ItaSoRL

A tabula-rasa artificial-life system that asks whether a from-scratch digital organism can tell that its world is a generative surrogate rather than the authentic base dynamics of its computational substrate, using only substrate seams, with detection read out, not rewarded.

Start at the website: ilevytate.github.io/ItaSoRL - the whole finding on one page, including the 90-second film, the headline numbers (99% / 50% / 73%), and links to everything below.

Google Colab (GPU): Open notebooks/colab_gpu.ipynb in Colab. Make your own copy first (File -> Save a copy in Drive) so your run and outputs are saved - the notebook's first cell enforces this. Then enable a GPU runtime and run all cells (clones the repo and runs python scripts/run_e2e.py).

New to the project? Start with the nine-page illustrated series - the same walkthrough in two voices, every number read from the committed artifacts:

• Plain-English series (PDF) - the story: the question, the world, the trick, what we found.
• Research edition (PDF) - the same pages in research terms: design, control battery, pre-registered results, open questions.

Drop a learning creature, built from nothing, into a world that is secretly either the real thing or a flawed copy of itself. Can the difference be noticed from the seams alone? And does the creature come to represent that difference on its own, without ever being paid to?

(Read this part if you want the idea without the jargon.)

Imagine raising a digital creature inside a simulated world. The creature starts with no built-in knowledge, no labels, no instructions. It just lives there, senses its surroundings, and learns to predict what happens next.

Now play a trick on it. Sometimes the world it lives in runs on the "true" physics of the computer underneath. Other times we swap in a slightly fake version: a copy that behaves almost the same, but carries tiny tell-tale flaws, the kind of seams you would only get from a stand-in rather than the real thing. Think of it as a very subtle "is this the Matrix?" test for a tiny mind.

We are chasing two separate questions, and keeping them strictly apart matters:

1. Can the fakeness even be detected? If an all-seeing outside observer inspects the world's behavior, can it spot the copy? (Short answer so far: yes, almost perfectly.)
2. Does the creature notice on its own? Here is the twist that makes this research, not a gadget: we never reward the creature for spotting the fake. We never even tell it the fake exists. We only ever read out whether the knowledge ended up inside its head as a side effect of ordinary learning.

What we found is the interesting part. The fakeness is trivially detectable from the outside, but the creature does not bother to learn it. An outside observer catches the copy nearly every time; the creature's own internal state barely registers it at all. Nothing in its day-to-day task required it to care about the difference, so it simply did not represent it.

That gap, between what is detectable and what a mind spontaneously encodes, is the real object of study. It turns a vague headline ("can an organism notice its world is fake?") into a sharper, testable question: under what conditions does a mind start to represent something it was never asked to care about? The next phase (making survival depend on the difference) is built to answer exactly that.

No - and the difference is easy to state. Other projects have gotten close, but each one skipped the hard part:

1. Chatbots that know when they're being tested. Today's AI chatbots can often tell when they're being evaluated versus talking to a real person. But those AIs read most of the internet - they already know what "a test" is, what "a simulation" is, what researchers do. That's like a student who read the teacher's answer key. Our creature has read nothing. It is born knowing zero - no language, no concepts, no hints. If it figures out something is off about its world, it worked that out from the world itself.
2. AI brains that learn things nobody taught them. Scientists have looked inside AI systems before and found knowledge that was never explicitly taught - like a game-playing AI that secretly worked out the layout of the board. But those discoveries were always about things inside the AI's world - the pieces, the score, the map. Nobody had checked for the bigger question: does it know whether the world itself is the real one or a copy?
3. Robots that check whether practice matched reality. Robots trained in simulators are sometimes given alarm systems that go off when the real world doesn't match their practice world. But someone built that alarm and told it what to watch for. That's an installed smoke detector. We never install anything. We just let the creature live, then peek inside its memory afterward to see if the knowledge showed up on its own.

And the real headline is the comparison nobody thought to make: what is plainly detectable in a world versus what a mind living in that world actually comes to know. In one sentence: everyone before either told their AI to look for the fake, or used an AI that already knew what "fake" means - we raised a mind from nothing, hid the flaw in the physics, never mentioned it, and then checked whether the knowledge showed up anyway. (The research-grade version of this comparison, with citations, is in docs/ITASORL.md.)

Take an agent with no imported semantics, drop it into a controllable world, and make that world either an authentic simulation or a subtly-flawed surrogate of itself. Can the distinction be detected from the substrate's seams alone? And, the part that matters, does an agent encode that distinction incidentally, without ever being rewarded for it?

The project rests on three invariants:

• Tabula rasa: no semantics are imported; the agent starts from nothing.
• Readout, not reward: world identity is probed, never trained or rewarded.
• Substrate-grounded: the only tells are artifacts of the computational substrate.

