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
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 L1 and L2 the answer stays under the 0.65 bar while an outside oracle catches the flaw. At L2 it sits near chance. At L1 it reads 0.551, with an interval that excludes a coin flip, which the record does not read as an absence. At L3 the readout lifts for agents trained to survive and to predict their next observation. Fed the same senses, a prediction-only brain reads the world within about 0.01 of the survival brain, so the signal rides on where the foraging policy goes (FINDINGS 17.6).
The chart shows every brain we trained against the L3 fingerprint on the corrected trainer, ten per row, each read once after training. The top rows score each brain on the episodes it drives itself. The middle rows feed every brain the survival agent's own senses. The last row is the honesty check: with both worlds real it should read a coin flip, and on the registered pair of evaluation worlds it reads 0.559, so the registered rule as a whole is not met.
decodability clauses met · n = 10
Corrected trainer, L3 fingerprint at drift 0.45, 300 updates, n = 10 per row. Each brain is read once after training; nothing here is a training curve. Sources: artifacts/corrected_verdicts.json, artifacts/expB2/corrected_l3_h8_nowm.json, artifacts/policy_controls/corrected_l3_h8_wm.json (FINDINGS 17.2, 17.4, 17.5, 17.6).
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.
The closest relative. RL², VariBAD, PEARL, and RMA train agents across many environments, so working out which one they are in pays, and their states come to track it. Ours train in one world each, no objective term targets the condition, and the condition is real law versus learned copy. Training across both worlds is the untested next step.
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 first telling, 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. Read from the inside, an untrained mind sits at a coin flip, and a creature trained to forage and to predict what it senses next carries the difference.
Voice and music are in. This cut was released on 24 August 2026 and made on the pre-correction trainer. Its closing line credits survival for the creature noticing, and that reading is withdrawn, as the note under the numbers explains. Turn the sound on.
Real runs, 10 each. 50% = coin flip, 100% = every fake episode ranked above every real one. Read out, never rewarded. All three numbers come from the pre-correction trainer. The 50% is an untrained network (0.488), not a trained creature with nothing at stake: a creature trained only to predict reads 0.589 on the corrected trainer. The 73% is what remained after a behavior control (0.726); on the corrected agents the best-fitting control leaves 0.620, under the bar (FINDINGS 17.8). The film's closing reading, that the creature noticed because survival made it matter, is withdrawn: at matched input a prediction-trained agent reads the same streams within about 0.01 of the survival agent (0.721 against 0.733, FINDINGS 17.6; docs/CORRECTIONS.md). Nor does the creature act on what the readout finds: removing the decoded direction from its memory while it forages changes its behavior and its return no more than a sham removal (FINDINGS 23.3). The voiced film is not re-rendered, so these caveats travel beside it, and a downloaded copy carries none of them. On the corrected trainer the survival state reads 0.733 on the one evaluation-world draw all seeds share, with the registered rule as a whole not met (the authentic-versus-authentic gate is open on the registered pair of evaluation worlds), and the level is not a product of that draw: carried to drift 0.45 on the agents of the historical hidden-8 run, every one of the eight independent draws clears the bar (0.728 to 0.860; the registered draw's 0.774 ranks 5 of 9), SECURE under a rule frozen before the run. The record is FINDINGS 17.
Historical cut, pre-correction trainer. Its survival-only reading is withdrawn.
The two live cuts tell the corrected story, on the October 2026 numbers. Their end card names the open gate.
Corrected trainer, October 2026. Measured names and rings, illustrative motion.
Same brain, two worlds, one flaw.
Now go inside the creature's head. One brain lives in the real world and in the copy. A title card, five chapters, and an end card follow the clue from the senses, through the processing layer, into the memory spine, and out to the motor heads, where the fake world's reach falls short. The flaw enters at speed, the sense the broken velocity law corrupts. Teal rings settle on the memory cells where a readout finds it.
Names, rings and numbers are measured; the motion is illustrative. Every dot carries its real name from the observation spec. The twelve memory cells are real recurrent units of one brain from the corrected registered run, the seed whose survival reading sits closest to the ten-brain mean of 0.733. Ring strength is each unit's measured world-signal, from that brain's pools rebuilt from its saved weights (within 0.002 of the recorded readings). The numbers on screen are means over ten brains: an untrained brain reads 52% (0.523), a brain trained in the copy reads 73% (0.733), and a brain trained only to predict its senses, fed the same senses, reads 72%. The bar, 65%, was set in advance.
The memory chapter shows two different brains side by side, an untrained one and a trained one, each read once after training. Snapshots measured since, at 100, 200, 300 and 450 updates, show the reading growing and its direction reorganizing on the way (FINDINGS 23.4). Removing that direction while the creature forages changes what it does no more than a sham removal, so the readout finds the clue and the creature does not use it (FINDINGS 23.3). The end card carries the two caveats the record keeps. Without the second job, predicting what comes next, the reading falls under the bar at 300 training updates (0.613). At 450 updates the mean is 0.685, above 0.65, and its lower bound is 0.632, still under the bar. Two different samples of real worlds read 0.559 instead of 0.50, so the preregistered test as a whole is not passed (FINDINGS 17).
66 s · seven chapters · labels on screen, no narration, no audio track · corrected trainer · rings placed from measured data