A personal research experiment · Solo research and engineering
Petri: a deterministic simulation trap with a hard validation protocol
Test whether the precursors of a complex ability can appear under ecological pressure alone, with claims drawn from internal probes, not behavior.
- determinism
- bit-exact across 2 OS processes
- throughput
- ~36,400 turns/sec
- speed margin
- ~218x under the <60s bar
- survival effect
- Cliff's delta 0.87 to 1.0, p < 2e-4 (5 seeds)
- NumPy (vectorized)
- PCG64 single-RNG
- Ablation gates
- Bit-exact replay
Problem
Petri is a personal research experiment: a deterministic simulation built to ask whether the precursors of a complex ability can appear on their own under ecological pressure, without being hand-coded in. The discipline that makes it interesting is where the claims come from. They are derived from internal-representation probes, not from behavior that merely looks right from the outside.
Constraint
Two constitutional clamps applied at once.
No reward shaping. The only internal signal available to an agent is prediction error plus homeostasis, so you cannot nudge the agent toward the outcome you want. And bit-exact determinism: every source of randomness in the world has to flow from a single RNG.
The wall was where these collided. Under a regime of real death and pure online learning, a lifespan is around 60 turns while competence wants roughly 45,000 turns of continuous learning. That is a learning gap of about 750x, a regime that is mathematically unlearnable if you take the constraints seriously rather than quietly relaxing them.
Architecture decision
The decision was to refuse reward shaping entirely and encode "physics is the referee" into the system.
The world is a vectorized NumPy simulation with single-RNG (PCG64) determinism, so a run is fully reproducible from a seed. On top of that sits a mandatory validation protocol: every positive finding has to pass ablation, probe-leakage, and bit-exact replay gates before it is allowed into a report. The sentence "it behaved as if it understood" cannot enter a report. Either a probe survives ablation and replays bit-for-bit, or the claim does not exist.
Result
Determinism and speed held with margin. Ten thousand turns run in two separate OS processes produced the same final hash, bit-for-bit, and the simulation runs at about 36,400 turns per second, clearing a sub-60-second bar by roughly 218x.
The measured effect that survived the protocol was a survival advantage: against a random policy, across five seeds, Cliff's delta ran 0.87 to 1.0 at p < 2e-4, consistent rather than noise. The point of Petri is not a headline claim. It is an apparatus where a claim has to earn its way past ablation and bit-exact replay before anyone is allowed to believe it.