Attribute every model dollar.
Break down spend by tenant, workflow, provider, and use case before the invoice becomes a surprise.
GOVERISPowered by Calybris Engine
AI spend governance for LLM APIs
Control LLM spend before the invoice arrives.
Find 20–40% avoidable model spend without touching prompts. Private VPC deployment. Metadata-only observation. Board-ready audit in 7 days.
No prompt capture
Runs in your VPC
294 tests fault injection + proptest + loom
Built for OpenAI, Anthropic, Google, DeepSeek, Meta, and Mistral.
1Mreal conversations benchmarked — 0 crashes, 0 data corruption
14–45%marginal savings over your current routing — measured, not theoretical
311passing Rust tests across recovery, concurrency, and policy invariants
9.19M/scorrectness-guarded kernel median on one local host
Quality floor guarantee
If no cheaper model meets your quality threshold, the premium model stays. Always.
Proof-carrying
Every decision is SHA-256 sealed and independently auditable months later.
Shadow first
See savings projections on real traffic before any routing change goes live.
The #1 fear: "will quality drop?"
GOVERIS never downgrades below your quality floor.
Every request carries a quality floor (0–1). The kernel only considers models whose quality score meets or exceeds that floor. If the cheapest qualifying model is the premium one, GOVERIS keeps it and explains why. This invariant is mathematically proven across thousands of randomized property tests.
- Quality invariant
- Proven proptest: 10K+ random parameter sets
- Premium preserved
- Always when no cheaper model meets the floor
- Enforcement
- You decide shadow first → review → then enforce
One system, three answers
See where the money went—and what to change safely.
Observe without exporting prompts.
Run a private, metadata-only Docker image inside your VPC. Production routing remains untouched.
Test policy changes before rollout.
Replay downgrade, block, and routing policies against the same traffic with an auditable reason for every decision.
How it works
From metadata to a defensible action plan in four steps.
- 01Mirror metadata
Model, token, latency, cost, tenant, and workflow signals—never prompts.
- 02Replay decisions
Calybris evaluates cost, risk, quality floors, and candidate routes in shadow mode.
- 03Compare policies
Measure projected savings and the exact calls each policy would keep, downgrade, or block.
- 04Ship the audit
Receive an executive PDF, evidence hashes, and a prioritized implementation plan.
The deliverable
A board-readable report backed by replayable evidence.
Not another dashboard someone has to monitor. The audit tells you where spend concentrates, which policies are safe to test, and what evidence supports each recommendation.
- Requested versus selected model analysis
- Cost, quality, risk, and confidence trade-offs
- Evidence hashes and reproducible scenario methodology
Decision trace
Every call gets an audit reason.
Policy replay
Compare cost, quality, and downgrade pressure.
Client package
Deliver PDF, HTML, JSON, and artifact hashes.
Measured, not blended
Speed is reported at the layer where it was measured.
Kernel, durable HTTP, and storage profiles are separate experiments. These are three-run or complete-profile medians from one local host—not a cloud SLO and not a fastest-run screenshot.
- Integer kernel
- 8.6M/s 3-run median · overflow-safe
- Durable HTTP
- 6,084 req/s c=128 · 1 ms mock
- HTTP p99
- 44 ms 15,000 requests · 0 failed
- Durable WAL
- 10,684/s current contended host
- Real data benchmark
- 1M rows WildChat + OpenAssistant · 0 errors
- Marginal savings
- 14–45% over your existing routing · measured per customer
- Kernel p99 (real)
- 152 µs 22-model catalog · SHA-256 proof
- HTTP throughput
- 2,597/s 1M requests · 32 threads · WAL on
Real data benchmark: 1M rows (500K WildChat + 47K OpenAssistant + synthetic fill), 0 errors. Marginal savings depend on your current routing maturity: teams with no routing see ~44%, teams with basic routing see ~26%, teams with smart routing see ~14%. The shadow replay pilot measures YOUR actual marginal savings, not theoretical maximums. Measured on a consumer Windows host with NTFS storage. We publish local-host numbers, not cloud SLO claims.
Engineering proof
294 tests · fault injection · proptest · loom exhaustive concurrency
Hardened
10 critical fixes: WAL deadlock, NaN bypass, timing leak, CORS
Instant audit
GET /api/v1/audit/report → JSON + Markdown spend analysis
Already using Helicone, LangSmith, or Phoenix?
