Reproduce

Don’t trust the website. Run it.

The strongest ZeroModel claims should lead to executable code, tests, or committed evidence packages. This page starts with the smallest cross-runtime proof and then points into the repository for deeper experiments.

Lua cross-runtime policy proof

The repository includes a bounded arcade policy that can be materialized in Python and exported as a dependency-free Lua module.

From a ZeroModel checkout:

python examples/lua_edge_policy.py
lua examples/lua/run_arcade_policy.lua

The independently reproduced run reported:

policy rows        112
generated module   21,235 bytes
python score       4
lua score          4
python steps       22
lua steps          22
runtime            Lua 5.4

The generated consumer carries the policy artifact identity and a separate consumer-plan identity.

What this demonstrates: bounded policy runtime portability and identity preservation across Python and Lua.

What it does not demonstrate: embedded-device latency, memory use, tiny-hardware performance, or cloud/edge symmetry.

Symbolic policy lookup

Run the bounded arcade policy example:

python examples/arcade_shooter_policy.py

This demonstrates deterministic policy lookup over the finite state table and provides the baseline behaviour later visual-address experiments must preserve.

Visual sign reader evidence

The visual reader’s current strongest bounded evidence is committed under:

docs/results/visual-sign-reader-hardening/
docs/results/visual-sign-reader-contract-hardening/

The recorded result reproduces:

112 / 112 canonical row recoveries
112 / 112 canonical action equivalence
2,401 / 2,401 four-target waves
31,213 visual decisions

The contract-hardening work separates canonical-only, exact-codeword, calibrated-nearest, and evidence-only acceptance semantics rather than treating every visual match as the same claim.

Controlled provider representation experiment

The fixed-provider PNG experiment is preserved under:

docs/results/controlled-png-representation-v1/

Its importance is methodological: provider configuration, prompt, parser, policy, fixture, and inference settings were held fixed while only the identified representation recipe changed. The tested implicit interventions did not pass the promotion gate.

Fixed-camera state claims

The current fixed-camera path is split deliberately:

canonical panel pixels
  → typed field evidence
  → compatible finite states
  → policy compatibility
  → execute common action or reject ambiguity

The deterministic compiler and canonical fixtures are executable today. Real optical capture performance remains a separate unmeasured boundary.

See:

Demo catalogue

ZeroModel also maintains an executable demo catalogue covering VPM artifacts, visual sign reading, Lua edge policy export, signs-not-directions traces, criticality verification, policy lookup benchmarking, Bertin pattern detection, FX triangular arbitrage, and Tiny Critic evidence.

Browse the repository demo catalogue

If a website claim cannot eventually point to something runnable, inspectable, or explicitly marked unmeasured, it should not be promoted as evidence.