Human Control Lab
Enter six synthetic investigations that separate human presence from meaningful control. See how machine filtering shapes evidence, how time compression turns approval into a bottleneck, how authority migrates after communications loss, and how decisions made before activation govern machine-speed behavior.
The analytical shift
A system may navigate automatically while a human retains release authority. A decision-support tool may never apply force, yet still decide which evidence reaches the operator. A local defense may complete a bounded response after activation because the available window is shorter than a meaningful review. The correct unit of analysis is the function, context, authority, information, time, and intervention path.
Control is disaggregated
These site definitions describe how authority is allocated in a particular function and context. They are not claims that every institution uses identical terminology.
AI or automation assistance
Human-in-the-loop
Human-on-the-loop
Preauthorized local action
Publicly unspecified
Investigation 01 · Upstream human removal
The human may retain the final button while an algorithm has already filtered the observations, merged the sources, ranked the candidates, and hidden everything below a threshold.
No gaze, pose, biometric, identity, or location data is collected. Selection exists only in this page session.
What the algorithm let me see
No person can inspect every observation in a high-volume stream. The question is whether filtering remains inspectable, reversible, provenance-aware, and able to preserve contradictions.
A model can be internally confident while relying on stale, correlated, incomplete, or out-of-distribution inputs.
The strongest influence may be the option, observation, or contradiction that never reaches the human interface.
Investigation 02 · The disappearing review window
The same fictional event runs through three authority architectures. Speed improves as handoffs disappear, but a faster chain can also apply an incorrect rule more consistently and leave less room for contextual judgment.
Policy written before the event
Human-directed
Review each observation and authorize each consequential action.
AI-assisted
Machine fusion and ranking; human authorizes after recommendation.
Preauthorized automation
A local rule engine acts only when the prior envelope is satisfied.
Human-Control Budget
Each dimension remains separate. The lab never converts them into a single morality or legality score.
Comparison complete
Investigation 03 · Authority after silence
Communications independence is not unrestricted lethal authority. When a link fails, navigation may continue while consequential action remains prohibited, expires, or enters a predefined safe state.
Bandwidth budget · 60 units
Select the information that must survive degradation. The values are fictional teaching costs.
0 / 60 units allocated
Predelegation contract
00Awaiting simulationSelect a bounded policy before the link is lost.
When the link failed
Investigation 04 · Nominal versus meaningful control
A mechanical click is not automatically a meaningful judgment. Queue volume, time pressure, interface defaults, confidence presentation, and the ability to request evidence can determine whether the human actually controls the decision.
Cognitive forcing functions
Synthetic queue
1 / 8Review window: generous
No response is transmitted. This is a transparent educational exercise, not a covert behavioral experiment.
I was in the loop—but was I in control?
Evidence Lab rule
Investigative reporting, official statements, independent analysis, legal interpretation, dispute, and unknown are different evidence states. The playable queue stays fictional and nonviolent; real cases belong in a source-analysis interface, not a reenactment.
Investigation 05 · Distributed authority
The final two seconds cannot reveal the whole decision architecture. Policy, data, thresholds, testing, mission boundaries, activation, supervision, updates, and auditability can all shape what the machine is permitted to do.
Select an initial attribution. The lifecycle remains hidden until you commit.
Accountability Shadow
The shadow grows when no person has adequate information, time, intervention ability, or reconstructable evidence.
My authority map
Investigation 06 · Pre-mission governance
What looks like a two-second machine decision may be the final state transition in a decision space shaped earlier by policy, engineering, testing, legal review, mission planning, configuration, and activation.
Fictional governance configuration
Choose how uncertainty produces review, restraint, reversible action, or bounded preauthorization. No target, weapon, route, or real operating parameter appears.
Two sensors disagree about one synthetic object.
WaitingA previously authenticated object loses its identifier.
WaitingThe review link fails while the event is still developing.
WaitingA protected corridor was activated after the plan was approved.
WaitingNavigation and sensor-based position estimates conflict.
WaitingPull the action backward through time
You programmed the decision space
Public evidence cards
Each card separates what public sources establish from what remains unknown. System sophistication, autonomous navigation, target recognition, decision support, and authority over force are separate questions.
A public official source documents the policy, architecture, status, or function. It does not prove every configuration or operational result.
An accountable public authority describes a program, policy, test, or control arrangement.
A supplier describes a capability. The statement remains a claim unless independently corroborated.
A researcher or investigative source interprets public evidence. The site keeps the interpretation separate from the underlying record.
Credible sources disagree, the account depends on anonymous reporting, or decisive evidence is unavailable.
The public record does not establish the operating mode, authority, threshold, abort path, or field behavior. Unknown is not guessed.
A timer, confidence score, object, threshold, result, or policy choice invented only for learning and not calibrated to a real system.
11 cases shown
Terminal defensive automation
A self-contained local-defense system publicly described as automatically detecting, evaluating, tracking, engaging, and assessing certain incoming threats.
Ship self-defense
A close-in ship-defense system combining autonomous Phalanx search and track functions with a guided interceptor launcher.
Selective rocket defense
A highly automated defensive chain that detects and tracks launches, estimates trajectories and landing areas, and concentrates interceptors on threats assessed to endanger defended zones.
Anti-radiation loitering munition
The manufacturer describes a system that can autonomously search for and strike emitting radar targets after human launch and mission setup.
Terminal target recognition
The manufacturer describes imaging-infrared Autonomous Target Recognition used to discriminate and recognize an intended target in the terminal process.
Post-launch sensing and guidance
Official sources describe semi-autonomous operation intended to reduce dependence on external networks in contested environments.
One-human-to-many teaming
The U.S. Air Force is developing autonomous aircraft functions while publicly retaining human authority over weapon release during current development and testing.
Upstream algorithmic gatekeeping
An AI-enabled intelligence and decision-support environment that processes large data volumes and helps surface, correlate, and present objects or options to human users.
Multidomain intelligence fusion
A U.S. Army ground-station program using AI and machine learning to process information from space, high-altitude, aerial, and terrestrial sources.
Distributed sensing and data products
A space-based layer publicly described as combining sensing, algorithms, processing, data fusion, and tactical data products for missile warning and tracking.
Sensor and effector integration
An Army command-and-control network intended to connect multiple sensors and effectors into a common integrated air-and-missile-defense picture.
No case matches the current evidence filter.
Continue the investigation
The Daily Break the Chain catalog now includes scenarios about hidden evidence, approval queues, lost-link policy, pre-mission authority, and accountability gaps. Each uses PHP-authoritative state, deterministic replay, and synthetic-only evidence.
Daily challenge
Inspect evidence, state confidence, select a proportionate control, and compare the path not taken.
Enter today’s challengeEvidence Atlas
Compare systems, doctrine, public control modes, evidence states, and unknowns without tactical coordinates.
Open the AtlasMethodology
Review how official records, manufacturer statements, independent analysis, disputes, and unknowns are classified.
Read the method