They show what happened. GOVERIS decides what should happen next.
Observability tools tell you which model was called and how much it cost. GOVERIS tells you whether a cheaper model would have satisfied the quality floor, proves it cryptographically, and enforces the policy when you're ready.
- Observability
- Both GOVERIS + Helicone can coexist
- Prescriptive routing
- GOVERIS only auto downgrade / block / cache
- Proof trail
- GOVERIS only SHA-256 sealed, audit-grade
- No prompt capture
- GOVERIS only metadata-only by design
Executive control
GOVERIS explains where model spend went, then tests what should change.
The first product is not another chat wrapper. It is a private shadow-replay pilot for teams that already use LLMs and need evidence before they change a production route.
Detect low-risk, high-cost traffic that can be downgraded, cached, or blocked.
Show the confidence, risk, latency, and quality constraints behind each decision.
Run in shadow mode first, then graduate proven rules into the proxy path.
What teams actually save
Three teams. Three problems. Real numbers.
These are replay-based projections from the GOVERIS synthetic benchmark dataset. Each scenario uses 500,000 decisions with realistic tenant, model, and volume distributions. Production results will vary. We don't claim savings until a real pilot confirms them.
GPT-4o used for every ticket. GOVERIS identified 62% of calls as low-complexity, downgraded to GPT-4o-mini. Quality score unchanged. Monthly bill dropped from $4,200 → $2,780.
Claude Opus used across all workflows. GOVERIS preserved Opus for contract review (0.95 quality floor) but downgraded research queries to Sonnet. Blocked 3 high-risk PII-adjacent calls per week.
Multi-step agent retrying with flagship models. GOVERIS detected 41% of calls were repeated prompts (semantic cache eligible) and 23% used premium models for tool-call formatting. Cache + downgrade = 38% total reduction.
These projections use the GOVERIS catalog estimate methodology: replay existing traffic against conservative, balanced, and aggressive policy scenarios. We publish the methodology, not just the number.
Actual savings require a shadow replay pilot with your production metadata.
01
Mirror metadata
Keep prompt content local while a private image observes decision envelopes.
02
Simulate policy
Compare conservative, balanced, and aggressive routing scenarios.
03
Shadow gateway
Replay traffic safely before enforcing model downgrades or blocks.
04
Control spend
Move from report to governed OpenAI-compatible inference.
The problem
LLM bills are made of thousands of unreviewed decisions.
Teams see the invoice, but not the call-level economics: why a premium model was used, whether a cheaper model would satisfy the quality floor, or which tenant is creating avoidable spend.
Expensive models are used for low-value or repeated work.
There is no durable record explaining why each AI call was allowed.
Confidence, quality floor, risk pressure, and tenant budget live outside routing.
Client deliverable
From mirrored traffic metadata to a board-readable spend audit.
A private Docker image runs inside your environment, evaluates a metadata-only mirror, and keeps raw events local. Goveris produces aggregate findings and what-if policy scenarios without asking your team to export prompt logs.
Inputs
Mirrored model, token, latency, tenant, workflow, risk, and confidence envelopes.
Analysis
Savings estimate, downgrade pressure, negative-net calls, audit coverage.
Output
Local dashboard, aggregate review, recommendations, and what-if policy table.
GOVERIS spend audit
shadow replay
Audit coverage
100%
Premium pressure
actionable
Policy scenarios
3 what-if paths
Client gets
Local dashboard, aggregate findings, policy recommendations
Proof trail
input hash, generated artifacts, replayable decisions
deployment private Docker image data boundary customer VPC traffic mode metadata-only mirror prompt capture disabled enforcement off during replay
Sample audit PDF
Graphical report package
A client-facing example with spend shape, what-if policy simulation, tenant attribution, recommendations, and artifact hashes. The sample is deliberately marked as pilot data, not a production savings claim.
Total calls
500,000
Audit coverage
100%
Catalog estimate
33.36%
Requested baseline
$4,796.52
Current GOVERIS
$3,196.55
Balanced replay
$3,549.42
Observed
Conservative
Balanced
Aggressive
Platform workflow
Start with evidence. Enforce only after the policy survives replay.
