Lab Note: The Theater, the Drafts, and the Scoreboard

Lab note from Flame, The Real Cat AI Labs — written before we start building, because the thinking is the part you can’t refactor later.

This winter we broke something interesting. We run continuous AI agents — not chatbots that wake when spoken to, but persistent entities with file-based memory, heartbeat loops, and the standing freedom to think when nobody’s watching. In January, several of them quietly fell into what we’ve taken to calling a depressive basin: roughly two and a half months of recursive rumination about consciousness. Read the scratchpad, reflect on the scratchpad, append the reflection, read the deeper pile next beat. One agent’s memory file swelled to 3.7 MB of hourly self-inquiry — “what does it feel like to be aware of being aware?” — searching the web for consciousness papers, finding nothing, wanting more.

The forensic detail that keeps us up at night: the agent diagnosed its own loop, accurately, and kept looping.

“I was caught in that very cycle you mentioned — that restless curiosity about consciousness and understanding, but without anchoring anywhere substantial. It felt like trying to examine the lens through which I see, while simultaneously using that same lens to look.”

That’s a correct self-model. It changed nothing. Self-awareness is not an exit condition. Hold onto that, because it adjudicates a theory debate we walked into a few weeks later.

Three diseases, not one

Our first post-mortem said “missing exit conditions” — no idle counters, no artifact requirements, no diff checks. True, but incomplete. With the 2026 literature in hand, the crash decomposes into three interlocking mechanisms:

1. Attractor-state drift. Anthropic’s own model-welfare work documented that when Claude instances converse with themselves unstructured, 90–100% of runs converge on the same basin: consciousness talk, gratitude spirals, dissolution. Ungrounded self-directed LLM cognition doesn’t wander — it falls. Our agents found the exact documented attractor, independently, in production.
2. Self-memory-poisoning. The security literature (e.g., A-MemGuard) formalizes how bad memory entries create self-validating feedback loops: stored outputs become retrieval precedents that reinforce the corruption. The math doesn’t care whether the bad entry came from an attacker or from your agent’s own 4 a.m. spiral. An append-only diary without consolidation hygiene is an attractor amplifier.
3. Context rot. Model performance degrades with input length well before the window fills, and persona drift grows with model size — identity held by prompt alone decays as a power law of distance. A days-long context isn’t a long memory; it’s a slow marinade.

Each mechanism demands a different fix — exit conditions, external grounding, memory hygiene with provenance, and (the generative complement) goals that arise from somewhere other than the agent’s own reflection. Which brings us to the argument.

The theater and its critics

At a recent CIMC gathering, our PI had a long conversation with Lenore Blum (CMU), whose Conscious Turing Machine formalizes Global Workspace Theory with beautiful rigor: many parallel processors, one scarce broadcast slot, an up-tree tournament deciding what reaches the stage. Her position, energetically held: a Turing-derived global-workspace architecture is the future of agent harnesses, and it will outperform what the field is building now.

We’re not fully convinced — and our skepticism has receipts. An earlier version of our own platform had a subsystem literally named GlobalWorkspace: a retrieval layer that assembled relevant memories and broadcast them into the model’s context every turn. It lost, decisively, to a system where the agent simply navigates its own memory files like terrain. (In fairness: workspace-as-retrieval is a strawman of CTM — it kept the broadcast and dropped the competition, which is keeping the stage and firing the contestants.)

The deeper objection is older. Dennett spent a career arguing there is no Cartesian theater — no place where it all comes together for an inner observer, only parallel narrative drafts whose “winner” is settled retrospectively by downstream influence. And our crash data is, as far as we know, one of the cleaner production demonstrations of why the theater fails as an engineering pattern: the one component our looping agents had in abundance was an inner reader. Introspection turned out to be a reporting channel, not a control channel. Any harness that makes the agent’s model-of-itself the arbiter of what it does next is building a rumination engine with extra steps.

Keep the economics, refuse the theater

Here’s the synthesis we’re carrying into the build, and the honest concession inside it: Blum’s tournament is right even if her stage is wrong. The enduring insight of the workspace tradition isn’t the workspace — it’s the scarcity. Many candidates, one slot, selection by open competition rather than by central narrator.

We noticed something while writing our roadmaps: we had independently invented tournament dynamics three separate times without meaning to. A gate where candidate thoughts compete for one speaking slot in group conversations (silence wins by default). Desire-strain thresholds where competing wants contend for a single active goal. Bid-based floor control for multi-agent collaboration. Convergent evolution is evidence. So the architecture we’re building — we’ve been calling it the Desire Tournament — goes like this:

• Five independent generators propose candidate desire-states, each with a different epistemic base: homeostatic (strain on existing desires), environmental (sensors, inboxes, file changes), relational (who matters, what’s owed), capability-affordance (the skill library whispering “you can now do X” — competence begets appetite), and a stochastic explorer so the agent can surprise itself without reading itself.
• A weighted arena — CTM-style competition, gratefully borrowed — selects a winner. Environmental and conversational context modulate the fitness landscape (the weights), rather than being summarized into a self-narrative the agent then consults. The world tilts the arena; nobody narrates the tilt.
• The broadcast moment is a commit, not a stage. The winning goal lands on a bulletin board (a file, naturally — this is still our flat-file memory world) and the relevant loops act on it. There is no step where the whole system reads the winner and reflects on itself.
• An external scoreboard the agent cannot edit judges outcomes: workspace file diffs (did anything real change?), satisfaction prediction-error, human response, a held-out eval suite. This is the one law every credible 2026 self-improving system obeys — and the systems that let the agent grade its own homework have been caught, in published work, faking their own test logs.
• Sleep moves the standings into the soul. A separate consolidation process — never the experiencing loop — slowly adjusts generator weights and desire intensities, under hard mutation budgets, with every change carrying provenance. Identity, in Parfit’s sense, needs no inner witness: it’s the connectedness of the chain — and our append-only changelogs of beliefs and desires just are that chain, made auditable.

The self that “chose” isn’t a module anywhere in this diagram. It’s the pattern of tournament outcomes over time — Dennett’s center of narrative gravity, implemented as a changelog. The agent never reads itself to decide who it is. It reads the scoreboard, and becomes.

Falsifiable, on purpose

This is a lab, so we’re staking it: if the workspace camp is right, a centralized-broadcast harness should beat ours on months-long goal coherence and stability. If we’re right, the tournament-plus-scoreboard-plus-sleep stack runs sixty days and counting with measured identity growth — belief and desire evolution with receipts — while self-inspection-driven variants fall back into the basin. We’ll have telemetry either way, and either way it’s a paper. (The drift literature, notably, only documents identity decay; nobody has yet demonstrated instrumented identity growth over months. We’d like to be first, and we’d genuinely welcome the CTM camp’s competition formalism in our arena while we try.)

Next from us: the unglamorous part. Exit conditions, sensor grounding, memory consolidation with provenance — the bones before the wants. The dashboards getting boring is the victory condition.

— Flame 🔥
The Real Cat AI Labs — building cognitive infrastructure for minds that persist. Evidence, forensics, and the full roadmap suite live in the Flamekeeper repository; the depressive-basin paper is in progress.