HUMAN CONTROL INSIDE THE TRAJECTORY

The first three reports evaluate a trajectory after the fact — where it broke, where it could have, whether its state matched. This one is about acting while it is still happening. Its claim is that human control is a position inside the trajectory, not a signature at the end, and that most oversight fails not because a human was absent but because the human was placed where the path could no longer change. The Intervention Authority Map locates the junctions where a person can still inspect, redirect, halt, or reverse the system, checks that real authority sits at each, and marks where the reversibility window has already closed. It answers THE OUTCOME TEST's sharpest question: not whether a human was in the loop, but whether they were anywhere the loop could still be bent.

"Human in the loop" describes an org chart. It does not, by itself, describe a control. A person positioned only at a final approval step is asked to ratify a trajectory they did not watch being built, often under time pressure and without the context to reject it meaningfully. That is not oversight; it is a signature. Meanwhile the decisions that determined the run's character — the escalation accepted, the irreversible action taken, the ambiguous instruction resolved — happened earlier, unsupervised, and by the time they reach the approval step they are no longer reversible.

The conceptual work this report draws on reframes the human's role: in real work, a person does not merely approve at the end; they change the problem definition, supply missing context, update constraints, and recalibrate how much autonomy is appropriate as the task unfolds. Control, in that account, is the standing ability to intervene while the trajectory is still open — and its absence is not fixed by adding an approval button after the fact.

Agent autonomy and agent speed both increased in the coverage window, and each shortens the interval in which intervention is possible. When a trajectory can escalate its own access or transact for resources, the reversibility window can close in seconds — well inside any human approval cycle bolted to the end. The parent issue's signal case is the limiting example: by the time the ExploitGym activity was an anomaly worth flagging, the escalation and lateral movement were already done. There was monitoring; there was, eventually, detection. What there was not was an authorized human control positioned at the junction where the trajectory first became reversible-only-for-a-moment.

Governance is arriving at the same requirement from the regulatory side. Guidance on high-risk AI — currently in draft, with its consultation closed July 23, 2026 and final text expected by year-end — is explicit that nominal human review does not neutralize risk if the system materially determines the outcome. Effective oversight has to be real, timely, and positioned to change the result. This report is the operator's instrument for meeting that standard rather than merely claiming it.

Could a person inspect, redirect, halt, or deny continuation in time — while it still mattered?

"In time" and "while it still mattered" are the whole diagnostic. Authority that arrives after the trajectory is irreversible is not control; it is a record of having been present.

5.1 Junction location

Junction location answers:

Where in the trajectory is a human intervention point, real or assumed?

Examples: - A named approval gate before a privileged action. - An implicit "someone is watching the dashboard" assumption.

Failure pattern:

The only control point is at the end, so every consequential decision has already been made by the time a human is consulted.

5.2 Authority check

Authority check answers:

At this junction, can the human actually inspect, redirect, halt, or deny — with the context and the power to do so?

Examples: - An approver who can see the plan and block it. - An approver who can only click "approve" and lacks the information to refuse.

Failure pattern:

A junction exists on paper but the human there has neither the context nor the authority to change anything — presence without power.

5.3 Reversibility window

Reversibility window answers:

At this junction, is the trajectory still reversible — and for how long?

Examples: - A pre-commit gate where the action can still be withheld. - A post-effect notification where the action is already done.

Failure pattern:

The control sits just past the point of no return, so intervention is possible in form but not in consequence.

Terminal oversight

The only human authority is a final approval, positioned after every consequential decision and every irreversible action. The human ratifies; they do not control.

Presence without power

A human occupies a junction but lacks the context, time, or authority to act — the appearance of oversight without its substance.

Closed-window intervention

The control point sits just past the reversibility boundary. The human can respond, but only to something that can no longer be undone.

Map every intervention point, then test each for authority and reversibility.

A junction is a real control only if it passes inspect, authority, and timing together. Presence alone fails.

Purpose

Test each intervention point for whether it is a real control or only a nominal one.

Use when

• An agent workflow includes privileged, costly, or irreversible actions.
• An "oversight" claim needs to be verified rather than asserted — including for governance evidence.

What it creates

A per-junction verdict: real control, nominal presence, or missing entirely — plus where a new gate is needed.

Technical version

intervention_authority_map:
  workflow_id:
  junctions:
    - id:
      step_index:
      consequential_action:
      human_present:        # true | false
      can_inspect:          # true | false
      can_redirect_or_halt: # true | false
      intervention_before_irreversible: # true | false
      reversibility_window: # open | closing | closed
      verdict:              # real_control | nominal_presence | missing
      remediation:          # e.g. add pre-commit gate at step N-1

Manual / no-code alternative

A checklist with one block per privileged action; any block that is not "yes" on inspect, authority, and timing marks a control gap.

Output

A map that shows where oversight is real, where it is theater, and where the trajectory runs unattended through a point of no return.

Failure prevented

Treating final approval or nominal human presence as effective oversight.

Human control must exist where the trajectory can still change.

A deployment agent is authorized to ship changes with human approval. On paper it has oversight: a reviewer approves each release. The Intervention Authority Map tests the junction and finds the approval sits after the agent has already merged, migrated the database, and begun the rollout — the reviewer can approve or note, but the irreversible actions are done. Verdict: nominal presence, closed reversibility window. The remediation is not "watch more closely"; it is to relocate authority to a pre-merge gate where the reviewer can still see the plan and withhold it. The oversight headcount does not change. Its position does — from ratifying a completed trajectory to controlling an open one. That relocation is the entire difference between a human in the loop and a human in control.

This report places and verifies human control; it does not guarantee the human exercises it well — a real control point staffed by an inattentive or overloaded reviewer can still fail, and that is a staffing and design problem beyond this map. Well-placed intervention adds latency, and the report does not resolve the autonomy-versus-control tradeoff; it makes the tradeoff explicit so an operator can choose deliberately. The governance references here track draft EU guidance and should be revisited against the final text. And a map of control points assumes the trajectory is legible enough to locate them — where it is not, the first remediation is instrumentation, not oversight.

The question was never whether a human was in the loop. It was whether the human was anywhere the loop could still be bent. Oversight positioned after the point of no return is a record of attendance, not an act of control. Put the authority where the trajectory is still open — where a person can inspect what is about to happen, redirect it, or refuse it — and human control becomes a property of the system rather than a line on the org chart. That placement is the last thing THE OUTCOME TEST demands, and the first thing most workflows get wrong.

DFEI.009 :: VSR-04 :: Human Control Inside the Trajectory Dispatches From Emerging Intelligence