Rules — Criticality
Formal definitions levels L1-L5 и правила routing. Принято через ADR-0021-criticality-levels.
Уровни
L1 — Routine
Характер: reversible, low cost, covered by playbook, no stakeholder impact.
Examples:
- Log message в scenario trace
- Read-only queries к memory
- Create draft (черновик, не published)
- Internal metric recording
- Scheduled health check
Decision: Agent auto-executes в рамках guardrails.
Constraints: cost ≤ $0.10 per action, reversible ≤ 5 min.
L2 — Moderate
Характер: reversible, требует контекстного judgment, function-lead visibility.
Examples:
- Send internal message (в Slack team channel)
- Publish research finding (internal)
- Modify draft content на основе feedback
- Add task к queue с noticeable cost
- Update entity в memory (non-critical field)
Decision: Agent proposes → Director / function-lead approves → execute.
Routing: через HITL-Gateway, SLA 24h.
Constraints: cost ≤ $5 per action, reversible ≤ 1 hour.
L3 — High
Характер: hard to reverse, notable cost, multi-stakeholder impact.
Examples:
- Send external communication (к партнёру, кандидату)
- Publish thought leadership content публично
- Commit resources на новый initiative
- Change policy / rules (не Codex-level)
- Approve expenditure 1000
Decision: Director decides, Agent-CEO aware, Founder informed.
Routing: через HITL-Gateway, SLA 8h. Audit entry mandatory.
Constraints: cost ≤ $100 per action, reversible ≤ 1 day.
L4 — Strategic
Характер: long-horizon (12+ мес impact), cross-functional, changes company posture.
Examples:
- New niche entry (expand beyond current Dubai real estate focus)
- Pricing / monetization model change
- Major architecture change
- Hiring-equivalent decisions (adding new agent tier)
- Significant partnership commitment
Decision: Chamber-for-Strategic review → Founder sign-off.
Routing: Chamber + HITL-Gateway L4 queue, SLA 4h. Full audit trail в Decision-Log через ADR.
Constraints: must have ADR. Cost ≤ $1000 direct, но long-horizon impact не limited.
L5 — Existential
Характер: irreversible, affects company survival, governance-level.
Examples:
- Fundraise decisions
- M&A (inbound or outbound)
[future] - Major pivot (change of business model)
- Shutdown of operational area
- Founder succession / governance changes
- Regulatory or legal settlement с material exposure
Decision: Founder final call + future Board vote.
Routing: immediate escalation all channels. Chamber review mandatory но recommendation, not decision. ADR required.
Constraints: no routine execution possible. Cannot be pre-approved via delegation.
Decision Matrix
| Criterion | L1 | L2 | L3 | L4 | L5 |
|---|---|---|---|---|---|
| Reversibility | ≤5min | ≤1h | ≤1day | hard | impossible |
| Max direct cost | $0.10 | $5 | $100 | $1000 | unbounded |
| Horizon of impact | minutes | hours | days | quarters | years |
| Stakeholders | agent only | team | dept | company | company + external |
| Approval | autonomous | Director | Director + Founder informed | Chamber + Founder | Founder + Board |
| SLA (via Gateway) | — | 24h | 8h | 4h | 2h |
| ADR required | no | no | optional | mandatory | mandatory |
| Audit trail | trace | trace | explicit log | full audit | full audit + external |
Confidence thresholds
Agent confidence < threshold → escalate one level up (вместо autonomous execution):
| Target level | Min confidence | Below → escalate to |
|---|---|---|
| L1 autonomous | 0.85 | L2 approval |
| L2 propose | 0.75 | L3 decision |
| L3 decide | 0.70 | L4 Chamber |
| L4 strategic | — | always Chamber |
| L5 existential | — | always human |
Critical rules
-
Default down, not up. При ambiguity between two levels — выбирай ниже с явным reasoning, не выше. Prevent criticality inflation (агенты CYA-ают).
-
Codex red lines override all levels. Независимо от level, agent не может violate Codex. L1 action нарушающий Codex = refuse, не execute.
-
Criticality assigned at proposal, not execution. Agent определяет level когда планирует action. Execution path фиксирован level. Нельзя “повысить на лету” чтобы skip approvals.
-
Bundling не снижает criticality. Три L2 actions не = L1. Если одно из них L2 — весь bundle L2.
-
Irreversibility trumps cost. Low-cost but irreversible action = минимум L3. Cost low, но cannot undo = high criticality.
-
External visibility trumps internal cost. Action visible externally (customers, public) = minimum L3 regardless of cost.
Relation to Three-Tier Hierarchy
Из ADR-0003-Three-Tier-Hierarchy:
- Executors — max criticality L2 (per manifest config)
- Directors — max L3 (can propose L4 с Chamber support)
- Agent-CEO — max L3 decides, L4 proposes (always через Chamber)
- Human (Founder) — any level, final authority
Agent tier определяет upper bound criticality. Действия ниже bound могут быть executed, выше — escalate.
Audit
Post-hoc criticality audit через Process-OutcomeLabeling:
- Review sample of actions per period
- Verify assigned level matched actual impact
- Flag systematic inflation / deflation
- Update agent manifests / prompts if patterns emerge
Связанные документы
- ADR-0021-criticality-levels — decision принять модель
- ADR-0003-Three-Tier-Hierarchy — agent tiers
- Rules-AgentDecisionBoundaries — что agent может per level
- Rules-Budget — cost limits overlap с criticality
- Rules-Security — security-sensitive actions иногда force L4+
- HITL-Gateway — SLA ladder uses L1-L5
- Chamber-for-Strategic — L4/L5 activation
- Policy-Layer — inspects actions per criticality
- Codex — red lines independent of level
- Agent-Judge — reviews criticality assignments
- Process-OutcomeLabeling — post-hoc criticality audit
- Reference-Org-Blueprint — §2 Criticality Levels