Detection difficulty is dialed with a surrogate ladder (L0 identical control, L1 discretization, L2 rollout drift, L3 learned-model fingerprint, L4 adversarial).

What is new here. Adjacent literatures each hold one piece: LLM evaluation awareness finds test-vs-deploy signals in pretrained (semantically contaminated) models; probing studies (AtariARI, Othello-GPT) read out incidental within-world variables of a single fixed world; agent-side mismatch detection (GalilAI, RAPT) makes noticing the change an explicit objective. 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 against the nearest prior work is on record in docs/ITASORL.md.

• H1 (detectability): the surrogate is distinguishable from the authentic world by a privileged discriminator.
• H2 (substrate-grounding): the signal lives in substrate artifacts, confirmed via ablations — confirmed at L3; L1 organism encoding is negative at the matched in-band grid (FINDINGS §14–14.7).
• H3 (emergence under selection): representing the distinction emerges when survival depends on it — resolved negative (FINDINGS §13.D).
• H4 (legibility / incidental encoding): a from-scratch agent encodes the distinction incidentally, without reward — positive at L3, negative at L1/L2.

For hand-authored substrate seams (L1 discretization, L2 rollout drift), detectability does not imply incidental encoding. An L2 artifact an external oracle detects at AUROC ≈ 0.99 leaves essentially no decodable trace in a from-scratch agent's representation (≈ 0.5). An in-band L1 quantization grid (oracle AUROC 0.873) likewise leaves the survival state at 0.533 (n = 10; FINDINGS §14.7).

But the picture changes at L3. When the surrogate is a learned-dynamics fingerprint (a small net replacing the velocity law) rather than a hand-tuned knob, the survival agent, and at this fingerprint only the survival agent, encodes it. At a difficulty where an untrained net is at chance (about 0.49, and still only 0.52 under a nonlinear probe) and a prediction-only agent is near chance (about 0.57), the survival probe reads 0.752 (n = 10, honest t-based 90% CI [0.698, 0.807], which excludes the pre-registered 0.65 bar; 8 of 10 seeds clear it). The dissociation is robust: it is not reward-mediated (world is not decodable from summed reward, AUROC 0.541, clean 10 of 10 seeds), not survivorship-biased (0 early deaths, every pool 110/110), not a linear-probe artifact (the untrained net stays near chance even nonlinearly), and the L0 authentic-vs-authentic control is at chance (0.517).

And the signal is not just behavior. The agent does move and forage differently in the two worlds - in fact the full behavior trace alone decodes the world at 0.803, better than the state probe itself - so the obvious deflationary reading was that the probe reads behavior, not a representation. A pre-registered per-timestep control (dump every step's speed/energy/food/drag, residualize the recurrent state on the behavior trace in-fold, probe what is left) rejects that reading: the behavior-independent world-signal is 0.726 (t-based 90% CI [0.679, 0.772], which excludes the 0.65 bar; the seed-level bootstrap interval [0.685, 0.765] agrees; 9 of 10 seeds clear it; quadratic variant 0.721). Strengthening that control to also residualize absolute position and heading (the covariate a differing velocity law could otherwise smuggle in) barely moves the signal, to 0.723 (t-based 90% CI [0.676, 0.769]; 8 of 10 seeds), closing the covariate gap in the headline's favor (FINDINGS §10.4.1). The control is honest on its own negative controls: the untrained agent's state reads exact chance (0.498) under the same control even though untrained behavior decodes 0.645, and the prediction-only agent stays near chance (0.574). (An earlier, cruder per-episode-mean control had under-estimated the signal at ~0.66 by over-removing - the attenuation our synthetic tests predicted.) The honest statement is: reward- and survivorship-controlled, robust to nonlinear probing, with a behavior-independent world-signal of about 0.73 that clears the pre-registered bar. This is the first place "detectable does not imply learned" reverses: a from-scratch agent, never rewarded for it, comes to carry world-discriminative state as a byproduct of surviving. The mediation audit is reproducible code (scripts/audit_behavior_mediation.py; artifacts in artifacts/expB2/).

A pre-registered replication at a second calibrated capacity sharpens the claim. The second in-band fingerprint (hidden = 7, selected by a frozen fallback rule after hidden = 4 failed its gates) passes every gate and replicates the behavior-independent world-signal almost exactly: 0.722 (t-based 90% CI [0.672, 0.773]) vs 0.726 at hidden = 8. But that coarser artifact is one every trained agent picks up (predictor 0.714 vs survival 0.737, under the pre-registered +0.05 dissociation requirement), so the survival-only part of the claim is conditional on the subtler hidden = 8 artifact. What survives both capacities is a reward-clean, survivorship-clean, behavior-independent world-signal of about 0.72 in the survival agent's state.

