Best Paper 2024, Webby Award 2024) and grounded in Clara Mancini's Animal-Computer Interaction framework. The cat is the primary user; the human is observer.

Ntalampiras et al. (2019) showed a proof-of-concept that the emission context of a meow can be algorithmically classified across three contexts (waiting for food, isolation, brushing). MeowTalk stretched this into translation claims beyond that scientific basis — as one of the app's own creators told the New York Times, "It's not pure science at this stage." Proof-of-concept overstretched to product, admitted by the product itself, is exactly the failure mode we are designing against. Every behaviour parameter in this simulator traces back to a published, peer-reviewed source.

What this is NOT

• Not a cat translator. We do not claim to decode cat language into human sentences. Vocalisation types are modelled from Schötz's Meowsic categorisation for simulation fidelity, not for "translation".
• Not a substitute for veterinary care. Stress scores are research instruments (Kessler & Turner 1997), not diagnostic tools.
• Not entertainment for humans at cats' expense. The ethics monitor enforces session caps, stress thresholds, and opt-out detection. It was built before the UI.
• Not a finished system. This is v0.1 of a simulator. The path from here to real-cat interaction requires ethics review, institutional oversight, and iterative validation.

Architecture

The Litterbox contains three independent subsystems:

1. SimCat — A state-machine model of domestic cat behaviour, parameterised by Litchfield et al.'s Feline Five personality dimensions and driven by Kappel et al.'s ethogram. Five named archetypes provide starting points; continuous interpolation is supported.

2. ChatCatAgent — A rule-based policy (v0) that selects actions from a constrained action space. Hard-coded ethical safeguards prevent the agent from escalating when the cat signals stress or withdrawal. Learned policies come later; this version establishes the safety envelope.

3. Ethics Monitor — A separate module that the agent cannot bypass. It tracks CSS trajectory, opt-out events, habituation curves, and time-in-stress ratios. It forces agent pauses at CSS >= 5 and locks sessions at CSS >= 6. Hard action-level invariants — e.g., the retreat-state restriction (side_glance / soft_purr only, intensity capped at 0.3 in RETREATING/LEAVING) — are enforced in EthicsMonitor.enforce(), the gate every action path passes through before the simulator sees the action. All events are logged; the dashboard exposes everything the monitor sees. See ADR 0009 for the enforcement architecture.

Running

pnpm install
pnpm dev        # opens the Litterbox in your browser
pnpm test       # runs ethics-regression, archetype-coverage, opt-out-detection
pnpm build      # produces a static build

Requires Node 20+.

Scientific sources

See CITATIONS.md for the full list of papers, DOIs, and how each is used. Key sources: Kappel et al. 2024 (ethogram), Litchfield et al. 2017 (Feline Five), Kessler & Turner 1997 (Cat Stress Score), Humphrey et al. 2020 (slow blink), de Mouzon et al. 2022 (cat-directed speech), Schötz (Meowsic vocalisations), Mancini & Nannoni 2023 (ACI ethics).

Ethics

See ETHICS.md for how Mancini & Nannoni's four principles (Relevance, Impartiality, Welfare, Consent) are operationalised in this system. The ethics monitor is not optional infrastructure; it is the core feature.

Research direction

Architectural and methodological decisions are tracked as ADRs. See docs/decisions/ for the full index. The v0.2+ work items are self-play RL against SimCat (ADR 0002) and habituation-rate calibration against real-cat data (ADR 0003); the resolved ADRs document what we found along the way — reward design in 0007 / 0008, and the ethics-enforcement architecture in 0009.

License

Apache-2.0. See LICENSE.