window.huggingface={variables:{"SPACE_CREATOR_USER_ID":"671a4041847919198a464328"}};d> Bounded State-Space Actuation Note v0, 2026-06-16 | Mian Zhang
Mian Zhang

Public Release - 2026-06-16

A model may propose. A bounded system decides whether action is allowed.

Artifact: Bounded State-Space Actuation Note v0. This public note turns model uncertainty into reviewable fields: state envelope, entropy drift, baseline divergence, deterministic fallback, receipt, and challenge route.

state_envelope entropy_drift baseline_divergence fallback_gate audit_receipt
Papers Evidence Map Concepts Counterexamples Registries
System boundary blueprint visual for bounded state-space actuation.

Artifact

Probability is not action permission.

This update records a control boundary for AI systems whose outputs may route into high-risk action. A fluent answer, a high-confidence string, or a large model does not by itself grant action authority.

The public object is the boundary shape: what is bounded, what is measured, when the system falls back, what receipt is kept, and what would challenge the claim.

1

State envelope

The permitted operating region for the current task, asset, route, or action class.

2

Action vector

The proposed action representation that must remain inside the permitted envelope.

3

Entropy signal

A public field for uncertainty or action spread, not a standalone proof of risk.

4

Baseline divergence

A measured distance from a deterministic or policy-defined baseline.

5

Fallback gate

The route to no-action, review, narrower scope, or deterministic execution.

6

Audit receipt

A record of the boundary state, trigger, fallback decision, and unresolved evidence.

Control Boundary

When uncertainty rises, the system should narrow authority.

A bounded actuation layer can treat model output as a proposal, not a command. If the action distribution becomes too diffuse, or the proposed action diverges from the baseline beyond a declared threshold, the public state should move toward no-action, human review, or a narrower deterministic route.

The relevant question is not whether the model can produce a persuasive explanation. The relevant question is whether the action boundary, fallback decision, and receipt are inspectable.

Concept visual of an entropy audit log and fallback trigger.

Audit

A stop decision should leave a receipt.

The public claim is not that a specific private runtime, latency, threshold, or deployment has been externally verified here. The claim is that high-risk AI action should expose a reviewable stop shape: trigger, boundary, fallback, receipt, and unresolved state.

Private logs, customer records, credentials, live execution traces, and account material remain outside the public layer.

Naming Boundary

Do not mix this control note with the paper index card.

The current public paper card labeled P07 remains WisdomBench-Embodied LIBERO / VLA, with its own DOI record and paper-specific boundary.

This 2026-06-16 note is a protocol-facing control boundary. Cite the paper DOI for paper-specific claims; cite this note only for the public actuation-boundary shape.

Matrix

Bounded actuation fields.

FieldRequired valueFailure modeRepair route
claim_idThe exact action-boundary claim.Boundary is described as a slogan.Add stable claim id.
state_envelopeAllowed state or action region.No defined operating region.Declare envelope and exclusions.
action_proposalAction vector, route, or candidate command.Text answer is treated as command.Separate proposal from permission.
entropy_signalUncertainty or spread measure such as H(A | S_t).Uncertainty is hidden.Expose signal and threshold status.
baseline_divergenceDistance from deterministic baseline, such as D_KL(P_model || P_baseline).No baseline.Add baseline and distance rule.
fallback_decisionNo-action, review, narrower scope, or deterministic route.System guesses through the gap.Route to refusal or review.
receiptTrigger, state, decision, and unresolved evidence.Stop decision leaves no audit trail.Add public-safe receipt.
challenge_routeWhere readers can contest threshold, baseline, or boundary.No public attack route.Add GitHub issue or registry row.
Concept visual of an evidence registry terminal for bounded action receipts.

Receipt Layer

The useful public record is the decision boundary, not private execution.

A strong public record can show that an action would require a state envelope, baseline, threshold, fallback, and receipt. It does not need to expose private runtime material to make that boundary inspectable.

Challenge

Challenge one actuation boundary.

Point to any claim where the state envelope is vague, the baseline is missing, the entropy or divergence trigger is not defined, or the fallback route permits guessing under uncertainty.

The strongest challenge names the action class, the missing boundary field, and the public-safe receipt needed to make the claim inspectable.