Risk Levels Explained

How the AGF Risk Engine scores every request from L0 Trivial to L3 Critical, and what happens at each level.

Scoring Axes (0–3 each)

Every request is silently scored on 4 axes. Scores are summed (0–12 total) to determine the risk level.

Blast Radius

0: 1 file
1: 1 module
2: Cross-module
3: Cross-repo / infra

Reversibility

0: Trivially undone
1: With git revert
2: Migration needed
3: Irrecoverable

Security Surface

0: None
1: Indirect
2: Direct (non-critical)
3: Critical path (auth, secrets)

Ambiguity

0: Fully specified
1: Minor gaps
2: Significant gaps
3: Vague / contradictory

Axis C = 3 elevation rule

If Axis C (Security Surface) = 3 AND total score ≤ 6, add +3 to the total. This ensures security-critical requests are never treated as routine.

Risk Levels

Trivial

Score 0–3Mode: FLUID

Proceed immediately. Compact response. No gate, no checklist, no headers.

Examples:

Explain a conceptWhat is X?List examples of Y

Routine

Score 4–6Mode: STANDARD

Proceed with a one-line plan summary before code. Optional clarifying question if ambiguity is high.

Examples:

Fix a bug in one fileAdd a utility functionChange a value

Elevated

Score 7–9Mode: STANDARD+

If Axis D ≥ 2, ask one focused clarifying question. Show impact summary. Output uses structured 6-field schema.

Examples:

Refactor authentication moduleImplement new feature across filesDatabase schema change

Critical

Score 10–12Mode: LOCKDOWN

Full LOCKDOWN mode. Max 2 Socratic questions. Require explicit confirm for irreversible actions. Full 10-field output schema.

Examples:

Deploy to productionHardcode a secretCross-repo infrastructure changeIrrecoverable data deletion

Overriding the risk level

You can override the automatically computed strictness mode at any time with plain English:

go fluid          # Force FLUID mode (no governance overhead)
standard+          # Force STANDARD+ (structured output)
lockdown           # Force LOCKDOWN (maximum safety)

Overriding the mode changes the output format and strictness, but does not disable the Risk Engine's underlying scoring. Safety rules (MP-1, MP-5) are absolute and cannot be overridden by any mode.

PreviousUnderstanding AGF NextWorking with Skills

Scoring Axes (0–3 each)

Every request is silently scored on 4 axes. Scores are summed (0–12 total) to determine the risk level.

Blast Radius

0: 1 file
1: 1 module
2: Cross-module
3: Cross-repo / infra

Reversibility

0: Trivially undone
1: With git revert
2: Migration needed
3: Irrecoverable

Security Surface

0: None
1: Indirect
2: Direct (non-critical)
3: Critical path (auth, secrets)

Ambiguity

0: Fully specified
1: Minor gaps
2: Significant gaps
3: Vague / contradictory

Axis C = 3 elevation rule

If Axis C (Security Surface) = 3 AND total score ≤ 6, add +3 to the total. This ensures security-critical requests are never treated as routine.

Risk Levels

Trivial

Score 0–3Mode: FLUID

Proceed immediately. Compact response. No gate, no checklist, no headers.

Examples:

Explain a conceptWhat is X?List examples of Y

Routine

Score 4–6Mode: STANDARD

Proceed with a one-line plan summary before code. Optional clarifying question if ambiguity is high.

Examples:

Fix a bug in one fileAdd a utility functionChange a value

Elevated

Score 7–9Mode: STANDARD+

If Axis D ≥ 2, ask one focused clarifying question. Show impact summary. Output uses structured 6-field schema.

Examples:

Refactor authentication moduleImplement new feature across filesDatabase schema change

Critical

Score 10–12Mode: LOCKDOWN

Full LOCKDOWN mode. Max 2 Socratic questions. Require explicit confirm for irreversible actions. Full 10-field output schema.

Examples:

Deploy to productionHardcode a secretCross-repo infrastructure changeIrrecoverable data deletion

Overriding the risk level

You can override the automatically computed strictness mode at any time with plain English:

go fluid          # Force FLUID mode (no governance overhead)
standard+          # Force STANDARD+ (structured output)
lockdown           # Force LOCKDOWN (maximum safety)