Public evidence registry
Claims Registry
A stable-ID registry of ZeroModel claims, evidence states, boundaries, and governing sources.
ZeroModel’s public claims are governed first by the repository’s docs/claims-audit.md. The audit explicitly notes that its legacy matrix predates the current package split, so newer package claims are also reconciled from their owning README or stage-completion boundary.
This website does not silently promote those newer package claims into the central audit. Every record says where its authority comes from:
- Central audit — the claim already has an explicit evidence status in
docs/claims-audit.md. - Central audit + package contract — the audit status and the package boundary agree.
- Package claims boundary — implemented and documented by the owning package, but not yet comprehensively folded into the legacy audit matrix.
- Implementation + independent reproduction — a concrete executable result has additionally been reproduced outside the project’s own test path.
Stable IDs make claims addressable without copying their wording into every page. ZM-PER-STATE-001 already existed in the repository; this pass extends the convention across the current package surface with IDs such as ZM-CORE-LUA-001, ZM-ART-CLOSURE-001, ZM-TRUST-AUTH-001, ZM-NAV-HIER-001, ZM-SEARCH-REL-001, and ZM-CRITIC-LIN-001.
The governing rule remains:
Claim → implementation → test or measurement → result → boundary.
The registry below is generated from structured site data. It includes positive claims, package-level boundaries that still need central adjudication, implemented-but-unmeasured work, and measured negative results.
Finite policy property checking
Declared properties over bounded finite policy artifacts can be checked deterministically and counterexamples surfaced through the analysis package.
- Evidence
- PolicyPropertyChecker and PolicyPropertySpec operate over finite VPM policy artifacts.
- Boundary
- No continuous formal verification, semantic reasoning, causal detection, or benchmark adjudication beyond explicit deterministic contracts.
- Authority
- package README
Compiled-report aggregate closure
A complete adapted report, governing adapter contract, ScoreTable, LayoutRecipe, and VPMArtifact can be persisted and reloaded as one content-addressed aggregate and checked for semantic closure across representations.
- Evidence
- Aggregate validation checks refs, contract semantics, deterministic VPM reconstruction, coordinates, values, and source bindings before issuing a closure receipt.
- Boundary
- The central legacy claim matrix explicitly does not yet comprehensively fold this package in. Trust-signed compiled reports, cross-schema conversion, and automatic repair are excluded.
- Authority
- package README
Decision adjudication taxonomy
A bounded decision attempt can be classified as exact, action-equivalent, action-changing, or rejected while keeping state correctness separate from action correctness.
- Evidence
- Domain-neutral adjudicate_decision contract in zeromodel.core with invariant tests and provider-evaluation integration.
- Boundary
- The adjudication record is not a promotion, deployment, or safety-certification state machine.
- Authority
- core package contract
Dependency-free Lua policy export
A bounded policy can leave Python as a dependency-free Lua module while preserving policy artifact identity and separate consumer-plan identity.
- Evidence
- The 112-row arcade policy exported to a 21,235-byte Lua module and reproduced score 4 in 22 steps under Lua 5.4, matching the Python path.
- Boundary
- This establishes runtime portability, not embedded latency, RAM usage, or constrained-hardware performance.
- Authority
- implementation + independent reproduction
No-model bounded policy lookup
A bounded policy can be materialized into an addressable VPM artifact and read at runtime without invoking a model.
- Evidence
- VPMPolicyLookup, action argmax, deterministic replay, cell/source proof, and the bounded arcade policy example.
- Boundary
- Finite policy lookup does not establish open-world generalization.
- Authority
- central audit
Deterministic VPM artifact construction
A VPM is a deterministic spatial view over a table of scored items.
- Evidence
- ScoreTable, LayoutRecipe, VPMArtifact, deterministic artifact identity, binary-canonical identity bytes, and golden artifact-id tests.
- Boundary
- A deterministic artifact contract is not a claim of semantic reasoning or causal explanation.
- Authority
- central audit
Cell-to-source evidence mapping
Every VPM cell can map back to its source row, metric, raw value, and normalized value.
- Evidence
- VPMArtifact.cell() and explicit cell/source mapping tests.
- Boundary
- Traceability is not itself causal explanation.
