bp Sphere · Governed Token Optimization
AI Runtime Efficiency Dashboard
Never spend a token that does not improve a governed business outcome.
LLM runtime is the last-resort synthesis layer after deterministic decision, policy, evidence, and action runtimes.
Evidence, replay, source lineage, and audit context remain outside the prompt unless requested; every token-heavy explanation must reference the governed decision replay or evidence pack that justified it.
Overall readiness
91.9%
Weighted enterprise efficiency score
Resolved without LLM
100.0%
80-95% of business decisions resolved without an LLM call; LLM is last-resort synthesis/explanation.
Prompt reduction
82.0%
812 → 146 tokens
Context compression
82.0%
812 loaded · 146 used
Cache / reuse
0.0%
opportunity_only
Token audit coverage
100.0%
Prompt/input/model/cost lineage required
Enterprise Efficiency Scorecard
| Area | Weight | Score | Weighted | Status |
|---|---|---|---|---|
| Architecture | 20% | 96.0% | 19.2 | green |
| Runtime execution | 20% | 90.0% | 18.0 | green |
| Agent implementation | 15% | 82.0% | 12.3 | amber |
| Prompt compilation | 15% | 94.0% | 14.1 | green |
| Context engineering | 10% | 92.0% | 9.2 | green |
| Model routing | 10% | 95.0% | 9.5 | green |
| Monitoring & FinOps | 10% | 96.0% | 9.6 | green |
Runtime Execution Mix
The target state is deterministic/cache/memory first, with LLM-heavy execution reserved for complex synthesis.
| Mode | Count | Share |
|---|---|---|
| Deterministic | 5,000 | 100.0% |
| Cache | 0 | 0.0% |
| Memory | 0 | 0.0% |
| Hybrid | 0 | 0.0% |
| Llm Heavy | 1 | 0.0% |
Agent LLM Classification
Configured agents
19
No-LLM / deterministic
15.8%
LLM assisted
16
| Agent type | Count | LLM use |
|---|---|---|
| Deterministic | 3 | Never |
| Email Writer | 1 | Sometimes / governed |
| Narrative | 3 | Sometimes / governed |
| Planner | 1 | Sometimes / governed |
| Retrieval Agent | 11 | Sometimes / governed |
Prompt, Context, and Routing Controls
LLM calls/action
0.0
Target < 1
LLM share
100.0%
Dynamic model routing enabled
Shared context reuse
91.8%
One evidence pack, many deterministic executors
AI spend
$0.0008
958 recorded tokens
| Work type | Preferred route |
|---|---|
| classification | small |
| narrative | balanced |
| complex reconciliation | large |
| simulation | reasoning |
| policy/threshold/SOR validation | deterministic |
Top Token Consumers
Unmapped
958
enterprise-decision-runtime · 100.0% of recorded tokens
World-Class Target Contract
| Metric | Target | Payload field |
|---|---|---|
| Business decisions resolved without LLM | 80-95% | request_mix.resolved_without_llm_pct |
| Prompt reduction via compilation | 60-90% | prompt_compilation.reduction_pct |
| Context reduction | 70-95% | context_engineering.compression_pct |
| Cache hit rate | 50-80% | cache_runtime.hit_rate_pct |
| Average LLM calls per user action | < 1 | request_mix.llm_calls_per_decision |
| Shared context reuse across agents | > 90% | multi_agent_efficiency.shared_context_reuse_pct |
| Token-level auditability | 100% | governance.token_audit_coverage_pct |
| Cost attribution by mission, agent, journey | 100% | governance.cost_attribution_coverage_pct |
Audit Proof
| FinOps source | FinOpsRuntime |
| Token ledger | LLMAuditRecord |
| Decision ledger | AgentExecution |
| Business value ledger | ValueEventRecord |
| Governance config | domains/finance/llm_governance.yaml |
| Price card | token_price_card.yaml |