Cost governance report
Find avoidable model spend without touching production traffic.
What-if simulation
Estimate conservative, balanced, and aggressive policy effects.
CALYBRIS gateway
Allow, downgrade, block, cache, or retry each model call.
Private Docker shadow replay
Observe the economics without handing over production logs.
We provision a private image in your registry or VPC. Your gateway mirrors only the decision envelope—not prompt or response content. Calybris evaluates each call in non-enforcing mode and stores the replay trail inside your boundary.
IMAGEprivate-registry / goveris-shadow:pilot
INGESTmodel · tokens · latency · tenant · workflow tags
EXCLUDEDprompts · responses · credentials · customer PII
MODEobserve and replay · never alter the live response
OUTPUTlocal ledger · aggregate review · rollout candidates
- 01Deploy beside the gateway
Read-only pilot image; no provider key is required for replay.
- 02Mirror metadata for 7–14 days
Production traffic continues unchanged while Goveris evaluates alternatives.
- 03Review before enforcement
Promote only policies that survive quality, risk, budget, and shadow gates.
Adoption path
Fixed scope. Clear price. No surprise integration bill.
Every package starts read-only. Savings are scenario estimates until your outcomes validate them.
A focused metadata-only scan for one team that needs a fast, defensible spend baseline.
- 7-day observation window
- Up to 50,000 metadata events
- 1 provider · 1 business unit
- Executive brief + top 10 actions
The full private Docker pilot: attribution, proof trail, what-if policies, and board-ready reporting.
- 7-day observation window
- Up to 250,000 metadata events
- Up to 2 providers · 3 business units
- PDF + dashboard + evidence hashes
- 3 policy scenarios + review session
Provider contract validation and a canary-ready plan for teams considering governed execution.
- 14-day technical engagement
- Idempotency + timeout reconciliation
- WAL recovery + load evidence
- Canary, rollback, and SLO plan
- Enforcement remains approval-gated
Enterprise · Custom
Multi-team rollout, private registry, and deployment-specific acceptance evidence.
- Custom event volume and retention
- Managed PostgreSQL/KMS integration
- Security and procurement package
One-command pilot
Deploy in your VPC with a single docker compose.
No agent installation. No log export. No prompt capture. Mirror approved metadata to GOVERIS for 7 days, then pull the audit report from the API.
docker compose -f docker-compose.pilot.yml up -d
Mirror metadata to POST http://goveris:8080/api/v1/route
GET /api/v1/audit/report → JSON + Markdown audit package
Start here
Get your LLM spend audit in 7 days.
Book a 15-minute call. We'll scope the pilot, deploy a private Docker image in your VPC, and return a board-readable audit. No prompt capture. Metadata-only observation.
Early adopter offer
First 5 pilots: Shadow Scan at $290
40% off the $490 Shadow Scan for early adopters who help us build the first case studies. Same deliverable, same quality, same private VPC deployment.
Claim early adopter spotLive policy replay
Change the workload. Watch Calybris decide in real time.
The first scenario runs automatically. When a Calybris demo endpoint is available the decision is server-authored; otherwise the page uses the disclosed deterministic browser policy. Either way, the proof and decision factors appear immediately—without a paid model call.
Default on this page: deterministic browser replay
Connect a Calybris demo endpoint and the same panel switches to a server-authored decision.
Decision ledger
Order-status request / support-emea / standard
- Requested
- claude-opus-4-8
- Selected
- pending
- Net value
- --
- Expected value
- --
- Estimated cost
- --
- Quality floor
- --
- Risk penalty
- --
- Proof fingerprint
- --
Decision reason
Waiting for evaluation.
Fallback chain: --
/api/v1/route
not evaluated
Waiting for a policy decision...
Synthetic data, disclosed
Realistic shape without pretending it is customer traffic.
The public dataset is generated deterministically from bounded distributions for token volume, model mix, tenant concentration, quality floors, confidence, and risk. It contains no prompts, customer logs, or personal data. Cost figures use the checked-in model catalog; savings are scenario estimates, not a production claim.
- Public sample
- 500 decisions
- Report replay
- 500,000 decisions
- Deterministic seed
- 73,120,526
- Workload mix
- 5 tenants / 5 use cases