Evidence before ambition

Right action. Wrong state. We measure the difference.

ZeroModel builds identified decision artifacts and the evidence needed to challenge them. It separates exact state recovery from action-equivalent error, preserves negative results, and can carry a bounded policy out of Python into a dependency-free Lua runtime.

Explore the claims Reproduce the evidence View the code

Validated · bounded

96.9%

Right action.

Normalized-pixel addressing produced 1,302 / 1,344 raw top-1 action-correct decisions on the bounded arcade fixture.

Measured · unsupported

75.0%

Correct state.

The same run recovered only 1,008 / 1,344 exact rows. A correct action can hide a wrong understanding of state.

Decision adjudication

Correct is not one outcome.

ExactCorrect state, correct action.

Action-equivalentWrong state, same action.

Action-changingWrong state, different action.

RejectedThe system refuses to decide.

Cross-runtime proof

The policy can leave Python.

A 112-row bounded policy can be exported as a dependency-free Lua module, executed in Lua 5.4, and preserve both policy artifact identity and consumer-plan identity across the runtime boundary.

reproduced result

policy rows        112
generated module   21,235 bytes
python score       4
lua score          4
python steps       22
lua steps          22
runtime            Lua 5.4
identity           artifact_id + plan_id preserved

Run the Lua proof Inspect the exporter

Bounded visual addressing

31,213 visual decisions. No symbolic runtime state ID.

112 / 112canonical row recoveries

112 / 112canonical action equivalence

2,401 / 2,401four-target waves

31,213visual decisions

Boundary: this validates an identified finite observation codebook under declared acceptance contracts. It is not open-world vision, natural-image robustness, or general image understanding.

Negative evidence

Five approaches failed. We kept the results.

Normalized pixels, global DINOv2 retrieval, ridge probes, registration plus locality, and discriminative local evidence all failed to establish a useful governed visual reader under their recorded test conditions. Controlled PNG interventions against a fixed local model also failed their promotion gate.

See what failed

Policy-safe ambiguity

You do not always need to know the exact state.

Validated · bounded

Ambiguous state. Same action.

ZeroModel can derive every state compatible with typed field evidence. If all compatible states imply the same policy action, it can execute that common action while preserving the ambiguity in the evidence.

Validated · bounded

Ambiguous state. Different actions.

If compatible states imply different actions, the same machinery rejects the decision and records the unresolved fields rather than inventing certainty.

Beyond the arcade

Identified decisions over other bounded domains.

FX triangular arbitrage

Identified bid/ask snapshots, costs, freshness checks, candidate cycles, rejection paths, mutation evidence, and replayable calculation traces.

Boundary: deterministic offline evidence, not a profitability or live-trading claim.

Tiny Critic

A declared-feature binary linear critic can be fit, evaluated on held-out synthetic fixtures, exported as an arithmetic-only payload, replayed, and rendered as a VPM ranking.

Boundary: not general text quality, truth detection, or hallucination detection.

Relation search

Declared relation-specific readouts can compile deterministic synthetic search fixtures with exact ranking, identified receipts, replay, and VPM inspection.

Boundary: not universal semantic search or superiority over cosine.

Public evidence surface

Every strong claim should have a status, implementation, evidence, and boundary.

The website is being rebuilt around the repository's claims audit rather than a separate marketing narrative. Validated claims stay bounded. Unsupported and refuted claims remain visible. Newer package claims are being reconciled into the same registry.

Open the claims registry Open the failure register Read the claim boundary