Human Operator Effectiveness

Human operator effectiveness is an ongoing program, not a feature with a completion bar. The goal is to keep fak understandable and steerable as the agent fleet, CLI surface, refusal vocabulary, and background loops grow.

The program’s job is to make a human better used:

• Use humans for judgment: policy choices, priority calls, missing witnesses, trust-boundary decisions, and review of repeated regressions.
• Let agents do routeable work: issues with a lane, a next action, and a witness.
• Keep learning bounded: one current focus, one practice step, and one explicit thing to skip.
• Keep evidence coherent: cached panes are useful only when their source stamps are visible.
• Keep check-ins cheap: an operator who read the last brief should start with what changed, resolved, or persisted.
• Keep the surface light: a growing system should not require a growing amount of operator ceremony.

Frontiers

The program is measured by frontiers and trends, never by “100% complete”:

Frontier	Surface	Meaning
Human attention	fak operator brief	Separates human, agent, watch, and background work so a person sees what actually needs them.
Learning pace	learning_agenda in the operator brief	Names the one thing to understand now, the practice step, the skip rule, and the drill-down sections.
Change compression	since_previous in the operator brief	Compares the current brief with a previous brief, showing new, resolved, and persistent attention-bearing items before detail.
Source coherence	coherence in the operator brief	Says whether cadence/program/milestone/heaviness reports describe one snapshot or a mixed cached state.
Surface pressure	fak operator heaviness	Reads the command surface, guard flags, refusal vocabulary, doc-map discoverability, and appeal channel.
Strategic tracking	fak program report	Tracks human-operator-effectiveness beside kernel and cache programs, using operator-heaviness lightness as the first deterministic metric.

Operating Rule

Do not make the operator read every transcript. A useful control plane should compress the fleet into:

1. The state of the system.
2. The choice a human can actually make.
3. The challenge worth understanding.
4. The work agents should keep doing.
5. What changed since the previous review.
6. The learning focus for this review.

If a surface cannot answer those five, it is background telemetry, not an operator page.

For live workers, use the narrower minimal long-running-agent signal projection: attention, latest witnessed outcome, current bounded move, and next check. Heartbeats, token motion, and transcripts remain drill-down rather than progress evidence.

Current Witnesses

• fak operator brief --collect --json folds the source reports into the operator envelope; --previous last-brief.json adds the since_previous change delta.
• fak operator heaviness --json emits heaviness_pressure and heaviness_debt.
• fak program report --json includes the human-operator-effectiveness program signal.

The current metric is intentionally modest: human_metric = max(0, 100 - heaviness_pressure), forced to 0 when hard heaviness_debt exists. It is a first proxy for how light the system feels to drive. Better future signals can add operator response time, repeated watch items, unresolved human choices, and brief read-time, but they must stay witnessed rather than self-reported.