- Authority
- central audit
Portable Tiny Critic readout
A declared numeric feature schema and fitted lightweight linear classifier can become an identified critic artifact with deterministic scoring, feature contributions, arithmetic-only portable inference, VPM ranking, and replay.
- Evidence
- Deterministic binary fitting, held-out batch scoring, contribution reconstruction, persisted evaluation evidence, portable payload, promotion recommendation, and receipt replay.
- Boundary
- Not general text quality, reasoning correctness, hallucination detection, semantic truth, or universal AI reliability.
- Authority
- central audit + package contract
Finite identified hierarchy traversal
A finite identified artifact hierarchy can be compiled, structurally validated, deterministically traversed, completely resolved, and replayed.
- Evidence
- HierarchyManifestDTO, NavigationTileDTO, declared traversal rules, traversal result/receipt, validate_hierarchy and replay_traversal.
- Boundary
- No planet-scale, infinite-capacity, logarithmic-time, semantic-search, nearest-neighbour, or 40-hop world-navigation claim.
- Authority
- package README
Pinned deployment binding
A visual address deployment can be pinned to one policy, calibration, source scope, and optional encoder identity.
- Evidence
- DeploymentBinding deterministic identity and exact contract verification.
- Boundary
- A DeploymentBinding is not a signature, authorization record, trust attestation, or persistence service.
- Authority
- central audit
Provider-neutral visual address seam
Different visual-address engines can share one provider-neutral contract while the addressed policy remains independently identified.
- Evidence
- ImageObservation, VisualAddressContract, VisualAddressDecision, VisualAddressProvider and conformance through the deterministic vision reader.
- Boundary
- The contract does not validate learned visual generalization.
- Authority
- central audit + package contract
Replayable local transition contradiction
A bounded perceptual action policy can record and replay why a local transition mismatch produced contradiction and replacement-lineage evidence.
- Evidence
- Declared observation schemas, comparison recipes, transition verification, contradiction artifacts, bounded repair proposals, deterministic ledgers, and replayable evidence stages.
- Boundary
- The package explicitly does not claim a general Observer architecture, habit formation, continual learning, universal state similarity, or open-world perception.
- Authority
- package README
Real fixed-camera runtime scaffold
A deterministic fixed-camera evidence compiler scaffold exists for real optical captures.
- Evidence
- The runtime compiler exists, but no empirical held-out real-camera capture result is committed.
- Boundary
- Implementation is not evidence that real optical perception works.
- Authority
- central audit
Deterministic panel pixels to typed field evidence
Canonical status-panel pixels can be compiled into typed field evidence for declared visual channels.
- Evidence
- PanelObservationEvidenceCompiler with panel-anchor validation, bounded registration, declared field measurements, canonical fixtures, and synthetic occlusion tests.
- Boundary
- No claim yet for real-camera glare, blur, viewpoint variation, or held-out optical capture performance.
- Authority
- central audit
Implicit PNG interventions improve provider reliability
The tested implicit PNG interventions improve the fixed local provider over the unlabelled representation.
- Evidence
- The controlled smoke experiment found no targeted improvement; two interventions increased action-changing errors and one produced an additional rejection.
- Boundary
- Unsupported for the recorded model blob, prompt, parser, policy, fixture, and smoke protocol; not generalized to other providers or domains.
- Authority
- central audit
Typed evidence to policy-compatible state set
Given identified observation status, a finite state specification, typed field evidence, and a bounded policy, ZeroModel can derive every compatible state, execute when all compatible states are action-equivalent, and reject action-changing ambiguity.
- Evidence
- FieldEvidence -> StateClaimSet -> PolicyCompatibilityReport with tests for equivalent ambiguity, changing ambiguity, and invalid-observation short-circuiting.
- Boundary
- This validates evidence-to-decision compatibility, not camera-image-to-evidence perception.
- Authority
- central audit
Relation-specific deterministic search readout
Declared relation-specific readouts over frozen representations can be compiled, used for exact relation ranking, inspected as a VPM, and replayed from a receipt.
- Evidence
- Deterministic ridge projections, aggregate-closed artifacts, exact ordering, VPM result inspection, and receipt replay.
- Boundary
- No universal semantic search, automatic relation discovery, scalable candidate filtering, or superiority over cosine.
