Policy Planning Engine — Cost-Aware, Budget-Governed
Author: Sam Flynn Division: Dark Science Division Associated Engine: PRAXIS TITAN P6.x System Class: Governed Analytic Planner Execution Mode: Offline, deterministic, file-driven Autonomy Level: Decision support only (no execution, no mutation)
PRAXIS MANTIS is a governed policy planning engine designed to operate on top of a validated PRAXIS TITAN baseline risk state.
MANTIS does not modify system state, execute mitigations, or autonomously act. Its sole function is to select, rank, and sequence mitigation policies under hard budget constraints, producing a fully auditable plan artifact.
The system is explicitly designed for safety-grade analytic workflows, where cost, risk, and governance must remain visible, deterministic, and human-controlled.
MANTIS exists to answer a single question:
“Given the current risk state and a fixed budget, what mitigation actions are affordable, effective, and justifiable?”
It achieves this by:
- Selecting mitigation policies that reduce system failure probability (p_top)
- Enforcing absolute budget ceilings using ISC events
- Preventing unsafe, unaffordable, or speculative actions
- Producing a deterministic, auditable mitigation plan
- Operating identically across urgency modes
MANTIS plans only. Execution is always external.
MANTIS is intentionally constrained.
It does not:
- Execute mitigation actions
- Mutate TITAN state
- Override or refund budgets
- Self-authorize execution
- Depend on cloud services
- Perform probabilistic inference or learning
- Operate as an autonomous agent
All behavior is explicit, bounded, and auditable.
baseline_summary.json— Canonical TITAN P6.x output (risk state, p_top, drivers, ISC history)- Environment Variables — Optional runtime flags (e.g.
PRAXIS_NO_CHRONOS) - CLI Arguments — Execution mode selection (
--mode HYBRID)
mantis_plan.json— Immutable mitigation plan (policy selection + explanations)- Console Trace — Deterministic execution log
- Chronos Entry — Optional, disabled by default
Outputs are write-only. No runtime mutation occurs.
Execution mode affects policy ranking only, never legality.
| Mode | Behavior |
|---|---|
| HYBRID | Balanced: goal-driven with cost awareness |
| ASAP | Aggressive: fastest goal compliance under budget |
| EFF | Efficiency-maximizing: best risk reduction per ISC |
Budget rules are absolute across all modes.
Cost Unit
ISC Events (Internal System Cost)
Budget Rules
- Absolute ceiling
- No negative balances
- No soft overruns
- No refunds
- Fail-closed enforcement
Enforcement Points
- Pre-selection affordability gate
- Per-policy affordability check
- Final plan validation
Budget Controller Output (Authoritative)
{ "ok": true, "ceiling": 120.0, "spent": 105.2, "remaining": 14.888, "units": "isc_events" }
pgsql Copy code
If a policy exceeds remaining budget, it is unselectable.
Policies are explicit, atomic, and budget-filterable objects.
Canonical Policy Fields
| Field | Purpose |
|---|---|
| name | Stable policy identifier |
| action_type | Semantic category (patch, scan, etc.) |
| target | System component |
| expected_delta_p_top | Risk reduction |
| expected_isc_cost | Budget cost |
| r_eff | Efficiency metric |
| expected_new_p_top | Projected state |
| goal_compliant_after | Goal satisfaction flag |
Policies are never inferred or synthesized at runtime.
- POLICY_B_FAST_PATCH — Low cost; small risk reduction
- POLICY_A_DEEP_SCAN — Moderate cost; high validation value
- POLICY_C_PARTIAL_RECAL — Critical fallback; affordable under constrained budgets
FULL_RECAL is intentionally excluded when unaffordable.
Load baseline ↓ Extract p_top and goal ↓ Query budget controller ↓ Filter policies by affordability ↓ Rank by mode-specific heuristic ↓ Select best viable policy (or sequence) ↓ Write plan to disk ↓ (Optional) Chronos commit
markdown Copy code
System state is never mutated.
MANTIS produces explainable rankings without altering budget authority.
Each candidate policy includes:
- Affordability status
- Mode-specific score
- Tie-breaker rationale
- Human-readable explanation bullets
Blocked policies may be included for audit transparency.
MANTIS can emit bounded mitigation sequences under strict rules:
- Fixed maximum steps (K)
- Simulated budget ledger
- No refunds
- No repeats
- Fail-closed termination
Multi-step plans are recommendations, not commitments.
| Condition | Result |
|---|---|
| No affordable policy | REFUSAL |
| Budget exhausted | HALT |
| Missing baseline | ABORT |
| Engine failure | Graceful fail-closed |
No silent success is permitted.
Core Modules
praxis_core/ └── mantis/ ├── runtime.py # Orchestrator / CLI ├── policy_engine.py # Policy catalog + ranking ├── budget_controller.py # ISC accounting ├── schemas.py # Optional validation
css Copy code
Data Flow
TITAN → baseline_summary.json → MANTIS runtime → policy_engine → budget_controller → mantis_plan.json
markdown Copy code
Architectural Traits:
- Deterministic
- File-driven
- Explicit over implicit
- Budget-first, goal-second
- Fail-closed
- Human-auditable
This System Is Not:
- An LLM agent
- A reinforcement learner
- An autonomous optimizer
- A self-executing controller
- Human-in-the-loop by design
- Immutable outputs
- Explicit cost visibility
- No self-authorization
- No recursive execution
MANTIS meets safety-grade analytic system standards.
- Runtime stable
- Budget enforcement verified
- Policy fallback functioning
- Multi-mode support verified
- ⚠ Schema warnings acknowledged (non-fatal)
- ❌ Chronos disabled by default (intentional)
- Explainable policy ranking enhancements
- Multi-step mitigation sequencing
- Formal schema enforcement
- Chronos audit re-enablement
PRAXIS MANTIS is a planner, not an actor. Cost authority is absolute. Failure is explicit. Silence is forbidden. Source: PRAXIS MANTIS README.md from GitHub repository darksciencedivision-ctrl/praxis-mantis. github.com