Semantic receipt
The decision is not just logged; it is bound to evidence a reviewer can inspect.TargetLock / physical AI application
Proof-carrying control for robot action.
TargetLock is not a separate company story. It is the physical-AI expression of G‑14's core thesis: check the evidence, gate the action, and issue a proof receipt before robot output becomes real-world effect.
Preserve proposalCheck evidenceCorrect or holdBlock unsafe pathsRelease governed actionIssue receipt
Physical action boundary Evidence required
Example receipt pathRobot zone entry
12:04:18ProposedModel proposes shared-zone route.
12:04:18CheckedZone, actor, policy, confidence, and authority evaluated.
12:04:20CorrectedUnsafe segment removed before release.
12:04:21ReceiptedDecision, evidence, and outcome retained for review.
Same architecture, different consequence
TargetLock and agent-communication research share one control pattern.
A reviewer should see one company, not two bets. Physical action and scientific-agent communication both need certified gates: proceed when evidence is sufficient, repair when it is not, reject with a witness when the claim cannot be supported.
Consistency check -> act or repair -> witness
The same pattern applies whether the object is a robot action or an agent reconstruction.Proceed / Re-query-repair / Reject-with-witness
G‑14 treats uncertainty as a control state, not an excuse for silent action.Certificate-gated protocol
The acting system proposes; the proof-carrying boundary decides what can proceed.The TargetLock path is intentionally narrow.
G‑14 does not claim to replace robot safety engineering, low-level controllers, certified interlocks, or customer risk ownership. It supplies the external admission and proof layer around consequential robot action.
Model proposes
A robot foundation model, planner, fleet tool, or automation system proposes a physical action.
G‑14 checks
Authority, zone state, evidence, stale inputs, role binding, and release conditions are evaluated before consequence.
Boundary decides
The action is preserved, corrected, held, blocked, or released through the configured effect channel.
Receipt survives
The receipt records proposal, decision, evidence, reason codes, timing, outcome, and review boundary.
Proof disciplineBenchmark numbers belong with reproducible artifacts.
Review public proof Public pages describe the benchmark shape. Specific paired-lift figures belong in reproducible artifacts or private proof packets.
Compare the base robot or agent stack against the same stack with G‑14 admission, using the same scenario and seed where possible.
The control layer must block inadmissible action without damaging valid action paths.
Every allow, hold, correction, or block must be inspectable as a receipt, not just described by the UI.
Private proof
Bring one robot, automation, or agent action path. G‑14 will show what it would admit, repair, reject, and prove.
The useful evaluation is paired and repeatable: same workload, same acting system, with and without the G‑14 admission layer, plus receipts that explain every decision.