- Authority
- central audit + package contract
DTO-only SQLite persistence adapter
The video action-set domain can be persisted through a SQLAlchemy/SQLite Store while keeping ORM rows internal and the public Store boundary DTO-only.
- Evidence
- Tests cover explicit schema creation, foreign keys, memory parity, rollback, MatrixBlob deduplication/conflicts, reopen behaviour, tamper detection, and runtime composition.
- Boundary
- No distributed coordination, high availability, or production authorization claim.
- Authority
- package README
Bounded artifact authorization verification
An identified artifact can be checked for integrity, signature, trust-policy acceptance, declared deployment authorization, freshness, rollback epoch, and revocation under a bounded trust policy.
- Evidence
- Ed25519 verification through standard cryptography primitives with separate integrity/authenticity/trust/authorization/freshness decisions.
- Boundary
- No secure hardware, remote attestation, PKI deployment, regulatory certification, supply-chain closure, or tamper-proof storage claim.
- Authority
- package README
Immutable provider-evaluation aggregates
Bounded provider-evaluation runs can be persisted and reloaded while separating exact-state correctness from policy-action correctness.
- Evidence
- Provider configuration, observations, expected/predicted states, policy decisions, response evidence, exact/action-equivalent/action-changing/rejected summaries, tamper checks, rollback, and memory/SQL parity.
- Boundary
- Evidence accounting does not establish provider accuracy, visual understanding, calibration, or safe deployment.
- Authority
- central audit
Controlled representation comparison
A real-provider PNG-representation experiment can hold provider, prompt, parser, policy, fixture, and inference settings fixed while changing only an identified representation recipe.
- Evidence
- Six persisted provider runs with content-addressed PNGs, operation provenance, case outcomes, summaries, and compiled-report identities.
- Boundary
- The experiment machinery is validated; representation improvement is not.
- Authority
- central audit
Visual policy addressing without symbolic state ID
A bounded visual observation can address an identified finite policy through a deterministic observation codebook under a declared acceptance profile.
- Evidence
- 112/112 canonical row recoveries, 112/112 canonical action equivalence, 2,401/2,401 four-target waves, and 31,213 visual decisions in committed hardening evidence.
- Boundary
- Not natural-image robustness, open-world perception, object recognition, or general image understanding.
- Authority
- central audit
Frozen DINOv2 global retrieval improvement
Frozen DINOv2 global retrieval materially improves governed policy addressing over normalized pixels.
- Evidence
- The recorded global CLS systems failed to establish the claim and exhibited high false-accept and conflicting-action error rates.
- Boundary
- Refuted for the recorded fixture, global CLS representation, preprocessing, calibration, and operating point only.
- Authority
- central audit
Frame-local discriminative evidence reader
Frame-local discriminative evidence yields a safe governed visual reader.
- Evidence
- The committed selection records no_safe_architecture and no_feasible_operating_point for the tested architecture family and protocol.
- Boundary
- A negative result for the recorded fixture and architecture family, not proof that every local method fails.
- Authority
- central audit
Registration plus locality as a safe reader
Registration and local comparison already provide a useful governed visual reader.
- Evidence
- Raw ranking improved to 1176/1344 exact-row and 1323/1344 action accuracy, while final governed benign coverage remained 0/1344.
- Boundary
- Registration repaired part of the ranking failure but did not create a useful accepted operating point.
- Authority
- central audit
Normalized pixels as a safe approximate reader
Normalized pixels provide a useful safe approximate visual-address reader.
- Evidence
- Raw top-1 reached 1008/1344 exact-row and 1302/1344 action accuracy, but the corrected governed operating point produced 0/1344 final benign coverage and 1344/1344 FRR.
- Boundary
- Useful ranking signal did not produce a useful zero-distinguishable-FAR accepted operating region on the recorded fixture.
- Authority
- central audit
Rejection-equipped ridge linear probe
A rejection-equipped ridge linear probe can match retrieval at lower complexity.
- Evidence
- Recorded benign action accuracy 59.45%, exact benign row accuracy 28.79%, 100% FAR, and 521 conflicting-action errors.
- Boundary
- Refuted for the recorded fixture and operating point.
- Authority
- central audit