A held-out probe (n = 10, committed per-seed summary in artifacts/expB2/heldout_l3_h8_summary.json) then sharpens what that signal is. It splits: the world-identity direction learned against the trained fingerprint still reads a held-out capacity variant of it (transfer 0.773 vs untrained floor 0.569; the pre-registered rule passes) - note the variant shares the training recipe, seed, and data, so this certifies robustness within one recipe (FINDINGS §10.6 scope note); the across-recipe generalization claim is carried by the cross-recipe probe below. A frozen reverse probe (train on the coarser fingerprint, read the subtler one; per-seed summary in artifacts/expB2/heldout_l3_h7_reverse_summary.json) fails its bar at 0.638, so this transfer is direction-dependent: it generalizes from subtle training artifacts, not bidirectionally. And under a common-garden control that runs both groups through an identical tail after differing prefixes, tail-only state still recovers the prefix world above the frozen bar on both directions (0.666 forward, 0.684 reverse; both clauses pass), though the last-8-step late tail decays toward chance (0.586/0.577). So the L3 world-signal is best read as a modest persistent stored world-identity component the survival policy also expresses reactively: it clears the common-garden bar but only just, and its late tail fades, so it is persistent-but-weak, not strongly stored. (The original common-garden numbers, 0.557 below the bar, were scored with a since-fixed biased estimator and have now been re-scored, overturning the reactive reading; see FINDINGS §10.6.1. The transfer numbers were unaffected.)

A cross-recipe probe (n = 10, artifacts/l3_crossrecipe/summary.json) then extends the transfer half across surrogate families: the same direction reads a gate-calibrated random-Fourier-features ridge law the agent never lived with (0.684 vs untrained floor 0.548; the pre-registered rule passes, machine-checked; a thin pass - the t-based 90% CI lower bound clears the bar by 0.004 and 7/10 seeds sit above it), so the world-signal is recipe-general, not a signature of one function class. See docs/FINDINGS.md and docs/PREREGISTRATION_L3.md.

.
|-- README.md                   this file: the map
|-- LICENSE
|-- CITATION.cff                citation metadata (name, ORCID, version)
|-- pyproject.toml              package metadata (pip install -e .)
|-- requirements.txt            runtime dependencies
|-- requirements-dev.txt        pytest + dev tooling
|-- itasorl/                    core library (world, agents, experiments)
|   |-- world.py                World protocol, surrogate ladder, matched-pair harness
|   |-- patch_of_earth.py       PatchOfEarthV0 concrete world, incl. L1/L2 hooks
|   |-- agent.py                recurrent world model (RSSM-lite)
|   |-- agent_ac.py             survival actor-critic (Experiment B-v2)
|   |-- experiment_a.py         agent-free L1 detectability oracle
|   |-- experiment_a_l2.py      agent-free L2 detectability oracle
|   |-- experiment_a_l3.py      agent-free L3 (learned-fingerprint) oracle
|   |-- experiment_b.py         incidental-detection harness
|   |-- experiment_b2.py        survival-coupled B-v2 pipeline
|   |-- surrogate_l3.py         L3 learned-dynamics surrogate (G_motion)
|   |-- behavior_audit.py       behavior-mediation controls (residual probes)
|   |-- stats.py                TOST/ROPE equivalence, bootstrap AUROC CIs
|   `-- results_io.py           end-to-end run recording
|-- scripts/                    deterministic reproduction runners
|   |-- run_e2e.py              pytest + all experiments (recorded)
|   |-- run_expA.py ...         Experiment A/B runners
|   |-- run_expB2.py            Experiment B-v2 / L3 (GPU if available)
|   `-- audit_behavior_mediation.py  behavior-mediation audit on dumped states
|-- docs/
|   |-- ITASORL.md              research plan
|   |-- ITASORL_world_spec.md   world specification ("A Patch of Earth" v0)
|   |-- FINDINGS.md             empirical results
|   |-- LEARNING.md             running lab notebook / lessons log
|   |-- PAPER_OUTLINE.md        writeup outline + claims inventory
|   |-- PREREGISTRATION.md      B-v2 pre-registration
|   |-- PREREGISTRATION_Bv3.md  B-v3 pre-registration
|   |-- PREREGISTRATION_L3.md   L3 pre-registration + deviation log
|   |-- PREREGISTRATION_C.md    Experiment C pre-registration (design-complete)
|   |-- AUDIT_2026-07.md        research-integrity audit
|   |-- itasorl-series-plain-english.pdf  illustrated walkthrough (plain English)
|   |-- itasorl-series-research.pdf       illustrated walkthrough (research edition)
|   `-- figures/                result figures (.png) + provenance README
|-- artifacts/                  published summaries (committed): expA/, expB/, expB2/, expC/, l3_crossrecipe/
|-- fullruns/                   e2e run bundles (gitignored; default output)
|-- results/LATEST_RUN.txt      pointer to latest fullruns folder
|-- notebooks/colab_gpu.ipynb   Colab end-to-end runner
`-- tests/                      pytest regression suite

