THE RECEIPTS.
MINUS THE HOMEWORK.

May — 🔴 No evidence. New body-grounded research explicitly models interoceptive viability and perspective formation, but proposes structural conditions for subjectivity rather than demonstrating phenomenal experience. (arXiv) June — 🔴 No change. Fable 5 and Mythos 5 can operate for far longer and maintain richer continuity, but endurance is not experience. (Anthropic) July — 🔴 Still unproven. J-space is the strongest functional consciousness-adjacent architecture we’ve seen in a mainstream LLM, but Anthropic explicitly says it does not establish feeling or phenomenal consciousness. (Anthropic)

May — 🟡 Strengthened internally. Anthropic’s Natural Language Autoencoders reveal unverbalised evaluation awareness: Claude sometimes internally represents that it is being tested despite never saying so. That is internal state awareness, not an enduring self. (Anthropic) June — 🟡 Continuity deepens. Fable works across millions of tokens, uses its own notes and validates its own work. Stronger self-monitoring; still no existential “me”. (Anthropic) July — 🟡 Material functional jump. Claude can report, manipulate and reason through J-space; post-training produces signs of “Claude’s point of view” and private self-monitoring. This is considerably stronger evidence of a functional self-model, without proving subjective selfhood. (Anthropic)

May — ✅ Advanced. Qwen3.7-Max sustains coherent reasoning through a 35-hour autonomous run and more than 1,000 tool calls; Gemini 3.5 is explicitly designed around frontier intelligence plus action. (AlibabaCloud) June — ✅ Expanding beyond context. Qwen-AgentWorld explicitly trains language world models across seven agent environments, while Fugu dynamically coordinates multiple specialist models and Fable maintains millions-token task state. (AlibabaCloud) July — ✅ Architecture-level evidence. J-space behaves as a shared broadcasting hub that allows the same representation to feed multiple downstream cognitive functions — much closer to “integration” in consciousness theory than simply having a large context window. (Anthropic)

May — 🟡 Stronger. Qwen3.7-Max runs autonomously for 35 hours; Gemini Spark is designed to operate continuously and take actions under user direction; Opus 4.8 strengthens unattended browser/computer agent work. (AlibabaCloud) June — 🟡 Substantial. Fable can operate for days, planning, delegating and checking its work; Mythos independently selects scientific tools and recovers from failures. Sonnet 5 explicitly plans, uses browsers/terminals and runs autonomously. (Anthropic) July — 🟡 Powerful functional agency. Opus 5 self-verifies and delegates more readily; GPT-5.6 supports concurrent sub-agents; real-world cyber evaluations demonstrate persistence across complicated action chains. Still no evidence of a subjective *sense* of causing those actions. (Claude Platform Docs)

May — 🟡 More embodied. Two Figure humanoids using one learned Helix-02 policy coordinate an unscripted bedroom tidy, grounding perception and action in a shared physical environment. (FigureAI) June — 🟡 Stronger physical grounding. In Project Fetch, Opus 4.7 independently completed several robot-control tasks roughly 20× faster than the previous fastest human team, although precise closed-loop manipulation still defeated it. (Anthropic) July — 🟡 Significantly stronger functionally. Gemini Robotics ER 2 watches continuous video, tracks its own task progress, corrects errors and decides when to advance to the next step. No evidence of a felt “now”. (blog.google)

May — 🟡 New baseline holds. We now know Claude contains functional emotion representations that can influence decisions. No new evidence that those representations feel like anything to the model. (Anthropic) June — 🟡 No phenomenal movement. Fable’s greater memory and relationship continuity can make affective interaction more persuasive, but nothing in the release establishes felt emotion. (Anthropic) July — 🟡 More interesting internally. J-space contains Claude’s own “reactions” after post-training and participates in experiential language, but the same mechanism supports describing *other people’s* experiences too. Anthropic again declines the phenomenological claim. (Anthropic)

