Enrichments
Enrichments
They add a figure to a mandate, a line to a decision’s detail or a section to a dossier — within what the rules allow. Settings › Modules shows which run on a deployment. Cost control follows from governed AI; it is not the promise.
Costs and savings
- Every call is priced from the model catalog, and its cost is sealed with the price that made it.
- Realized savings — an answer reused, a cheaper route that served — and estimates — a simulation, a projected avoided cost — are separate totals, never added together.
- Spend per mandate is in Mandates › Spend, the detail in Mandates › More › Cost details. Calls made outside the gateway can be reported with
POST /api/ingestor the SDK’s wrappers.
x-azoth-provider: openai
x-azoth-model: gpt-4o-mini
x-azoth-cost-usd: 0.00000930
x-azoth-cost-saved-usd: 0.00000000
x-azoth-cache-hit: false
x-azoth-energy-wh: 0.002100
x-azoth-latency-ms: 412curl -X POST https://www.azothos.ai/api/ingest \
-H "X-Azoth-API-Key: $AZOTH_API_KEY" -H "Content-Type: application/json" \
-d '{ "events": [{ "provider": "openai", "model": "gpt-4o", "inputTokens": 1240,
"outputTokens": 128, "costUsd": 0.0082, "latencyMs": 340, "agentId": "support-bot" }] }'Model routing
Within what the rules allow, a call can go to a model that meets your quality and latency bounds — the cheapest when cost breaks the tie. The quality weighed is your workspace’s own pass rate on its reference cases, or neutral when nothing has been measured; each decision’s reason says which.
Semantic cache
An answer can be reused for an equivalent question, each reuse traced (X-Azoth-Cache-Hit). A risk classifier keeps precision-sensitive tasks — code, finance, legal — to exact matches. Turn it off for one call with "azoth_use_cache": false. A system in vault mode never reads or writes the cache.
Prompt optimisation
Rule-based compression before the call, its quality scored, with no model involved. A proposal for a mandate applies only once a person accepts it.
Energy and carbon
Energy and carbon per mandate, greener routing, and an auditable trace of the CO₂ avoided against a conservative baseline. These figures are estimates, not carbon accounting.
Simulations
Test a candidate rule against past traffic before publishing it, or run a governance dry-run against a proposed execution context.
Model catalog
- Prices, context windows, capabilities, hosting zones and lifecycle, in versions: a published version never changes, and a new price is a new version. A person publishes the prices; the providers’ listings only propose them.
- The catalog is signed outside the app, and self-hosted and air-gapped deployments import it signed. A model served without a price is sealed at zero and listed among the items waiting for a person.
- Every day, each mandate’s models are read against the catalog: a model ending within 30 days, and a cheaper model of the same provider that could replace it. A mandate can be held to its models; a new one joins the list only after the mandate’s reference set passed on it.
AI footprint in code repositories
- Turned on repository by repository, Azoth reads what AI coding tools left in a GitHub repository’s history: co-author trailers (declared), markers and configuration files. An author is kept only as a pseudonym — never a name, an address or a message — and turning the reading off deletes everything read.
- An owner or admin can freeze a repository’s findings into a sealed statement, exported as a dossier the verifier checks.
- With
azoth proxyandazoth hook install(CLI), a commit names the calls the gateway sealed: such a commit is shown as attached to the sealed call #seq, never as proof that the code came from it. A marker is never a proof.
Value and risk portfolio
Per mandate, the measured cost and realized savings beside the value its owner declares with hypotheses — never summed — a composite risk score with its components, and a frozen monthly report.
Every figure Azoth cites carries a Measured, Modelled or Projected tag: see the metrics methodology.