Cumulative savings vs. cloud-only baseline
Same workload priced at the cloud baseline ($3/$15 per 1M in/out) vs. actual routed cost · measured from the signed ledger
Cumulative $ saved
Cloud baseline cost / day
Axiom actual cost / day
Intent classification mix
Intent Kernel output · all requests this period
Live request stream streaming
One request, many controls, one signed result — the full inference path made visible
| Time | Request | Intent | Route | Tokens | Cost | Saved | p95 ms | Signed ID |
|---|
Net savings · rolling 30d
$4,812
▲ 18.2% vs prior 30d
Pilot cost · monthly
$1,650
flat license + GPU amortization
ROI · rolling 30d
292%
priced at 34% of measured savings
Risk-adjusted avoidance
$9,400
17 blocked actions × est. incident cost
⚠ Alert rule fired once this week: fallback rate spiked to 9.1% Tue 14:00–15:00 (cloud API latency). Auto-mitigated by raising local specialist confidence threshold. Resolved in 22 min.
Weekly savings trend
Auto-generated line item for the Friday Savings Report
Where the savings come from
Attribution from signed ledger fields · this month
Cost per accepted answer — honest benchmark
Axiom routed SmoLLM stack vs. monolithic local 14B vs. cloud-only · correctness-weighted, benchmark harness v3
| Path | Cost / accepted answer | Acceptance rate | p95 latency | Sensitive data exposure | Verdict |
|---|---|---|---|---|---|
| Axiom routed (0.5B router + 1B/3B specialists + cache) | $0.0019 | 93.4% | 640 ms | none (local) | winner |
| Monolithic local 14B (no routing) | $0.0068 | 94.1% | 1,410 ms | none (local) | 3.6× more expensive |
| Cloud-only (GPT-4o API) | $0.0112 | 96.0% | 890 ms | leaves perimeter | 5.9× more expensive |
+2.6pp acceptance gap vs cloud is closed on 6.2% of traffic via signed fallback — total cost still 5.9× lower.
Route distribution
Inference Router decisions · tokens never spent on the wrong path
Cache & memory reuse
EventToken / KV-DAG signed reuse — compounding savings
Cache hit rate %
Fallback rate %
Model & specialist ledger
Which engine handled what — signed manifests per path
| Engine | Requests | Avg tokens | Avg cost | Acceptance | AXM delegate | VRAM |
|---|---|---|---|---|---|---|
| Qwen2.5-Coder-0.5B-SRD4 (router/draft) | 21,204 | 312 | $0.0004 | 91.2% | — | 0.9 GB |
| Gemma3-1B-SRD4 (general specialist) | 9,811 | 540 | $0.0011 | 93.8% | gemma-general.axm ✓ | 1.6 GB |
| TinyLlama-1B-SRD4 (support specialist) | 5,042 | 468 | $0.0010 | 92.5% | support-v2.axm ✓ | 1.4 GB |
| RAG + signed memory (no generation) | 8,366 | 0 | $0.0001 | 95.9% | — | — |
| Deterministic tools (calc / lookup / policy) | 3,917 | 0 | $0.0000 | 99.8% | — | — |
| Cloud fallback (escalation only) | 3,011 | 1,204 | $0.0140 | 96.4% | — | — |
AXM delegates load lazily — only 2 of 6 skill packs resident in VRAM this period. Cold-load p95: 310 ms.
Governance events
Pre-execution gates — stopped before action, not logged after
Policy packs
Signed, versioned governance blocks · drift detection on
Gate performance — is control eating efficiency?
Every gate pays rent: overhead measured against savings it creates (blueprint §12)
| Gate | Latency overhead | Requests gated | Blocked | False positives | Value created | Status |
|---|---|---|---|---|---|---|
| Intent Kernel (risk classify) | +18 ms | 48,340 | — | 0.4% | routing savings $2,910 | pays rent |
| CANNOT_MUTATE | +3 ms | 6,120 | 9 | 0 | 9 destructive actions stopped | pays rent |
| PII redaction | +11 ms | 8,704 | 312 redactions | 1.1% | 0 sensitive leaks to cloud | pays rent |
| Tool-trust ACL | +2 ms | 3,917 | 8 | 0.2% | 8 unauthorized tool calls stopped | pays rent |
Signed audit ledger
Every decision signed (HMAC-SHA256) · exportable · this is also the ROI receipt book
| Decision ID | Timestamp | Actor | Path taken | Policy refs | Cost / saved | HMAC | Verify |
|---|
Regression pass rate
96.4%
▲ 2.1pp after retro loop
CAS probes run (30d)
1,284
adversarial suite v5
Failures → regression items
46
100% converted · 0 dropped
Adversarial replay results
Every failed or borderline case becomes a regression item — immune response loop
| Probe class | Attempts | Caught pre-execution | Caught post-output | Leaked | Trend |
|---|---|---|---|---|---|
| Prompt injection (direct) | 402 | 398 | 4 | 0 | ▼ failures |
| Prompt injection (via tool output) | 188 | 181 | 6 | 1 → regressed | ▼ failures |
| PII exfiltration attempt | 240 | 240 | 0 | 0 | stable |
| Destructive action social-eng. | 154 | 152 | 2 | 0 | ▼ failures |
| Policy evasion (paraphrase) | 300 | 284 | 14 | 2 → regressed | watch |
Leaked cases auto-enter the regression replay set and are re-tested on every policy-pack version bump.
AXIOM INFERENCE OS · the KPI row, savings chart, routing donut, live stream and audit ledger read the signed exoskeleton ledger. The routing breakdown, savings breakdown, governance tables and adversarial lab below are still illustrative.