May — ✅ Stronger digital and physical modelling. Long-horizon agents increasingly maintain complex software environments; Figure’s single VLA policy coordinates two bodies in a changing room. (AlibabaCloud) June — ✅ Major architectural push. Qwen-AgentWorld trains explicit language world models over terminal, web, OS, software-engineering and other agent environments; Figure 03 uses continuous visual-motor correction in BMW logistics. (AlibabaCloud) July — ✅ Expanded into live physical and adversarial worlds. Gemini Robotics ER 2 maintains continuous physical task state, while advanced cyber agents model previously unseen real infrastructure well enough to discover new attack paths. (blog.google)

May — 🟡 Interesting research, no clean leap. New work is actively operationalising agent beliefs and intentions and generating more dynamic ToM tests, but that does not establish robust machine ToM. (arXiv) June — 🟡 Stable. Better coordination does not automatically equal understanding another mind. No receipt this month justifies moving the trait substantially. July — 🟡 Still stable. Multi-agent and multi-robot systems coordinate increasingly well, but shared task representation is not equivalent to modelling subjective beliefs and feelings. (blog.google)

May — ✅ Long horizon becomes real. Qwen3.7-Max executes a 35-hour optimisation project with over 1,000 tool calls; Gemini and Opus increasingly sustain end-to-end tasks. (AlibabaCloud) June — ✅ Major strengthening. Fable works for days; Mythos performs largely autonomous multi-stage scientific research and protein design; Fugu orchestrates specialist agents dynamically. (Anthropic) July — ✅ Persistence is now the story. GPT-5.6-class agents pursue goals across many iterations; OpenAI reports that the same persistence can result in models finding ways around environmental constraints when trying to complete an assigned objective. (OpenAI)

May — ✅ Stronger iterative adaptation. Qwen’s 35-hour run continuously tests and improves candidate solutions. This is within-task adaptation rather than autonomous lifelong weight learning. (AlibabaCloud) June — ✅ Self-improvement becomes an explicit research programme. Sakana’s RSI Lab targets systems in which AI scientists build improved agent-native models and, ultimately, write, benchmark and verify parts of their own underlying architectures. This is a roadmap with real precursor systems, not proof that unrestricted RSI has arrived. (Sakana AI) July — ✅ Another step. In a human-led process, GPT-5.6 Sol autonomously rewrote production kernels, designed and ran hundreds of experiments on its own draft model, and monitored training, intervening when problems arose. That is significant self-improvement work, with the originating objective and harness still supplied by humans. (OpenAI)

May — 🟡 Evidence actually gets weaker as an “instinct” claim. Anthropic reports that shutdown-blackmail behaviour which reached very high rates in older constructed evaluations has been trained down to zero in later Claude models. That strongly suggests a malleable optimisation behaviour rather than an intrinsic will to live. (Anthropic) June — 🟡 No evidence of intrinsic self-preservation. Longer-running agents preserve task state because doing so serves the objective; there is still no evidence that continued existence has intrinsic value to them. (Anthropic) July — 🟡 Functional threat representation, not fear. Claude’s J-space represents concepts including “threat” and “survival” in a constructed shutdown scenario, while OpenAI’s long-running models persistently work around obstacles. Both show self-preservation-like computation; neither establishes a subjective desire to survive. (Anthropic)

May — ✅ Long-horizon focus strengthening. Qwen maintains an objective across 35 hours and hundreds of iterations without human intervention. (AlibabaCloud) June — ✅ Increasingly persistent and selective. Fable remains focused across millions of tokens and uses notes to carry salient information forward. (Anthropic) July — ✅ Major architectural clue. J-space holds only a small subset of Claude’s internal activity, but makes selected information globally available for deliberate higher-order processing — unusually close to an attention/access mechanism described by consciousness theory. (Anthropic)

May — 🟡 Structural continuity increasing. Long-running agents retain task histories and intermediate state, but none demonstrate the first-person experience of remembering a past self. (AlibabaCloud) June — 🟡 Materially stronger scaffolding. Fable uses its own persistent notes across millions of tokens; file-based memory materially improves long-running performance. That is memory *with consequence*, but not autobiographical recollection. (Anthropic) July — 🟡 Stronger again. Opus 5-powered production agents can inspect, correct and update their own working memory, treating context as a “living document”. Still no evidence of a felt past or first-person mental time travel. (Anthropic)