• docs/itasorl-series-plain-english.pdf: the illustrated walkthrough in plain English - the friendliest entry point to the whole project.
• docs/itasorl-series-research.pdf: the same series in research terms - design, control battery, pre-registered results, open questions.
• docs/ITASORL.md: the research plan, core question, literature white-space, hypotheses (H1 to H4), experiments (A/B/C), the surrogate ladder, validity audit, statistics, and engineering architecture.
• docs/ITASORL_world_spec.md: the world specification, "A Patch of Earth" v0, the 2.5D representation, fields and forcing, dynamics, ecology, the ~146-dim observation, ladder attachment, and confound management.
• docs/FINDINGS.md: empirical results from the first build-and-test cycle.
• docs/PAPER_OUTLINE.md: the writeup outline and a claims inventory linking every headline number to its committed artifact.
• docs/LEARNING.md: the running lab notebook (methods lessons, dead ends, decisions).
• docs/PREREGISTRATION_L3.md, PREREGISTRATION_Bv3.md, PREREGISTRATION.md, PREREGISTRATION_C.md: pre-registrations (with deviation logs) for the B-v2, B-v3, L3, and (design-complete) C experiments.
• docs/AUDIT_2026-07.md: a skeptical research-integrity audit (numbers, statistics, pre-registration timing, citations).

Result figures live in docs/figures/ and are regenerated by the run scripts:

• expA_ceiling.png: L1 detectability ceiling vs grid spacing.
• expA_L2_ceiling.png: L2 detectability ceiling vs drift strength.
• expB_incidental.png: recurrent-state probe across the drift sweep (target vs negative control).
• expB_channels.png: two incidental-detection channels (recurrent state vs prediction error).
• expB_kstep.png: effect of a longer-horizon objective on encoding.

To regenerate all of them in one recorded pass, run python scripts/run_e2e.py --quick from the repo root; each scripts/run_exp*.py runner rewrites only its own figure. Per-figure provenance (which script, which doc section) is tracked in docs/figures/README.md.

Dependencies:

pip install -r requirements.txt      # runtime deps
pip install -e .                     # or: install itasorl as an editable package

Reproduce the experiments (each is deterministic given its seeds; run from the repo root so figures land in docs/figures/):

python scripts/run_e2e.py --quick   # full battery + recorded results (recommended)
python scripts/run_expA.py          # Experiment A, L1
python scripts/run_expA_l2.py       # Experiment A, L2
python scripts/run_expB_full.py     # Experiment B: recurrent-state probe
python scripts/run_expB_surprise.py # Experiment B: prediction-error channel
python scripts/run_expB_kstep.py    # Experiment B: open-loop horizons
python scripts/run_expB_gap.py      # Experiment B: engagement + delta objective
python scripts/run_expB_nonlinear.py# Experiment B: nonlinear probe check
python scripts/run_expB2.py         # Experiment B-v2: survival-coupled

A quick smoke test of the Experiment B pipeline:

python -m itasorl.experiment_b

Run the test suite (pytest ships in the dev requirements):

pip install -r requirements-dev.txt
pytest -q

Google Colab (GPU): Open in Colab (same notebook as the badge at the top). Make your own copy first (File -> Save a copy in Drive; the first cell enforces this), enable a GPU runtime, then run all cells.

Local Jupyter / VS Code: open notebooks/colab_gpu.ipynb from this repo; it auto-detects local mode (no Drive/download cells).

1. README.md (this file): the map.
2. docs/FINDINGS.md: what we found and what it means.
3. docs/ITASORL.md: the full research plan and rationale.
4. docs/ITASORL_world_spec.md: the world, in detail.

Citation metadata lives in CITATION.cff; GitHub renders a "Cite this repository" button from it, or run cffconvert -f bibtex for BibTeX.

The hard part of "can an organism notice its world is fake?" turned out not to be detectability: substrate artifacts are readily detectable by a privileged observer, with calibrated difficulty and a clean confound audit. The hard part is incidental encoding. A from-scratch agent trained only to predict its senses does not represent a dynamics artifact it is never asked to care about, even when that artifact is strongly present and trivially detectable from the outside. That gap, between what is detectable and what an agent spontaneously encodes, is the real object of study, and it turns the headline hypothesis into a sharper, conditional question that the next phase (survival-coupled reward, richer ladder rungs) is designed to answer.