When does a mind learn what it was never asked to care about?
ItaSoRL raises a digital organism from scratch inside a simulated world. Sometimes that world is authentic. Sometimes it is a subtly flawed copy. We never tell the organism, and never reward it for noticing: we read its internal state and ask when the difference gets encoded anyway.
A flaw in the world can be trivially detectable from the outside and still leave no trace inside the mind. That gap is the object of study.
The organism lives in "A Patch of Earth": a small, deterministic, Earth-like world. Sometimes it runs on authentic dynamics; sometimes on a surrogate carrying tiny substrate seams. Difficulty is dialed on a ladder from L0 (identical control) to L4 (adversarial).
Three invariants hold everywhere: tabula rasa (no imported semantics), readout not reward (world identity is probed, never trained), and substrate-grounded (the only tells are artifacts of the computation itself).
The experiments
Detectability ceiling
Agent-free first: can a privileged oracle spot the surrogate from the seams alone? Yes, almost perfectly - with calibrated difficulty and a clean confound audit.
Incidental encoding
Does a prediction-only agent encode the flaw it was never asked to care about? Robust negative at L2: the artifact an oracle catches almost perfectly leaves essentially no decodable trace, across three readout channels.
Survival changes it
Against a learned-dynamics fingerprint, the survival agent encodes the world where prediction-only sits near chance; survival-specificity holds at the subtler of the two tested fingerprints. Confirmed at n = 10 and audited: reward- and survivorship-controlled, with a ~0.73 behavior-independent signal, and held-out probes show the signal transfers across fingerprint instances (0.773) and even across surrogate recipes (0.684), though only from subtle training artifacts (the reverse direction fails the bar at 0.638). Re-scored with a corrected estimator, the common-garden control passes the frozen rule on both directions (0.666 forward, 0.684 reverse), so it reads as a modest persistent stored world-identity component the policy also expresses reactively, weak and tail-decaying rather than strongly stored (see FINDINGS 10.6.1).
Emergence under selection
Make survival depend on the difference and watch whether representing it emerges across generations. The L4 adversarial rung of the ladder waits here.
Detectable is not the same as learned.
After training, a linear probe reads the agent's internal state and asks: is world identity in there? At L2 the answer is flat at chance, no matter how loud the artifact is outside. At L3, survival pressure is the first condition that lifts the curve.
Recent work
Others told their AI to look for the fake, or used minds that already knew what "fake" means. We measure what a from-scratch mind comes to know unasked.
Chatbots can tell a test from a real conversation - but they read the internet, so the concept of "being tested" was imported, not discovered. A student with the answer key. Our organism is born knowing nothing.
Probing studies (AtariARI, Othello-GPT) found knowledge nobody taught - but always about things inside one fixed world: the pieces, the board, the score. Never about which world is running.
Robots get engineered alarms that fire when the world stops matching their training (GalilAI, RAPT) - an installed smoke detector, detection as the designed objective. We install nothing and only read.
To our knowledge, this is the first study to measure the gap between external detectability and spontaneous internal encoding of simulator authenticity - authentic base dynamics vs a learned surrogate - in tabula-rasa recurrent agents, with substrate-level artifacts as the only cue and world identity read out post-hoc, never rewarded.
the axis-by-axis cross-reference, with citations →Detectable all along. The whole finding, in 90 seconds.
A creature is born knowing nothing. We copy its world and change one rule of physics. An outside watcher catches the fake almost every time. The creature's own mind stays at a coin flip, until telling real from fake decides whether it eats.
Voice and music are in. This is the cut released on 24 August 2026. Turn the sound on.
Real runs, 10 each. 50% = coin flip, 100% = always right. Read out, never rewarded.
Same brain, two worlds, one flaw.
Now go inside the creature's head. One brain lives in the real world and in the copy. Five chapters follow the clue from the senses, through the processing layer, into the memory spine, and out to the motor heads. The flaw enters at speed, the sense the broken velocity law corrupts. Teal rings settle on the memory cells that end up holding it.
Nothing on screen is decoration. Every dot carries its real name from the observation spec. The twelve memory cells are real recurrent units of one representative brain, the seed closest to the published 0.752 mean, and ring strength is that unit's measured world-signal from the saved held-out pools. On-screen numbers are the canonical ones: 0.488 floor, 0.65 bar, 0.752 survival, 93% watcher gate; vision masked 69%, every sense masked 50%.
58 s · five chapters · narration on screen, no audio track · rings placed from measured data
Open source. Reproducible. Run it yourself.
Every experiment is deterministic given its seeds. Clone the repo, or open the Colab notebook on a GPU runtime and run all cells.