AI-Production

The production-rigor layer for AI features, token metering with spend caps, a CI eval gate, PII/moderation guardrails, and a versioned prompt registry.

The AI-Production family is the layer between "the model call works in dev" and "the model call survives production": every call is metered and capped, every prompt change is scored against a regression gate before it ships, and every input/output crosses one guardrail boundary.

HTTP/1.1 402 Payment Required

{
  "error": {
    "code": "spend_cap_reached",
    "message": "Spend cap reached: circuit breaker open",
    "details": { "scope": "ai:complete" }
  }
}

The cap is enforced before the provider is called, in integer credit units, fail-closed, over budget returns 402, never an unbounded charge.

What's in the family

  • ai-kit: the metered inference gateway — infer, inferStream, embed, and embedMany, all behind the same reserve-before-provider-call chokepoint, composing prompt-registry, ai-meter, guardrails and ai-config behind Vercel AI SDK v7.
  • ai-meter: Postgres-atomic reserve/reconcile token metering, per-tenant spend caps, and a circuit breaker. Integer credits only, no floats.
  • ai-evals: a regression gate for prompt and model changes: defineEval() scores a dataset, gateAgainstBaseline() fails the build on a real score drop.
  • guardrails: a fail-closed guard around every model call: PII redaction, a swappable moderator, and an unconditional secret-shape gate.
  • prompt-registry: append-only prompt versioning with a mutable alias pointer, so promoting or rolling back a prompt is a pointer move, not a redeploy.
  • credits: the integer credit wallet the spend caps debit against: append-only ledger, debit-before-spend, fail-closed 402 on an empty balance.
  • field-crypto: per-tenant field encryption, shared with the Compliance, Local-first, and Provenance families, for any prompt input or output you store.

Install

bunx --package @caisson-sh/cli@latest create-caisson --name caisson-app --module <id@version>
cd caisson-app
bun install

Repeat --module <id@version> once per module you want. Every module installs from public npm, no token needed; see Getting started for the full flag reference.

How it composes

Metering, guardrails, and the eval gate sit at the same seam: a call to a model provider is metered and checked against the cap before the provider is reached. The eval harness runs the same prompts offline in CI, so a regression fails the pull request instead of a customer's session:

$ bun run eval
FAIL  prompts/summarize@v3
  faithfulness  0.71   gate >= 0.80
1 regression — exit 1. Build blocked.

The prompt registry versions the string both the live call and the eval read, and the credit wallet is the ledger the spend cap debits against.

Composing with the base

AI-Production reads the tenant id from @caisson-sh/tenancy-rls's bound context, so a spend cap is never checked against the wrong tenant's budget. Guardrails and the eval harness run independent of billing, wire @caisson-sh/billing separately if a spend cap should also gate a subscription tier.

License

Every module here is Apache-2.0 and installs from public npm.

verify-pack

The out-of-band verifier for a Caisson audit evidence pack — validates the signed file manifest, the pack seal, the receipt chain, and every per-row anchor signature, without trusting executable code from inside the pack.

ai-kit

The AI Production Kit — one metered gateway for language inference, streaming, and embeddings, enforcing reserve-before-provider-call accounting behind Vercel AI SDK v7.