ARTI / RESEARCH STATUS

Every research claim carries its evidence.

This matrix separates public contracts, experimental results, architectural interpretations, and open hypotheses. Status is grounded in ARTI 1.7.0 at 8e734d6; new evidence must update both limits and the verified commit.

CONTRACTRESULTINTERPRETATIONHYPOTHESISPLAN

CLAIM BOUNDARY

This is not a marketing feature list.

Contracts can be proven by tests; task capabilities require training and controls. Implementing a mechanism does not automatically prove the research interpretation around it.

ClaimKindStatusCurrent evidenceBoundary

Pulse forms a fixed K workspace

CONTRACTSUPPORTED 1.X

Public shape, mask, gradient, and device tests.

Fixed capacity does not guarantee task information survives.

Inspect basis →

UnFold preserves original input values

CONTRACTSUPPORTED 1.X

Every original input index appears exactly once in the hard gather.

It is not the inverse of Fold and cannot restore lost information.

Inspect basis →

FusionPulse combines compact workspaces

CONTRACTALPHA

Joint context, feature-wise survival, and a fixed output K.

Shared-object or manifold recovery remains an interpretation.

Inspect basis →

Runtime vocabularies use local output slots

CONTRACTALPHA

Call-time vocabulary tensors and local-slot APIs.

Permutation consistency requires a model trained for the contract.

Inspect basis →

Observer frames can be inverted explicitly

CONTRACTEXPERIMENTAL INTEGRATION

Paired rotations and a caller-supplied operator bank.

ARTI does not discover sources or verify physical correctness.

Inspect basis →

PhaseMixer mixes feature operators by state and coordinate

CONTRACTEXPERIMENTAL

Public ARTIPhaseMixer, operator weights, and receptor gain.

It is not observer inverse, sensor fusion, or physical phase recovery.

Inspect basis →

Membrane separates context, visibility, and emission

CONTRACTALPHA

Public/inner streams and caller-defined readability.

It is not encryption, a security boundary, or a secret reasoning channel.

Inspect basis →

Task representations remain stable across segmentation

HYPOTHESISTRAINING GATE

Requires train-A/test-B, random re-partitioning, and pooling baselines.

Not a product claim until the gate passes.

Inspect basis →

FusionPulse recovers shared object structure

INTERPRETATIONOPEN RESEARCH

Requires leave-one-view-out, neighborhood, and task ablations.

The tensor contract alone cannot prove this interpretation.

Inspect basis →

Membrane improves prompt-injection robustness

HYPOTHESISNO PUBLIC CLAIM

Requires attack success, normal-task utility, and permission-leakage benchmarks.

The site does not present this as a security capability.

Inspect basis →

VERIFIED SNAPSHOT

ARTI 1.7.0 · 8e734d6

Stable Candidate is not LTS. Every Alpha, Experimental, and Hypothesis label keeps its own boundary.

Understand the conceptGitHub