DIA Agent Autonomy Gradient¶

Document Status¶

Field	Value
`policy-id`	`DIA-AUTON-LEVELS-001`
`type`	Implementing act (Level 3 of the normative hierarchy)
`version`	0.1.0-draft
`basis`	Art. II.3-4, II.8, V.10, V.13 of the DIA Constitution

1. Purpose of the Document¶

The Constitution (Art. II.3) requires that "the greatest power of the system pass through the human." At the same time, Art. V.10 requires an agent to have a kill switch, limits, and an explicit trust mode. This document defines four levels of agent autonomy, operationalizing these principles without creating a narrow human-in-the-loop bottleneck for every operation.

The autonomy gradient allows a human to set the frame, not click "OK" on every step. Power passes through the human because the human defines the gradient.

2. Levels of Autonomy¶

A0 - Proposal (Propose & Wait)¶

Parameter	Value
Description	The agent prepares a proposal. It takes no action without explicit approval from the operator.
Default when	Decisions that change policies, contracts, permissions, publications, or sensitive data. Any action with irreversible or hard-to-reverse effects.
Reporting	The proposal is shown to the operator with rationale, options, and risk assessment.
Reversal	Not applicable - the action was not taken.
Examples	Editing a public document; changing a node policy; sending a message on behalf of the user; escalating a case; modifying an agent contract.

A1 - Act & Notify¶

Parameter	Value
Description	The agent takes action but immediately informs the operator. The operator can reverse the action within a defined time window.
Default when	Actions with low risk and high reversibility that require speed but are not routine.
Reporting	Immediate notification with a description of the action, rationale, and reversal instructions.
Reversal	Possible within a defined window (default: a federation parameter, e.g. 15 minutes). After the window expires, the action is treated as approved.
Examples	Routing a task to another node; updating a memarium cache; answering a low-stakes network query; logging a sensorium event.

A2 - Act Within Budget¶

Parameter	Value
Description	The agent acts autonomously within an explicitly defined budget: limits of time, token cost, scope of operations, and number of actions. It reports after the fact.
Default when	Routine, repeatable operations with a predictable scope and low unit risk.
Reporting	Aggregate report (periodic or after the budget is exhausted) with metrics: number of actions, cost, deviations from the norm.
Reversal	Individual actions may be hard to reverse, but the budget limits the scale of harm.
Budget limits	Defined in the agent contract: `max_cost`, `max_time`, `max_actions`, `scope_whitelist`, `scope_blacklist`. Exceeding any limit -> automatic stop and operator notification.
Examples	Answering routine queries; aggregating sensorium data; maintaining a memarium index; monitoring node health metrics.

A3 - Emergency Mode¶

Parameter	Value
Description	The agent acts at maximum speed in a situation of direct danger to life or sudden serious harm. It leaves a full trace. Post-hoc review is mandatory.
Activated when	Only when the conditions of Art. II.8 are met: direct danger to life or sudden, direct, and serious harm to health.
Reporting	Full, unredacted trace of all actions, stored locally and (if possible) replicated.
Reversal	Not the priority during the crisis. After the crisis -> mandatory review and possible correction.
Time limits	A3 mode has a defined maximum duration (a federation parameter). After it expires, the agent automatically returns to level A0 (fail-closed).
Activation	Automatic (based on sensorium signals or crisis-pattern detection) or manual (operator). Automatic activation requires a separate confirmation in the log.
Examples	Coordinating aid during a blackout; alert about threat to life; first-contact medical triage; securing a whistleblower's communication channel under direct threat.

3. Rules for Assigning Levels¶

3.1. Agent Contract¶

Each agent declares in its contract (Art. V.10):

autonomy:
  max_level: A2           # Maximum level the agent is designed for
  default_level: A1       # Default level at startup
  emergency_capable: true # Whether the agent is capable of A3
  budget:
    max_cost_tokens: 1000
    max_time_seconds: 3600
    max_actions_per_cycle: 50
    scope_whitelist:
      - "memarium.read"
      - "memarium.index"
      - "sensorium.aggregate"
    scope_blacklist:
      - "policy.modify"
      - "reputation.vote"
      - "publish.*"

3.2. The Operator May Lower, Never Raise¶

A node operator may set an agent to a lower level than the contract's max_level. The operator may not set a higher one. Example: an agent with max_level: A2 may be restricted to A0, but an agent with max_level: A1 may not receive A2.

Rationale: autonomy level is an architectural constraint, not a convenience parameter. An agent designed for A1 does not have the budget mechanisms required by A2.

3.3. A Federation May Tighten¶

A federation may impose a max_level lower than the agent contract (e.g. in CORP_COMPLIANT mode all agents are A0). It may not weaken the limits from the contract.

3.4. Escalation Upward¶

An agent may not escalate its own level of autonomy (zero self-authorize, consistent with Art. V.13). Escalation requires:

A0 -> A1 or A2: operator decision.
Any -> A3: operator or automatic crisis detection with an explicit log entry and mandatory post-hoc review.
A3 -> return: automatic after expiration of the time limit (fail-closed to A0).

4. Matrix: Operation Type x Autonomy Level¶

The table below is the default. Federations may tighten it (move left), but may not loosen it (move right).

Operation category	Minimum level	Rationale
Policy / contract change	A0	Irreversible, affects governance
Publication / external communication	A0	High reputational stakes
Reputation vote	A0	Affects trust routing
Modification of sensitive data	A0	Privacy, dignity
Routing a task to another node	A1	Reversible, but requires awareness
Updating cache / index	A2	Routine, budgetable
Answering a routine query	A2	Routine, budgetable
Sensorium aggregation	A2	Routine, budgetable
Protection of life / crisis triage	A3	Art. II.8

5. Audit and Monitoring¶

5.1. Decision Traces¶

Each action of an agent, regardless of level, generates a log entry with at least:

timestamp,
autonomy_level at the moment of the action,
action_type,
scope (which resources were used),
cost (if measurable),
justification (for A1 and A3 - explicit; for A2 - available on request).

5.2. Budget Review (for A2)¶

An agent operating at A2 generates a budget report containing:

usage of the cost limit (%),
usage of the time limit (%),
number of actions vs. limit,
deviations from the norm (anomalies).

The report is available to the operator on request and is generated automatically when >=80% of any limit is exhausted.

5.3. Post-Hoc Review (for A3)¶

After A3 ends, a review is mandatory and includes:

the full action trace,
assessment of whether A3 activation was adequate (whether the threat was real),
assessment of proportionality of the actions taken,
identification of side effects,
recommendations for calibration of activation thresholds.

The review is documented and available for audit.

6. Relation to Other Documents¶

Constitution Art. II.3-4: The autonomy gradient operationalizes the principle that "power passes through the human."
Constitution Art. V.10: Autonomy levels extend the agent contract with an explicit autonomy_level parameter.
Constitution Art. V.13: This document concretizes the ban on agents independently escalating privileges.
Constitution Art. IX: A3 mode is the formalization of Art. II.8 and Art. IX.3.
Constitution Art. XIV: Every use of A3 is a constitutional exception, even if it has a simplified activation path due to time pressure.